[brand-header]

# Comfy Labs Online -- User Manual

## 1. What Comfy Labs Online is

Comfy Labs Online is a small, invite-only web app for making images, reference sheets and short video clips with local AI models, and for making finished images bigger. Everything runs on one machine: Chipp's RTX 5090. The card does one job at a time, so every render waits its turn in a single queue shared by everyone, including Chipp's own tools.

There are seven workflows: **CONCEPT IMAGE**, **OMNI IMAGE**, **EDIT IMAGE**, **SHEETS**, **H3 VIDEO**, **UPSCALE IMAGE** and **UPSCALE VIDEO**. What you make lands in your own **Gallery**, which only you can see. **Libraries** keeps your favourite renders, your uploads and your project Moodboards in one place.

## 2. Getting in

### Step 1: the email code (Cloudflare)

Open the link Chipp sent you. A Cloudflare page asks for your email address and mails you a one-time code. Enter it. This happens once per browser; the sign-in then lasts about a month.

### Step 2: your invite code (first visit only)

The first time you arrive, the app asks for the invite code Chipp gave you and a display name.

![The invite code screen](ComfyLabs-Online-Manual_images/invite_8.svg)

1. **Signed in as** -- the email you verified with Cloudflare.
2. **Invite code** -- the code from Chipp, in the form `XXXX-XXXX-XXXX`. Each code works once.
3. **Display name** -- the name shown on your jobs. It cannot contain `[]()!<>` or start with a dash.
4. **Join** -- creates your account and opens the app.

Too many wrong codes in an hour lock the form for a while. If your code fails, ask Chipp for a new one.

### Your work is private

Your Gallery shows only your own renders, and your Libraries only your own favourites, projects and uploads. Other users never see them. The admin can open any user's Gallery and Libraries to help out (section 8), and every change an admin makes there is recorded under the admin's name.

### Log out

**LOG OUT** at the top of the sidebar, beside your name, ends your Cloudflare session on this browser. Next time you will be asked for an email code again (not the invite code).

## 3. The screen

![A tour of the screen](ComfyLabs-Online-Manual_images/tour_8.svg)

1. **Workflow buttons** -- one per workflow, with a one-line description. The one you are in is orange. Click another to switch; your form is kept.
2. **Library** -- **Gallery** (your own renders), **Libraries** (Favorites, projects, uploads and Moodboards), **My Jobs** (everything you ran), **Assets** (your saved sheets) and **Docs** (this manual).
3. **Header: page and model** -- the page you are on and the model it will use. The header stays pinned while you scroll.
4. **5090 status** -- in the header on every page: the card's state (FREE or BUSY), who is on it, how many jobs wait, and its memory (section 4).
5. **Your name** -- the display name everyone sees.
6. **LOG OUT** -- see section 2.
7. **Model pill** -- the model(s) the current settings will run.
8. **The workflow form** -- everything below the model pill belongs to the workflow.

### Tooltips

Almost every control has a small **?** next to it. Hover it for what the control does and, where it matters, how long it takes on the GPU.

![A tooltip](ComfyLabs-Online-Manual_images/tooltip_8.svg)

1. **?** -- hover to read the tooltip.
2. **The tooltip** -- here, what KREA 2 Turbo is good at and its GPU time.

## 4. Sharing the 5090

One click on **Render** (or **Edit**, **Render sheet**...) is one job. Jobs run one at a time in the order they were queued. Friends and Chipp's own local tools share the same card, so a job can wait even when nobody else is on the site.

The header, pinned to the top of every page, shows the card and your own job:

![The header with the 5090 status and the working indicator](ComfyLabs-Online-Manual_images/queue_header_8.svg)

1. **Page and model** -- where you are.
2. **State** -- the pill reads **FREE**, **BUSY**, **FINISHING** (the last job is done and is handing back the GPU) or **UNKNOWN** (gpu-share did not answer). The dot is teal when the card is free and orange when it is busy.
3. **Who is on the card** -- a friend's name and workflow, or "Chipp: Local job" for Chipp's own tools. The line under it gives the running time and estimate (local jobs carry no estimate, "unknown") and how many jobs are waiting. Hover it to read a line cut short.
4. **GPU memory** -- current use of the card's 32 GB, as a compact readout and bar (hover it for the full figure). Not visible in this shot: while your working indicator is showing, the status block is too narrow for the memory bar, so the header drops it.
5. **Working indicator** -- while your job waits, it says who is ahead and when it should start, for example "Queued #3 behind Chipp and Bob, starts in ~2:10". If a local job of unknown length is ahead, it says "wait unknown". Once your job starts, it shows the running step. For about 10 seconds after the job finishes, it reads "Done:" followed by the job's title. Click it to open the result.

On a narrow window the header drops the memory bar first, then the second line, then who is on the card; the state pill always stays.

While anything waits, or something is wrong, a queue frame at the top of the workflow page shows it:

![The queue frame with a notice, the queue and your jobs](ComfyLabs-Online-Manual_images/queue_panel_8.svg)

1. **Notices** -- shown only when something is wrong; here, the queue is paused by the admin. Other notices you may see: "queue worker is down" or "ComfyUI offline". Jobs keep their place and wait; nothing is lost.
2. **The queue** -- every waiting job with its position, estimated start and estimated duration.
3. **Your jobs** -- your own jobs, with their position and **Cancel**.

About the times:

- Every time is an **estimate**, shown as a range ("~38-50 s"). It comes from the last measured runs of the same kind of job. "No estimate yet" means that combination has not been measured.
- The first image from a model includes loading the model, often 20-60 s. Later images from the same model are much faster.
- **Cancel** removes a queued job at once. A running job is stopped at the next safe point. In CONCEPT and OMNI the images that already finished are kept as a partial result.

## 5. The workflows

Two controls appear in several workflows and work the same way everywhere:

**Pick from a library.** Wherever a workflow takes an image (or, in UPSCALE VIDEO, a video), an expander such as **Pick from a library** or **Pick a reference from a library** lets you reuse something from your **Favorites** or one of your projects (section 7). Choose the **Library**, page through it with **Newer** and **Older**, and press **Use** under the item you want. The original stays in your library unchanged. Each workflow remembers the library you last picked from. You can always upload directly instead.

**Moodboard style.** CONCEPT IMAGE, OMNI IMAGE, SHEETS and H3 VIDEO have an optional **Moodboard** box, and EDIT IMAGE has the **Apply Moodboard style** tool. It opens on **No Moodboard**, which keeps your prompt exactly as written. Pick one of your project Moodboards (listed by project name and version, for example Smoke Board v1) to add its written visual style to what the model receives; you can read the result under "What each model receives" (H3: "What H3 receives"). The version is fixed when the job is queued, so **Re-run** renders with the same one, shown as "(from the job)". If the Moodboard has a newer version since, **Use current Moodboard version** switches to it. With no Moodboards yet, the box says so; you build them on the Libraries page (section 7).

![Pick a reference from a library and the Moodboard box](ComfyLabs-Online-Manual_images/pick_library_3.svg)

1. **Pick a reference from a library** -- opens the picker (here on OMNI IMAGE).
2. **Library** -- **Favorites** or one of your projects.
3. **Use** -- puts that item into the workflow.
4. **Newer** / **Older** -- page through the library; the line beside them shows the page and how many items it holds.
5. **Moodboard** -- open, it lists **No Moodboard** and each of your project Moodboards with its version.

### CONCEPT IMAGE -- fast ideas, five models side by side

Use it to try an idea quickly, or to see how five local models read the same prompt: KREA 2 Turbo, Z-Image Turbo, ERNIE Turbo, FLUX.2 Klein 4B and Ideogram 4.

![CONCEPT IMAGE: models](ComfyLabs-Online-Manual_images/concept_models_8.svg)

1. **Model tick box** -- tick one model, or several to compare them in columns. A model whose files are missing is disabled with the reason.
2. **Aspect** -- 1:1, 16:9, 9:16, 4:3, 3:4, 3:2 or 2:3, per model.
3. **Size** -- **Draft** (about 1 MP, fastest) or **Full** (the model's top size: KREA native 2K, Ideogram 2048x1152, the rest about 2 MP).
4. **Model notes** -- exact pixels, what the model is good at, and the measured time per image and with the model load.
5. **Images per model** -- 1 to 4. Image number k uses the same seed on every model, so rows compare fairly.

![CONCEPT IMAGE: prompt and Render](ComfyLabs-Online-Manual_images/concept_prompt_8.svg)

1. **One prompt / Idea to N variations** -- one prompt for every image, or one idea turned into N different prompts (N = images per model). With variations, row k uses prompt k on every model.
2. **Prompt** -- put any words the picture should show in double quotes: `a sign that reads "KEEPER'S WALK"`.
3. **Refine prompt** -- a short GPU job (a local language model) rewrites your prompt for the ticked models and lists what it changed. The result goes back into the box; edit it freely.
4. **What each model receives** -- the exact prompt each model gets. Ideogram 4 gets a structured caption the app writes from your prompt.
5. **Render** -- queues one job for all images.
6. **Estimate** -- how many images and roughly how long.

**Moodboard** (above the prompt) -- optional; see **Moodboard style** at the start of this section.

Step by step:

1. Tick the models, set aspect and size for each.
2. Choose images per model.
3. Write the prompt (or an idea), optionally **Refine prompt**.
4. Press **Render** and watch the queue panel.

If you change the models (or the idea) after a Refine or Variations run, the helper's result is marked out of date and **Render** is blocked until you run it again or keep it as your own text.

**Results.** Open a finished job from **My Jobs** or the gallery:

![CONCEPT IMAGE: a result](ComfyLabs-Online-Manual_images/concept_result_8.svg)

1. **Close this result** -- back to an empty result area; your form stays.
2. **One column per model.**
3. **Image details** -- model, seed, size and GPU time.
4. **Download** -- the full-size PNG.
5. **Final** -- opens OMNI IMAGE with this image's prompt, for a high-quality final render on Qwen-Image-2.1.
6. **Send to** -- hands the image to another workflow (section 6).

**Re-roll** at the bottom repeats the whole job on new seeds.

**Tips and honest limits:**

- KREA and Z-Image spelled quoted words correctly in every test; KREA at 2K can add extra, misspelled signs around them.
- ERNIE sometimes doubles the subject (two lighthouses) and misspells secondary text. Klein garbles longer signage.
- Ideogram's safety filter can return a grey "blocked" frame; the app flags it on the result.
- Refine prompt can drop details from your prompt. Read the change list and put back what matters.

### OMNI IMAGE -- one prompt, references, Qwen vs H3

Use it for a finished still, with or without reference images, on **Qwen-Image-2.1**, **MiniMax H3**, or both side by side (**A|B**).

![OMNI IMAGE: model and references](ComfyLabs-Online-Manual_images/omni_form1_8.svg)

1. **Model** -- Qwen-Image-2.1, MiniMax H3, or A|B (both, same seeds, side by side).
2. **A filled reference slot** -- its number is its token: this image is `@1`.
3. **Remove** -- takes the image out of the slot.
4. **The next empty slot** -- drop an image or **Browse files**, or open **Pick a reference from a library**. Up to 10 references for Qwen, 9 for H3 and A|B. PNG, JPG or WebP up to 20 MB. With no references it is plain text to image.
5. **Prompt** -- write `@1`, `@2`... where each reference belongs. An `@N` with no image blocks Render and says why.

Single views work best as references. A full multi-panel sheet makes the model draw the subject two or three times.

The optional **Moodboard** box sits above the prompt (see **Moodboard style** at the start of this section).

![OMNI IMAGE: prompt help, size and Render](ComfyLabs-Online-Manual_images/omni_form2_8.svg)

1. **Refine prompt** -- rewrites your prompt for the selected model, keeps every `@N`, and lists what changed.
2. **Generate from images** -- writes a prompt from your references (needs at least one).
3. **What each model receives** -- the final prompt per model, including the wording the app adds.
4. **Aspect** -- with exact pixels per model.
5. **Size** -- Draft or Full.
6. **Renders per model** -- 1 to 4.
7. **Transparent PNG** -- Qwen only: a cut-out on a transparent background.
8. **Advanced: seed** -- random seeds or a fixed one.
9. **Render** -- one job for the whole batch.

![OMNI IMAGE: an A|B result](ComfyLabs-Online-Manual_images/omni_result_8.svg)

1. **Close this result.**
2. **One column per model** -- each row shares a seed.
3. **Image details** -- model, seed, size, time.
4. **Download / Send to SHEETS** -- Send to SHEETS copies the image into your uploads and opens SHEETS with it added as a reference.
5. **Qwen-Image-2.1 wins / MiniMax H3 wins** -- pick the better image of the row. Picks feed the win rate the admin sees.
6. **Pick status.**

**Re-roll** repeats the settings on new seeds. "What each model received" shows the exact prompts used.

**Tips and honest limits:**

- MiniMax H3 is weak on text inside the image (a boat's name came out garbled).
- Qwen with many references can drop a subject: a 10-reference test duplicated one person and left out another. Use fewer, clearer references when every subject matters.
- Refine prompt can drop setting details (place, time, weather). Check the change list.
- H3 stills take longer: the card restarts its model for every H3 image (about 45 s) plus 1-3 min per image.

### EDIT IMAGE -- change an image, keep the rest

Use it to change part of an existing image on Qwen-Image-2.1: an instruction, a relight, the brand noir regrade, an extended canvas, a transparent background or one of your Moodboard styles. The result keeps the source's exact size.

**Getting an image in:** upload one, open **Pick from a library**, or send one here from another workflow (**Send to EDIT**). Images with transparency are flattened onto white first.

![EDIT IMAGE: the form](ComfyLabs-Online-Manual_images/edit_form_8.svg)

1. **Tool** -- **Instruction**, **Relight** (six presets or your own description), **Regrade** (the locked brand noir grade or a custom one), **Outpaint**, **Remove background**, **Apply Moodboard style**.
2. **Instruction** -- describe the change in plain words; everything you do not name is kept. `@1` is the image you edit; `@2`, `@3`... are references.
3. **Must stay exactly** -- things the model must not touch.
4. **Add reference images** -- up to 9, for things to bring into the picture. The budget line shows how much of the 12 MP image budget you use.
5. **Pick a reference from a library** -- reuse an image from your Favorites or a project.
6. **Limit to a painted area** -- only change where you paint (Instruction, Relight, Regrade).
7. **Advanced** -- mask grow and feather, the seed, and "What the model receives".
8. **Edit** -- queues the edit.

**Apply Moodboard style** restyles the image with one of your project Moodboards (palette, light, texture, finish) while keeping its subject, identity and composition. Pick the Moodboard in the **Moodboard** box that appears with this tool; **Must stay exactly** works here too. A Moodboard works with this tool only; the other tools ignore it. It cannot be limited to a painted area.

**Painting an area:**

![EDIT IMAGE: painting the area to change](ComfyLabs-Online-Manual_images/edit_mask_8.svg)

1. **Limit to a painted area** -- switch it on to show the canvas.
2. **Your paint** -- paint over what should change.
3. **Brush** -- brush size.
4. **Clear mask** -- start the painting again.
5. **Canvas buttons** -- undo, redo and clear (Ctrl+Z also undoes a stroke).
6. **Painted share** -- how much of the image is painted.

Every pixel outside the painted area is copied from the original. The app checks this and fails the job if even one differs.

**Extending the canvas:**

![EDIT IMAGE: Outpaint options](ComfyLabs-Online-Manual_images/edit_outpaint_8.svg)

1. **Tool: Outpaint.**
2. **Note** -- the model repaints the whole frame, so the original area can shift slightly in tone. It is not pixel-exact.
3. **Extend by** -- to an **Aspect ratio**, or a percentage **Per side**.
4. **Aspect** -- the new shape.
5. **Keep the image at** -- where the original sits in the new frame.
6. **What the new area shows** -- a short description of the new content.

**Remove background** makes a transparent PNG. If the result has no real transparency, the job fails and says so.

**History, before/after and Undo:**

![EDIT IMAGE: history and before/after](ComfyLabs-Online-Manual_images/edit_history_8.svg)

1. **Original** -- the source you started from.
2. **Current step** -- each edit is one step and becomes the image you edit next.
3. **Undo / Start over** -- Undo makes the previous step current again and deletes nothing; later steps stay selectable. Editing an earlier step starts a new branch.
4. **Before/after slider** -- drag to compare.
5. **Step details** -- tool, size, seed, time.

Under it: **Download PNG**, **Try again** (same edit, new seed) and **Prompt** (what the model received).

**Tips and honest limits:**

- An outpaint's original area can change tone (see above). A paste-back of the original was tried and left a visible seam, so it was removed.
- The brand noir regrade is deliberately very dark, as in the example above.
- Typical times: about 60 s for an instruction, relight or regrade at 2 MP; about 40-50 s for an outpaint or remove background.

### SHEETS -- character, prop and location reference sheets

Use it to turn 1-4 reference images into one identity sheet (turnaround, expressions, callouts, palette) on Qwen-Image-2.1, and keep it as a versioned Asset.

![SHEETS: sheet type and references](ComfyLabs-Online-Manual_images/sheets_refs_8.svg)

1. **Sheet type** -- Character, Prop / vehicle, Location - interior, Location - exterior.
2. **Add reference images** -- one to four single views of the subject. **Pick a reference from a library** adds one from your Favorites or a project.
3. **Each reference** -- shown with its `<imageN>` number.
4. **Role** -- what the image is for (Identity / face, Full body, Outfit, Extra angle...).
5. **Remove.**

![SHEETS: zones and facts](ComfyLabs-Online-Manual_images/sheets_zones_8.svg)

1. **Zones** -- the panels the sheet will contain (turnaround, outfit callouts, palette, expression grid, height guide, hands study, action poses).
2. **Name** -- printed in the sheet's header band.
3. **Age** -- a number; the sheet is told to show that age.
4. **Must-not-drift details** -- the features that must match in every panel.

Fill the facts yourself, or press **Describe from my images** (a short GPU job, about 10 s) to pre-fill them.

The optional **Moodboard** box (above the prompt) adds a Moodboard's palette, light, texture and avoid rules to the prompt's style paragraph only; the layout, roles and facts stay as the template writes them. See **Moodboard style** at the start of this section.

![SHEETS: options, prompt and Render](ComfyLabs-Online-Manual_images/sheets_render_8.svg)

1. **Style** -- match the references, or a set style.
2. **Text** -- **App labels** (the app draws the name, type, style, zone legend and palette in clean bands) or **No text**.
3. **Background** -- always pure white.
4. **Image budget.**
5. **Generate prompt** (optional) -- reads your references and adds what they show (colours, materials, marks) to the template. You can skip it; the template alone works well.
6. **The prompt** -- see or edit the full prompt.
7. **Render sheet** -- queues the sheet (about 1-2 min).
8. **Describe from my images** -- fills the facts from your references.
9. **Estimates.**

![SHEETS: a result](ComfyLabs-Online-Manual_images/sheets_result_8.svg)

1. **Close this result.**
2. **Show the raw render (no bands)** -- the model's image without the app's header and footer.
3. **Header band** -- name, type, style, creator and date, drawn by the app.
4. **Footer band** -- zone legend and exact palette swatches.
5. **Details** -- seed, panels found, palette hex codes.
6. **Download sheet PNG.**
7. **Prompt** -- what the model received.

![SHEETS: follow-ups](ComfyLabs-Online-Manual_images/sheets_actions_8.svg)

1. **Re-roll** -- same prompt, new seed.
2. **The one change** -- describe one thing to fix ("the back view is missing the scarf").
3. **Fix** -- re-renders with only that change.
4. **Save as Asset** -- keeps the sheet as a versioned asset. Here it is already saved as version 1. A Re-roll or Fix of a saved sheet is added to the asset as a new version automatically.
5. **Open Assets** -- see section 7.

**Tips and honest limits:**

- The expression grid sometimes comes back as 2x4 (8 faces) instead of 3x3.
- Location references full of people can leak a person into one panel. People-free references render clean.
- Exterior sheets tend to repeat the wide view's angle in the play-area panels.
- The model never draws per-panel labels (it garbles text); the app draws the bands instead.
- Use single views as references, never a whole sheet.

### H3 VIDEO -- a short clip with sound

Use it for one MiniMax H3 clip per click, with native stereo sound: from text alone, from a first and/or last frame, or from 1-9 reference images. The page is one scroll in three parts, **SOURCE** (what the clip starts from), **DIRECT** (what happens, the camera and the sound) and **OUTPUT** (size and length), then a **Request summary** with the one **Render** button. A clip you open shows below as its **RESULT**.

#### SOURCE: what the clip starts from

![H3 VIDEO: SOURCE with a first frame and two reference images](ComfyLabs-Online-Manual_images/h3_form1_8.svg)

1. **First frame**, **Last frame** and **Reference images** -- tick the inputs you want.
2. **The first-frame slot** -- the chosen frame, its file name and, when its shape differs, the note that it will be center-cropped. Upload one or switch on **Pick from a library** to fill it.
3. **Remove** -- clears the frame.
4. **@N** -- each reference card's number in the prompt, with its role.
5. **Role** and **note** -- what to keep from this image (both optional).
6. **Up** / **Down** / **Remove** -- reorder or drop the reference.
7. **Add reference images** -- drop more images here, or switch on **Pick references from a library** below it.

Tick what you want to give H3; leave all three boxes unticked for text to video:

- **First frame** -- the clip opens on this image.
- **Last frame** -- the clip ends on this image.
- **Reference images** -- 1-9 images named @1...@9 in the prompt (identity, props, setting, style).

Each ticked frame gets its own slot: upload a PNG, JPG or WebP, switch on **Pick from a library**, or send one here from another workflow. A frame of another shape is center-cropped to the clip's size, and the slot tells you so. **Remove** clears the frame; the file stays in your uploads. A ticked box without an image keeps Render off, and the request summary names the fix, for example "Add the first frame, or untick First frame."

Without reference images, H3 uses its first/last-frame model: text only, a first frame, a last frame, or both. With **Reference images** ticked it uses its reference model instead, and any first or last frame you add anchors the opening or the ending of the clip. Those frames are **not** @N images; your @N numbers stay the same.

**Reference images.** Add up to nine with **Add reference images** (drop one or several; they become the next @N slots in order), with **Pick references from a library**, or by a hand-over from another workflow. Each card shows its @N and the image, with:

- **Role** (optional) -- **Person/identity**, **Clothing/prop**, **Location/setting**, **Style/look** or **Object**; the default is **No role**.
- **Note** (optional) -- up to 80 characters on what to keep from it, for example "the red coat".
- **Up** / **Down** -- swaps it with its neighbour; the prompt's @N tags swap with it.
- **Remove** -- later slots move up, so check the @N tags in the prompt. The file stays in your uploads.

Roles and notes appear above the prompt as a legend (for example "@1 Person" and "@2 station.png" on one line) and are given to **Refine prompt** and **Generate from images**; the render itself never sees them. In the prompt, write what each image contributes -- for example, "the woman from @1 carries the kettle from @2 in the railway station from @3." H3 receives `<Picture 1>`, `<Picture 2>`... in place of those tags. These are identity, object, setting or style guides, **not** automatic first or last frames.

Re-roll and Re-run retain available frames, references, roles and notes. If any original reference image is missing, Re-run clears all reference slots rather than silently assigning a later image to the wrong @N; choose the images again in order and check the prompt. Once you have uploaded references, **Manage uploaded files** beside them shows how much of your 1 GiB quota they use and lets you delete old ones. Images used by a queued or running job cannot be deleted, and a finished clip's Re-run needs its images again. **All uploaded files** there opens Libraries > Uploaded files (section 7). This page does not offer reference video or audio uploads.

#### DIRECT: what happens, the camera and the sound

In page order: the **Prompt** with its helper buttons right under it, the **Camera move**, the **Sound** (Audio, Silence, Dialogue), the optional **Moodboard**, and **What H3 receives**.

#### Camera move

Instead of writing the camera move into the prompt, you can pick one. The app writes it as plain prose, the way H3 likes it, and puts it first in what H3 receives.

A one-line summary under the **Camera move** title says what the section adds (for example "Slow dolly in" with "look: 35mm film", or "None (write the move in the prompt)"). Open **Choose move and look** to change it. At its top are two checkboxes:

- **On** -- untick it to switch the whole section off. H3 then gets no camera sentence and no look from here, whatever was picked, and your picks come back when you tick it again.
- **Randomize** -- each render gets a camera move drawn at random instead of the one you picked. The drawn move shows in place of Group and Move, and goes first in what H3 receives. A new move is drawn after each Render, never the same one twice in a row. The draw leaves out the experimental moves and the moves with a blank to fill. With a first frame it also leaves out moves that set their own opening view. Speed and Look still apply. **Re-roll** keeps the move the clip was made with.

![H3 VIDEO: the camera move picker](ComfyLabs-Online-Manual_images/h3_camera_8.svg)

1. **Group** -- the kind of move: dolly, pan and tilt, tracking, orbit, crane, zoom and lens, drone, POV, reveal, stylized, lateral, compound. The last group, **Experimental: may not finish in a short clip**, holds moves too big for a 3-4 s clip (hyper-zoom, macro zoom, hyperlapse, bullet time, barrel rolls, a fast 360 orbit); they say so when picked.
2. **Move** -- the default, **None: I write it in the prompt**, adds nothing, so the clip is made exactly as before.
3. **Speed** -- **As described** keeps the move as written, with its own pace words ("gradual, deliberate and suspenseful"). Very slow / creeping, Slow, Steady, Fast, Accelerating or Decelerating replaces the move's speed word ("A slow pan" becomes "A fast pan") or is added in front of a move that has none, and drops the pace words so they never contradict your speed. Speed is off for the moves whose pace is the move itself (marked "its own pace, no Speed" in the table below: a whip pan, a snap zoom, a fast dolly in, a static tripod, an FPV dive, a hyperlapse...).
4. **The blank** -- a few moves need a detail: the slow pan asks what it **Reveals**, the tilts ask **From ... to ...** ("from her muddy boots to her face"). Render stays off, with a note, until it is filled in.
5. **Look** -- an optional film or camera look, added as one sentence after your prompt: 35mm film, 65mm IMAX film, VHS camcorder, Kodak Ektachrome, infrared, CCTV, or underwater housing. **None** adds nothing.
6. **The camera sentence** -- exactly what H3 gets first, with your speed and blank filled in.

With a move picked:

- The **Prompt** asks for the action, then the light. Leave the camera out of it: if the prompt also names a camera move (a phrase such as "push-in", "dolly in", "pan left", "tilt up", "truck right", "zoom in", "tracking shot", "handheld camera", "drone shot", "orbits around", "crane up", "arc shot", "pedestal down", or "the camera moves / pans / rises..."), a warning says so, because two moves fight each other. Ordinary words are left alone: a delivery truck, a frying pan or a head that tilts is not a camera move.
- **What H3 receives** shows the full text in this order: the camera sentence, your prompt, the look, the Moodboard style if you picked one, the dialogue, then the audio line.
- With a first frame (on its own, with a last frame, or anchoring reference images), a move that sets its own opening view (marked "sets the opening view" in the table below) shows a warning: your first frame already is the opening view, so H3 may ignore the move or jump away from your picture.
- **Refine prompt** keeps the move word for word: it is told the move is fixed and rewrites only the action, the light and the setting.
- **Suggest motion from the first frame** is told the move you picked and writes the action only; it never picks a move for you. With no move picked, both work as before.
- **Re-roll**, **Continue** and the Gallery's **Re-run** bring back the move, speed, blank and look with the rest of the settings. If the clip had a move or a look, they also turn the section **On** and show that exact move with **Randomize** off.

##### Every camera move

The moves come from a camera cheat sheet (`docs/reference/Camera_Prompt_Cheat_Sheet.md`), rewritten as plain prose for H3. Two duplicate pairs were merged (the dolly zoom and the "vertigo combo"; tracking with an orbit and "tracking + orbit"), and the bird and car POV shots were made general. The three walking tracking shots also set the action: the subject walks.

| Group | Move | What the camera does |
|---|---|---|
| Dolly / push / pull | Slow dolly in (push) | A slow dolly in: the camera physically moves forward through space toward the subject, and the background perspective widens as the depth increases. |
| Dolly / push / pull | Slow dolly out (pull) | A slow dolly out: the camera physically moves backward away from the subject, which grows smaller in the frame as more of the surroundings come into view. |
| Dolly / push / pull | Fast dolly in (rush) | A fast dolly in: the camera rushes forward toward the subject's face, with a sudden sense of urgency. (its own pace, no Speed) |
| Dolly / push / pull | Dolly zoom (vertigo, zolly) | A dolly zoom: the camera physically moves backward while the lens zooms in at the same time, so the background warps and stretches while the subject stays the same size in the frame. |
| Pan & tilt | Slow pan | A slow pan: the camera turns to the right on a fixed point, revealing [reveals]; gradual, deliberate and suspenseful. (you fill in: Reveals) |
| Pan & tilt | Whip pan left | A whip pan to the left: the camera whips violently sideways with strong directional motion blur between two points. (its own pace, no Speed) |
| Pan & tilt | Tilt up | A slow tilt up: the camera tilts vertically upward on a fixed axis, [from ... to ...]. (you fill in: From ... to ...) |
| Pan & tilt | Tilt down | A slow tilt down: the camera tilts vertically downward on a fixed axis, [from ... to ...]. (you fill in: From ... to ...) |
| Tracking & following | Leading shot (backward tracking; subject walks toward camera) | A leading shot: the subject walks forward toward the camera while the camera moves backward at exactly the same speed, keeping the subject centred; a steady, even pace. |
| Tracking & following | Following shot (behind; subject walks away) | A following shot: the subject walks away from the camera and the camera follows behind at the same speed, keeping the same framing. |
| Tracking & following | Side tracking (parallel; subject walks sideways) | A side tracking shot: the subject walks from left to right and the camera trucks alongside at the same pace, keeping them in profile. |
| Tracking & following | Tracking with orbit | A tracking shot with an orbit: the camera travels alongside the moving subject while curving in a subtle orbit around them; dynamic movement around the subject. |
| Orbital & arc | 180-degree arc (half orbit) | A half orbit: the camera travels on a curved track 180 degrees around the subject, from a front view to a back view. |
| Orbital & arc | Slow cinematic arc | A slow arc shot: the camera moves in a wide curve around the subject, revealing their side profile; gentle and unhurried. |
| Crane & pedestal | Crane up | A crane up: the camera rises upward and backward on a jib arm, ending in a high-angle shot looking down at the subject. |
| Crane & pedestal | Crane down | A crane down: the camera descends from a high bird's-eye view down to the subject's eye level; smooth and controlled. (sets the opening view) |
| Crane & pedestal | Pedestal up | A pedestal up: the whole camera rises vertically, straight up, from waist height to eye level, without tilting. |
| Crane & pedestal | Pedestal down | A pedestal down: the whole camera lowers vertically, straight down, from eye level to waist height, without tilting. |
| Zoom & lens | Smooth zoom in | A smooth zoom in: the camera stays perfectly fixed in place while the lens focal length increases, magnifying the subject. |
| Zoom & lens | Snap zoom (crash zoom) | A snap zoom: a sudden, extremely rapid zoom straight into the subject's eyes, an instant, high-impact change of framing. (its own pace, no Speed) |
| Zoom & lens | Rack focus | A rack focus: the shot opens with the foreground subject sharp and the background blurred; halfway through the focus shifts, so the foreground blurs and the background turns sharp. |
| Zoom & lens | Focus pull reveal | A slow focus pull: the shot starts completely out of focus in soft bokeh, and the lens pulls focus until the subject is razor-sharp; a gradual reveal. (sets the opening view) |
| Drone & aerial | Epic drone reveal | An epic drone reveal: the camera starts low behind a ridge, then rises vertically while tilting down to reveal the subject and the vast horizon beyond. (sets the opening view) |
| Drone & aerial | Drone flyover | A steady high-altitude drone flyover: an aerial shot moving straight forward, passing high over the subject and the wide landscape below; smooth and even. |
| Drone & aerial | Top-down (god's eye) | A slow top-down drone shot: the camera hangs directly overhead, pointing straight down at 90 degrees, and rotates. (sets the opening view) |
| Drone & aerial | FPV drone dive | A fast FPV drone dive: aggressive, agile camera movement diving down the side of a building toward the subject. (its own pace, no Speed) |
| Drone & aerial | Wide drone orbit | A wide drone orbit: a sweeping aerial circle around the subject that shows the immense scale of the surroundings from above. |
| POV & immersive | First-person POV walk | A first-person POV walk: the camera moves forward with the slight bob and sway of a person walking, approaching the subject. |
| POV & immersive | POV flight (flying creature) | A steady POV flight: the camera is locked to the head of a flying creature soaring over the landscape, with clean, stable forward motion and no jitter; its wings are visible at the edges of the frame, moving in rhythm. |
| POV & immersive | Vehicle-mounted POV | A vehicle-mounted shot: the camera is rigged low on the side of a moving vehicle, looking along its flank at the ground and a spinning wheel; the road rushes past at speed. |
| POV & immersive | Handheld / documentary | A handheld camera: it moves with organic human jitter, slight instability and a subtle breathing motion, as if held by an operator with no stabiliser. |
| Reveal & transition | Reveal from behind (wipe) | A sliding reveal: the view starts blocked by an object in the foreground, then the camera slides sideways to reveal the subject hidden behind it. (sets the opening view) |
| Reveal & transition | Fly-through | A fly-through: the camera physically flies through a small gap, window or opening to reveal the scene behind it. |
| Stylized & special | Dutch angle | A Dutch angle: the camera is tilted sideways on its roll axis for the whole shot, so the horizon line runs diagonally. |
| Stylized & special | Worm's eye tracking | A worm's-eye tracking shot: the camera moves along the ground looking up at the subject, so everything towers overhead. |
| Stylized & special | Static tripod (no camera move) | A static tripod shot: a fixed lens and no camera movement at all; the camera stays perfectly still while the subject and the surroundings provide all the motion. (its own pace, no Speed) |
| Lateral / slider | Truck left | A truck left: the camera physically slides sideways to the left on a track, with strong parallax as the foreground moves against the background. |
| Lateral / slider | Truck right | A truck right: the camera physically slides sideways to the right on a track, with strong parallax as the foreground moves against the background. |
| Compound movements | Crane reveal (rise + pan) | A crane reveal: the camera rises vertically while panning horizontally to reveal the wide landscape or cityscape. |
| Compound movements | Push in + tilt up | A push in with a tilt up: the camera moves forward while tilting upward, for a dramatic reveal of the character or building. |
| Compound movements | Tracking + handheld | A handheld tracking shot: the camera tracks the subject with slight organic handheld instability, for documentary realism. |
| Compound movements | Crane down + push in | A crane down with a push in: the camera descends from a high angle while pushing forward toward the subject, landing at eye level. (sets the opening view) |
| Compound movements | Spiral rise (orbit + pedestal up) | A spiral rise: the camera orbits the subject while rising vertically at the same time, a spiralling upward reveal. |
| Experimental: may not finish in a short clip | Cosmic hyper-zoom (space to street) | A cosmic hyper-zoom: the shot begins in outer space looking at Earth, and the camera zooms in with extreme, continuous speed, plunging through the layers of the atmosphere and the clouds, resolving the city and diving down between the buildings. (sets the opening view; its own pace, no Speed) |
| Experimental: may not finish in a short clip | Extreme macro zoom (face to pupil) | An extreme macro zoom: the camera begins on a detailed portrait and zooms in continuously with extreme magnification, past the pores of the skin and deep into the iris and pupil. |
| Experimental: may not finish in a short clip | Hyperlapse | A hyperlapse: the camera moves forward through the scene while time runs accelerated, with sped-up motion, light trails and compressed day-night cycles. (its own pace, no Speed) |
| Experimental: may not finish in a short clip | Bullet time | A bullet-time shot: ultra-slow motion with the subject frozen mid-action while the camera orbits around them. (its own pace, no Speed) |
| Experimental: may not finish in a short clip | Barrel roll / vortex | A barrel roll: the camera spins 360 degrees clockwise on its own axis while moving forward, a disorienting vortex. |
| Experimental: may not finish in a short clip | FPV dive + barrel roll | An aggressive FPV dive: the camera dives steeply and makes a 360-degree barrel roll halfway down, for extreme energy. (its own pace, no Speed) |
| Experimental: may not finish in a short clip | Fast 360 orbit | A fast 360-degree orbit: the camera circles the subject in one continuous full loop while the background spins behind them. (its own pace, no Speed) |

#### Prompt, sound and Moodboard

![H3 VIDEO: prompt and sound](ComfyLabs-Online-Manual_images/h3_form2_8.svg)

1. **Prompt** -- the camera move first, then the action, then the light (with a move picked above, just the action and the light). Restate the place, time and weather.
2. **Audio** -- describe the sound ("steady sea wind, one gull cry").
3. **Silence** -- no sound at all.
4. **Dialogue (optional)** -- one line for one speaker. The clip is then one continuous shot, so lip-sync holds.
5. **What H3 receives** -- the full prompt.
6. **Refine prompt** -- for text/first/last modes, the existing camera-first H3 craft suggests a better motion prompt and audio line. In **Reference to video**, it uses MiniMax H3's [full-reference guidance](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md): image roles, retention analysis, shot-by-shot motion, soundscape and music. Its six sections keep editable @N labels, and the Audio box keeps your original sound instructions if the suggestion drops them. With a first frame, the local VLM sees that unnumbered opening image before @1; check how the refined prompt uses both.
7. **Suggest motion from the first frame** -- offered when you have a first frame and no reference images. **Generate from images** appears instead with reference images: the local VLM drafts a prompt using every chosen reference with its role and note, and starts from your first frame if one is set. Neither button renders the video; read and edit its output first. **Keep as my prompt** hides the note about where the prompt came from.
8. **The @N legend** -- each reference's number and role, above the Prompt.
9. **Moodboard** -- optional; see below.

Under Dialogue, the optional **Moodboard** box adds one of your project Moodboards' visual style to what H3 receives, after the shot and look and before dialogue and sound; your @N references are untouched. See **Moodboard style** at the start of this section. **Continue** keeps the clip's Moodboard for the next shot.

#### OUTPUT: size, length and seed

![H3 VIDEO: OUTPUT and the Request summary](ComfyLabs-Online-Manual_images/h3_form3_8.svg)

1. **Size** -- 16:9 native 1344x768 (best quality), 9:16 portrait 768x1344, 1:1 square 768x768, 16:9 720p 1280x736, and three **draft** sizes: 16:9 draft 672x384, 9:16 draft 384x672 and 1:1 draft 384x384. A draft renders the same full clip at a quarter of the pixels, with any inputs, in about a third of the time. Use drafts to try prompts, camera moves, frames, references and seeds, then render the keeper at full size or upscale it (see **Upscale this take** below).
2. **Length** -- a slider in seconds and frames (24 fps, 17-frame steps), up to H3's maximum of 15.08 s (362 frames), capped at what the card has been measured to handle for this size and these exact inputs.
3. **Current limit note** -- under the slider: the longest length for these inputs and, in plain words, what limits it and how to get more. Hover the slider's **?** for the measured detail.
4. **Advanced: seed** -- **Fixed seed** off gives a new random seed each clip; on, you choose the **Seed** to repeat a clip.
5. **Request summary** -- the state (here READY) and the facts of the request; see below.
6. **Estimated GPU time.**
7. **Render** -- queues the clip.

#### Request summary and Render

Right above **Render**, the **Request summary** shows the state of your request:

- **READY** -- everything is in place; **Render** is on.
- **NEEDS INPUT** -- something is missing.
- **UNAVAILABLE** -- this length has no safety proof for this combination of size and inputs.

After the state come up to five facts (what runs, for example "Reference to video: 3 images + first frame"; the size; the length; the camera move and look if any; the sound), then **Estimated GPU time**. When it is not READY, up to three lines name exactly what to fix, for example "Write a prompt." or "Add at least one reference image, or untick Reference images."

**Render** queues one clip as one job. Time grows with length: at 16:9 a 4.5 s clip takes about 10 minutes of GPU time and a full 15 s clip about 50 minutes, more with many reference images (the estimate above Render shows it). Long clips hold the shared card for everyone, so try ideas at a draft size first. Your draft stays in the form after Render, so you can change it and render again; changing it never changes the queued job.

**Length limits:** One rule sets every limit, live, for the exact clip in the form: its size, prompt, reference images and first/last frames. The slider stops at the longest length the production 5090 has already rendered safely with at least as much work. "Safely" means a full 20-step render that kept at least 4 GB of video memory free and ComfyUI's shared GPU memory flat. An admin raises a limit only by rendering the next 17-frame step that way. H3's own maximum is 15.08 s (362 frames). Measured limits on 3 October 2026:

| Clip | 16:9 and 9:16 | 1:1, 720p and drafts |
|---|---|---|
| Prompt only, or one reference image | 15.08 s (362 frames) | 15.08 s |
| First frame, last frame, or both | 14.38 s (345 frames) | 15.08 s |
| 4 wide reference images | 14.38 s (345 frames) | 15.08 s |
| 9 wide reference images | 14.38 s (345 frames) | 15.08 s |
| 9 wide reference images + first and/or last frame | 13.67 s (328 frames) | 15.08 s |
| 9 square reference images | 12.96 s (311 frames) | 15.08 s (720p: 14.38 s) |

What makes a clip "more work" is its size, its length, and everything H3 has to read: each reference image (a square one costs more than a wide one), a first or last frame, and a long prompt. So **fewer reference images, wide rather than square ones, and a shorter prompt leave room for a longer clip**. The heaviest clip, 9 square references at 16:9, stops at 12.96 s: the next step (13.67 s) failed while turning the finished clip into video, so longer is refused. If the note under the slider says a combination has not been measured yet, remove the first/last frame or ask the admin to measure it. If none passed, the request summary says UNAVAILABLE with the reason and Render stays off. Use **Continue** to chain clips beyond 15 s.

#### RESULT

![H3 VIDEO: a result](ComfyLabs-Online-Manual_images/h3_result_8.svg)

1. **Close this result.**
2. **The clip** -- plays with sound.
3. **Clip details** -- mode, size, frames and seconds, seed, sound track, render time.

![H3 VIDEO: actions and frame slider](ComfyLabs-Online-Manual_images/h3_actions_8.svg)

1. **Safety proof** -- the checks run around the render (for the curious).
2. **What H3 received.**
3. **Download** -- the MP4 (H.264 with stereo sound, 24 fps), the main action of a result.
4. **Re-roll** -- same prompt, frames and settings, new seed.
5. **Continue** -- takes the last frame and opens a new clip that starts on it with the same prompt, sound, size and Moodboard. With reference images it also retains your @N images; it becomes available only where a safe first-frame-plus-references length has been measured. If the original reference clip was longer than that limit, the next form shortens the length and explains why. Nothing renders until you press **Render**.
6. **Frame slider** -- pick any frame of the clip.
7. **Upscale this take** -- with its size box, **1080p** or **2K**; see below.
8. **Send clip to UPSCALE VIDEO** -- see below.

**Upscale this take** (beside Download, with its size box **1080p** or **2K**) makes this exact take bigger: the same framing, motion, timing and sound. It queues an UPSCALE VIDEO job that enlarges the clip's saved raw H3 data with the learned latent upscaler, and the result appears in UPSCALE VIDEO and your Gallery. 4K is not offered. It is the quick way to finish a draft you like. When the clip cannot be upscaled this way, the button is off and a line says why, for example a clip saved without its raw H3 data ("Send the clip to UPSCALE VIDEO for a pixel upscale instead"). **Send clip to UPSCALE VIDEO** opens UPSCALE VIDEO with this clip selected, to pick another engine or size.

![H3 VIDEO: using a frame](ComfyLabs-Online-Manual_images/h3_frame_8.svg)

1. **The chosen frame.**
2. **Save this frame as a still** -- adds it to the gallery as an image.
3. **Send this frame to OMNI** -- use it as an OMNI reference.
4. **Admin: Certify next rung** -- admin only (section 8).

**Send this frame to UPSCALE IMAGE** sits beside these and opens UPSCALE IMAGE with the frame as a still.

**Preview** and its **Render at full size** and **Re-roll preview** buttons are gone: pick a draft size instead. Clips made earlier with the old Turbo preview still play and download; their result says so and has Re-roll and Continue off. Re-run such a preview from My Jobs to render it at full or draft size. An old preview "Upscale this take" job cannot be re-run; upscale a finished clip as described above.

**Tips and honest limits:**

- H3 is weak on text inside the picture.
- Refine prompt can drop place, time and weather words; the app puts those sentences back and says so in the change list.
- Every H3 clip restarts the card's model first, so even short clips take minutes.

### UPSCALE VIDEO -- make a clip bigger without changing its timing

Choose one of **your** videos with **Pick a video from a library**, upload MP4/MOV/WebM up to 100 MB (**Upload video (MP4, MOV or WebM)**), or send a clip from H3 VIDEO or Gallery. The original stays unchanged. A clip picked from your library or sent from the Gallery is already on the server, so it has no 100 MB limit (up to 1 GiB). Source limits: constant frame rate, at most 5 minutes / 18000 frames / 60 fps and 4096 pixels per side; portrait clips, including phone video with rotation metadata, stay portrait. Uploaded videos are private, like everything you make. **Manage uploaded videos** frees space under your 1 GiB quota; completed results remain, but Re-run needs the source again. Inputs of queued or running jobs cannot be deleted. **All uploaded files** there opens Libraries > Uploaded files (section 7).

If Chipp's Agent rendered a new H3 clip, it belongs to the separate private **Agent** account rather than Chipp's gallery. An admin can open **Import my Agent H3 clips** and press **Copy to mine** to copy an eligible clip **and its raw video latent** into the admin's own private Gallery, without altering the Agent original. Older H3 videos made before raw-latent saving cannot get their exact original latent back from an MP4. New H3 renders (full and draft sizes) and admin certification renders save it.

**Choose the method before pressing Render; the app never switches engines automatically:**

- **Faithful (FFmpeg Lanczos):** enlarge existing pixels without inventing detail; any eligible clip, 1080p, 2K-width or 4K.
- **SeedVR2 3B video:** local temporal detail restoration, 1080p or 2K-width; may change textures. **SeedVR2 7B video** is the heavier, slower option for more detail. Check faces and text carefully. Neither offers 4K until that model and canvas pass a 5090 memory/quality check.
- **NVIDIA RTX Video Super Resolution:** local neural enhancement in small frame batches, usually more conservative than diffusion. It can still reconstruct details; source-to-output enlargement is limited to 4x per side.
- **Real-ESRGAN (per frame):** local Photo or Illustration/animation model on the NVIDIA GPU; useful for defined lines and graphics, with possible flicker in moving fine detail. Native 4x processing limits eligible source sizes.
- **H3 learned latent upscale:** local learned enlargement of a *new H3 VIDEO clip with its saved raw latent*, at 1080p or 2K-width, up to 4x. It keeps the same take and restores the original soundtrack. H3 VIDEO's **Upscale this take** queues exactly this method. Older H3 clips without raw data can use any eligible pixel method instead.

![UPSCALE VIDEO: H3 learned latent upscale with a different latent file](ComfyLabs-Online-Manual_images/upscale_video_latent_3.svg)

1. **Upscale method** -- here **H3 learned latent upscale**.
2. **Output preset** -- the target size.
3. **Advanced: use a different latent file** -- see below.
4. **Upload a .latent file.**
5. **Latent file** -- which latent to enlarge; "The clip's own saved latent" by default.
6. **Render UPSCALE VIDEO.**

With **H3 learned latent upscale** picked, **Advanced: use a different latent file** lets you choose the raw H3 latent to enlarge: for a clip whose own latent is missing, or one made outside this app. Upload one with **Upload a .latent file** (it stays private) or choose one of your own under **Latent file**; the default, "The clip's own saved latent", uses the clip's own. The file must match the clip exactly (its width, height and frame count), or Render stays off. The app cannot tell whether a matching file is the same take: the wrong latent upscales a different video and keeps this clip's sound, so check you picked the right one.

There is no cloud option: every method runs locally on the 5090.

The result is a separate MP4 with the original number of frames, timing and soundtrack (compatible AAC copied, other source audio converted to AAC), plus a poster. Download it or **Upscale again with other settings**; a silent source stays silent. Choose Faithful when identity, lettering or image content must stay closest to the original. The local tiled H3 re-sampler is *not* one of these methods; it needs separate seam, motion, audio and memory proof first.

### UPSCALE IMAGE -- take a keeper to print size

Use it when you have an image you like and need it bigger: for print, a poster, a wallpaper or a close crop. It makes one image 2x, 3x or 4x larger (or to an exact long edge) on one of four engines, or on two side by side (**A|B**). It does not change what is in the picture; it adds pixels.

**The image to upscale.** Upload one (PNG, JPG or WebP, up to 20 MB), open **Pick from a library** and choose one of your images, or send one here from another workflow with **Send to UPSCALE** (section 6). One image per job. Transparent PNGs stay transparent.

![UPSCALE IMAGE: the image and the engine](ComfyLabs-Online-Manual_images/upscale_engine_8.svg)

1. **The image** -- its file name, size in pixels and megapixels.
2. **Choose another image** -- puts this one away so you can upload or pick another. Nothing is deleted.
3. **Detail** -- "Adds real texture and sharpness with the SeedVR2 3B diffusion model. Best for photos and renders; it invents fine detail, so faces and small text can change slightly. About 10-30 s." The default, and the right first try for most photos and renders.
4. **Fast** -- "Sharpens with NVIDIA RTX Video Super Resolution without inventing detail: the most faithful to your image, a little crisp at 4x. About 5-20 s." Pick it when the image must stay exactly as it is (faces, logos, small text), or for the largest sizes.
5. **Real-ESRGAN** -- "Crisp, clean edges with Real-ESRGAN. Best for illustrations, UI and flat graphics; choose the Photo or Illustration model. It smooths fine texture and invents some edge detail. About 10 s."
6. **Max detail** -- "Like Detail with the larger SeedVR2 7B model: the richest invented detail and colour, so faces and text can change more. The slowest: about 40 s at 4096 px." Pick it for a hero image where extra invented detail is welcome.
7. **A|B** -- "Runs Detail and Fast on the same image in one job and shows them side by side with a slider, so you can pick. About as long as Detail plus a few seconds." Use it when you are not sure which suits the image.
8. **Real-ESRGAN model** -- shown only for Real-ESRGAN. **Photo** (realesrgan-x4plus) for photos, renders and textures; **Illustration / anime** (realesrgan-x4plus-anime) for drawings, anime, UI, flat colour and line art.

The same descriptions are in the **?** next to **Engine**. If the server cannot run an engine, its description ends in "UNAVAILABLE:" and the reason; the other engines keep working.

![UPSCALE IMAGE: size, options and Render](ComfyLabs-Online-Manual_images/upscale_size_8.svg)

1. **Make it** -- **A scale** or **A long edge**.
2. **Scale** -- 2x, 3x or 4x on both edges. With **A long edge** you type the exact pixels of the longer side instead, up to 4x the image; the other side follows the shape of the image. Output sides are rounded to multiples of 8 px.
3. **Colour correction** -- Detail, Max detail and A|B only: keeps the colours of your image (SeedVR2 can shift them slightly). On by default.
4. **Save as** -- **PNG** keeps every pixel (big files); **JPEG (quality 92)** is much smaller. A transparent result is always saved as PNG, and the page says so.
5. **Advanced: seed** -- Detail, Max detail and A|B only: SeedVR2's seed (default 42). The same seed, image and settings give the same result.
6. **The output line** -- the exact result before you render, for example "Output will be 2752 x 1536 (4.2 MP, 2x), about 11-15 s." For A|B it adds "twice (Detail and Fast)".
7. **Render** -- one job, one image (two for A|B).

With a transparent PNG, a **Keep transparency** box appears (on by default): the transparent areas stay transparent. The colour is upscaled by the engine and the transparency is enlarged separately. Turn it off to flatten the image onto white.

**Size limits (long edge, in pixels):**

|Engine|Image in, at most|Result, at most|
|---|---|---|
|Detail|2048|4096|
|Max detail|2048|4096|
|Fast|any|8192|
|Real-ESRGAN|2048|8192|
|A\|B|2048|4096|

Every engine goes up to 4x, and each side of the image must be at least 64 px. A request over the limits does not queue: the output line turns yellow, **Render** is greyed out, and a red line says why. When the result would be too big, it names the largest scale that engine allows for this image (for example "The largest scale Detail allows for this image is 2x"). When the image itself is too big for the engine, it names the engine that can take it, for example "Try Fast (up to 8192 px)." Fast takes an image of any size.

![UPSCALE IMAGE: an A|B result](ComfyLabs-Online-Manual_images/upscale_result_8.svg)

1. **Close this result** -- hides it here. The job stays in My Jobs.
2. **Compare** -- which two images the slider compares: **Detail vs Fast**, or your image (**Source**) against either one. A single-engine result compares your image against the result.
3. **View** -- **Whole image**, scaled to fit, or **100% centre (640 px)**: the middle of the result at full size, where you see the detail. Your image is resized to match, like a plain resize.
4. **The before/after slider** -- drag the handle left and right. The labels in the top corners name each side.
5. **Image details** -- engine, scale, size before and after, render time.
6. **Download PNG / Send to** -- one pair under each image. Download saves the upscaled image, full size. **Send to** hands it to **EDIT**, **OMNI** or **SHEETS**; nothing renders until you ask there.
7. **Upscale again with other settings** -- puts this job's image and settings back in the form, so you can try another engine or size.

**Typical times on the 5090** (render only; waiting in the queue comes on top):

- Detail: about 9 s for a small image at 2x, 12-20 s up to 8 MP, 25-27 s for 4096 x 4096.
- Max detail: about 15 s for 2048 x 2048, 42 s for 4096 x 4096.
- Fast: about 5 s up to 8 MP, 8-16 s for 8192 px results. The first Fast job after a quiet spell takes about 20 s while the model loads.
- Real-ESRGAN: about 10-16 s at any size.
- A|B: about 16 s for a 2x of a 1376 x 768 image.

**Tips and honest limits:**

- Detail and Max detail invent detail. Check faces, hands and small text at **100% centre** before you use the result; if they changed too much, use Fast.
- Fast adds no new detail, so a soft image stays a little soft, only bigger.
- Real-ESRGAN smooths fine texture such as skin and fabric. It is the one to use for flat graphics.

## 6. Moving work between workflows

- **Send to** (CONCEPT results): **OMNI** (as a reference), **EDIT** (as the image to edit), **H3 VIDEO** (as the first frame), **SHEETS** (as a reference), **UPSCALE** (as the image to upscale).
- **Final** (CONCEPT): opens OMNI IMAGE with that image's prompt, for a finished render on Qwen-Image-2.1.
- **Send to SHEETS** (OMNI results), **Send this frame to OMNI** and **Continue** (H3 VIDEO).
- **Send to UPSCALE IMAGE**: under each OMNI result, beside Download on a SHEETS result, in the EDIT viewer, on crops you kept in Assets, and in the Gallery Details of your images. On H3 VIDEO it is **Send this frame to UPSCALE IMAGE**, under the frame slider.
- **Send clip to UPSCALE VIDEO**: on an H3 VIDEO result and in the Gallery Details of your videos. **Upscale this take** on an H3 VIDEO result queues the upscale straight away.
- **Send to** (UPSCALE results): **EDIT**, **OMNI** or **SHEETS**.

![The Send to menu](ComfyLabs-Online-Manual_images/sendto_8.svg)

1. **OMNI** -- adds the image as an OMNI reference.
2. **EDIT** -- starts a new edit on it.
3. **H3 VIDEO** -- uses it as the first frame.
4. **SHEETS** -- adds it as a SHEETS reference.
5. **UPSCALE** -- opens UPSCALE IMAGE with it as the image to make bigger.
6. **Send to** -- opens and closes this menu.

The image is copied into your uploads and the target workflow opens with it in place. A workflow that is not installed is greyed out.

## 7. Gallery, Libraries, My Jobs, Assets

### Gallery

Your own renders, newest first, 24 per page. Nobody else sees them.

![The Gallery](ComfyLabs-Online-Manual_images/gallery_8.svg)

1. **Scope** -- **All my renders**, **Favorites**, or one of your projects. With a library chosen, the page lists what is in it, its uploads included.
2. **All my renders** -- with a library in scope, switch it on to list every render while the star still adds to that library; off lists only what is in the library.
3. **Workflow** and **Kind** -- show only what one workflow made, or only images or videos.
4. **Columns** -- 4 or 3 across.
5. **Video cards** -- click the frame (or Details) to play the clip.
6. **Card details** -- creator, title, workflow, date, then the resolution (for example `1376 x 768 px`), kind, length for videos, and item number.
7. **Details** -- opens the item.
8. **Star** -- adds the render to the library in scope, or to **Favorites** when **All my renders** is chosen; a filled star removes it again (the file stays). Hover it to see which library it changes. Your Favorites are made the first time you star something.
9. **Trash** -- opens the item's Details with the delete question showing (see Delete below).

Click any image thumbnail to see the full-size original over the page. Click anywhere to close it.

![Gallery: the detail view](ComfyLabs-Online-Manual_images/gallery_detail_8.svg)

1. **Close.**
2. **The full image** (or video).
3. **Creator, title, workflow and date.**

Below the image is the prompt, then:

![Gallery: settings and actions](ComfyLabs-Online-Manual_images/gallery_detail2_8.svg)

1. **Settings** -- the job's settings.
2. **Job number and file size**, and the libraries the item is in ("In no library (the star adds it)" when none).
3. **Download.**
4. **Open in (workflow name)** -- opens the job on its workflow page.
5. **Send to UPSCALE IMAGE** (images) or **Send clip to UPSCALE VIDEO** (videos).
6. **Delete** -- see below.

**Delete** (on your own items; an admin sees it on every item) removes the item from the gallery and deletes its file from the server for good. It asks first: press **Delete for good** to go ahead or **Keep it** to back out. It cannot be undone. The job stays in My Jobs, marked "Deleted from the gallery". An asset you saved from it keeps its own copy. If a job that is still queued or running uses the file, Delete refuses until that job finishes.

An upload card (shown when a library is in scope) has the same star and trash. Its **Details** show the file's facts, its **Libraries** (see below), **Download** and **Delete**, which removes the upload from every library and its file for good.

### Libraries

**Libraries** in the sidebar keeps the things you want to reuse: your **Favorites** and any number of **projects**. A library holds your renders (starred in the Gallery) and your own uploaded images and videos. One item can be in Favorites and several projects at once. Every workflow's **Pick from a library** reads from here (section 5). The page is private to you; the line at the top shows the storage your uploads and Moodboards use. Uploads here have no quota.

At the top, **View** switches between **Libraries** (this part) and **Uploaded files** (below).

![Libraries: a project and an open upload](ComfyLabs-Online-Manual_images/libraries_3.svg)

1. **Library** -- **Favorites** or one of your projects. Browsing a library never selects it for a render.
2. **New project** and **Create project** -- type a name (up to 60 characters) and make an empty project.
3. **Project name**, **Rename** and **Delete project** -- shown for a project. Renaming keeps its items and Moodboard.
4. **Add images or videos to** (the library) -- PNG, JPEG and WebP images up to 20 MB; MP4, MOV and WebM videos up to 100 MB, several at a time. An upload always goes into Favorites too.
5. **Kind** -- show all kinds, only images or only videos.
6. **Manage** -- opens an item with its facts. **Libraries** lists every library that holds it; change the list and press **Save libraries**. An upload must stay in at least one library. **Delete upload** removes an upload from every library and deletes its file for good, after **Delete for good** / **Keep it**. **Close** puts the item away.

Page through a library with **Newer** and **Older**. An empty library tells you to star renders in the Gallery or upload here.

**Deleting a project.** **Delete project** first shows exactly what goes: the project and its memberships, the uploads made in this project (from every library and from disk, with their size), and its Moodboard with all its versions. Rendered outputs stay in the Gallery, and uploads made in Favorites or another project stay. Press **Delete for good** or **Keep it**. If a queued job still needs one of those files, the project cannot be deleted until that job finishes or is cancelled.

#### Moodboards

Each project can have one **Moodboard**: a written visual direction (Soul.md) made from 4-10 reference images of the project. Workflows apply it only when you choose it for a job (section 5); it is text, not a trained model.

![Libraries: a project's Moodboard](ComfyLabs-Online-Manual_images/moodboard_3.svg)

1. **Reference images** -- choose 4-10 images of this project that share one look. They are copied into the new version, so later changes to the library do not alter it.
2. **Build Moodboard** (later **Build a new version**) -- queues one GPU job that studies the chosen images and writes Soul.md as a new version. It runs in the shared queue and becomes current when it is ready; the page updates when it finishes. A failed build says why.
3. **Current version** and the version list -- each version says whether it was built from images or is an edit of another version, and whether it is ready, pending, failed or current.
4. **Make current** -- workflows offer this version from now on. Jobs already queued keep the version they were given.
5. **Soul.md version** -- which ready version to read or edit.
6. **Soul.md** -- the text, in seven sections: Identity, Palette, Lighting and Contrast, Composition and Lens, Texture and Finish, Motion Language, Avoid. Keep the title and the seven `##` sections, at most 800 characters each, and do not write `@1`-style tokens.
7. **Save as new version** -- saves your edit as a new version and makes it current; the version you edited stays unchanged. **Discard changes** puts back the saved text.

#### Uploaded files

**View: Uploaded files** lists every file you gave a workflow: uploads, hand-off copies, staged videos and latents (and, for the Agent, its inputs). Uploads you made into Favorites or a project are managed in the Libraries view instead. The H3 VIDEO and UPSCALE VIDEO upload managers open this view with **All uploaded files**.

![Libraries: Uploaded files with the delete confirmation](ComfyLabs-Online-Manual_images/uploaded_files_3.svg)

1. **Kind** (images, videos, latents, other), **Usage** (**Used by active job**, **Referenced by past jobs**, **No recorded job references**) and **Sort** (**Newest first**, **Oldest first**, **Largest first**). "No recorded job references" does not mean unused: an open form may still hold the file.
2. **Search** (part of a file name or a date folder) and **File modified between** (a date range).
3. **The count line** -- how many files match and their size, and how many you selected.
4. **Select this page**, **Select all N matching** (every file matching the filters, on every page), **Clear selection** and **Delete selected...** (at most 500 files at once). The selection is kept across pages.
5. **Each file** -- name, size, date and how jobs use it, with **Select**, **Open** (the file in full with its facts) and **Delete**.
6. **The confirmation** -- lists exactly what goes. Finished results and saved copies stay, but a Re-run or an open form will need the file uploaded again; it cannot be undone. Nothing can be deleted while a queued or running job uses one of the files; **Deselect these** drops the files in use (or changed since you selected them) so you can confirm the rest. Press **Delete N files for good**, or **Keep them** to back out.

After a delete, a line says how many files went and how much space came back, with **Per-file results** (each file's outcome: Deleted, Already missing or Failed). **Dismiss** hides it.

### My Jobs

Everything you ran, newest first, with status and timings.

![My Jobs](ComfyLabs-Online-Manual_images/myjobs_8.svg)

1. **Job number, workflow, title and status** (DONE, QUEUED, RUNNING, FAILED, CANCELLED).
2. **Timings** -- submitted, started, finished, time waited and GPU time. A failed job shows its error here.
3. **Open** -- shows the result on its workflow page.
4. **Re-run with changes** -- fills the workflow form with this job's settings so you can adjust and render again.

### Assets

Your saved SHEETS. Only you can change them.

![Assets: the library](ComfyLabs-Online-Manual_images/assets_list_8.svg)

1. **Type tabs** -- Character, Prop / vehicle, Location - interior, Location - exterior.
2. **Versions** -- how many and which is current.
3. **Open.**

![Assets: an asset](ComfyLabs-Online-Manual_images/assets_detail_8.svg)

1. **Back to the library.**
2. **Version** -- view any version.
3. **Set as current** -- make the viewed version the current one.
4. **Show the raw render (no bands).**

Further down: **Download sheet**, **Download raw**, **Download facts**, the facts and prompt, and the references.

![Assets: panels found in the sheet](ComfyLabs-Online-Manual_images/assets_panels_8.svg)

1. **Suggested front view** -- the app finds the panels in the sheet and suggests the front view.
2. **Kept** -- this panel is already a single-view reference.
3. **Keep as single-view reference** -- keeps a panel as its own image.

Kept panels appear under **Single-view references**. Use these, never the whole sheet, as references in other workflows. **Crop by hand** at the bottom lets you cut your own crop.

## 8. For the admin

The **Admin** page is in the sidebar for admins only. Its tabs: Online now, Sign-ins, Jobs, Queue, Users & invites, Usage, System health, Notifications.

### Other users' Gallery and Libraries

Admins see an **Owner** box at the top of the **Gallery** and of **Libraries**. It starts on your own name ("(you)"); pick another user to see their renders or manage their libraries. The Gallery also offers **All owners**, which lists everyone's renders (the star is off in that view, because there is no single library to add to). On Libraries, a note on every panel reminds you whose libraries you are changing, and every change is recorded under your name. **Copy to my Favorites** on another user's item copies the file into your own Favorites so your workflows can use it; the original stays theirs, and the copy is recorded. Workflow pickers only ever list your own libraries.

### Adding someone (two steps)

1. In the Cloudflare Zero Trust dashboard, add their email to the Access policy **Friends Only** (Access controls > Policies > Friends Only > Edit > Include Emails).
2. On **Users & invites**, fill in the form at the bottom and press **Create invite**. The app shows the invite code once, with a ready-to-send invitation email under it. Send them that email.

The screenshot shows the tab just after an invite was created for "Arik" (the form below is empty again, ready for the next one).

![Admin: Users & invites](ComfyLabs-Online-Manual_images/admin_users_8.svg)

1. **The code** -- shown once. Only a hash is stored, so copy it (or the email below) before you press Done.
2. **Before sending** -- a reminder to add their email to **Friends Only** in Cloudflare first. Without it they cannot get past the email code.
3. **Subject** -- hover the box for its copy button (top right).
4. **Body** -- the invitation email in plain text: their first name, the site address (https://comfy.land-page.com), how to get in with their email and this invite code, when the code expires, what to know about the shared site, a link to this manual, and your display name. Hover the box for its copy button.
5. **Open in email app** -- opens a new message in your mail app with the recipient, subject and body filled in. Some mail apps cut long mail links short; if the message arrives cut short, copy the text from the boxes instead.
6. **Done** -- hides the code and the email. Neither is stored or written to the log, so they cannot be shown again.
7. **First name** -- optional; used for the greeting ("Hi Arik,"). Left empty, the email starts with "Hi,".
8. **Their email** -- optional; the address they will sign in with. It must be a valid email address (it is saved in lower case). It goes into the email's sign-in step and becomes the recipient of **Open in email app**. Left empty, the email tells them to use the address the invite was sent to.
9. **Note** -- who the invite is for; only admins see it. Left empty, it becomes the first name and email, so the invite list stays readable.
10. **Expires in days** -- default 14.
11. **Create invite** -- makes the code and the email above.

Further down the tab: **Invites** (every invite: state, note, dates, who used it; revoke unused ones there), **Users** (everyone with an account, status, online and last seen) and **Manage user**.

![Admin: managing one user](ComfyLabs-Online-Manual_images/admin_manage_8.svg)

1. **Manage user** -- the selected user.
2. **Display name / Rename.**
3. **Block** -- pauses their access at once and cancels their queued and running jobs. **Unblock** reverses it.
4. **Set admin flag** -- admin rights also need the email in `config.toml` `[auth] admin_emails`.

### Removing someone

1. **Block** them in the app.
2. Remove their email from **Friends Only** in Cloudflare.
3. Optionally, revoke their Cloudflare session in Zero Trust so they are signed out everywhere.

### Queue control

![Admin: Queue](ComfyLabs-Online-Manual_images/admin_queue_8.svg)

1. **Queue state** -- running or paused.
2. **Pause online queue / Resume online queue** -- pausing stops new online jobs from starting; it never touches Chipp's local agent tickets.
3. **Local agent job** -- admins see the agent name and purpose (blurred in this manual); friends see "Chipp: Local job".
4. **A queued online job.**
5. **Up / Down / Cancel** -- reorder queued online jobs, or cancel any job.

### Usage

![Admin: Usage](ComfyLabs-Online-Manual_images/admin_usage_8.svg)

1. **Group by** -- Day (last 30 days) or Week (last 12 weeks), in UTC.
2. **Jobs and GPU minutes per user.**
3. **Chart** of the same.

Further down: the OMNI A|B win rate per model, and per-workflow counts with median GPU time and wait.

### System health

![Admin: System health](ComfyLabs-Online-Manual_images/admin_health_8.svg)

1. **ComfyUI** -- up or down, with version.
2. **GPU** -- memory in use and who holds the card (lease text blurred in this manual).
3. **Worker** -- running, with heartbeat. "Worker down" means no heartbeat for 30 s.
4. **Disk** -- free space on the data drive.
5. **Disk usage by user.**

Below it is the last 200 lines of today's log.

![Admin: Restart ComfyUI](ComfyLabs-Online-Manual_images/admin_restart_8.svg)

1. **Confirm box** -- tick to allow the restart.
2. **Restart ComfyUI** -- queues the restart through the 5090 queue. It waits its turn, so it never interrupts someone else's job.

### Notifications

The **Notifications** tab lists what the app sent to Chipp: new invite redemptions, first sign-ins, failed jobs, and worker or ComfyUI down. **Send test notification** checks the channel.

### H3: Certify next rung

On H3 VIDEO results, admins see **Admin: Certify next rung**. It runs one full 20-step test render at the next longer length (one 17-frame step) for a chosen size, prompt only. A pass raises the cap for every request that is no more work than the test; a fail closes that length and everything heavier. Use sparingly: near 15 s each attempt holds the card for about 50 minutes. Reference images and first/last-frame combinations are measured with the admin's campaign script (`scripts/h3-certify-campaign.py`), which queues one test at a time so other jobs run in between.

### Agent (MCP) jobs are private

Chipp's local agents can run workflows through the app's local API. Their jobs run as the user **Agent** and go through the same queue. Like every user, the Agent has its own private Gallery and sees only its own renders. An admin can look at them by choosing **Agent** in the Gallery's **Owner** box. To work on an Agent H3 clip, copy it with **Import my Agent H3 clips** in UPSCALE VIDEO (section 5).

## 9. Troubleshooting (admin)

|What you see|What to do|
|---|---|
|A workflow page shows an exception during beta|Send Chipp a screenshot of the full error and what you clicked, privately. Error details can include local paths or private information; do not post them publicly.|
|The site shows **Bad gateway (502)**|The app is not running. Start the scheduled task: `Start-ScheduledTask -TaskName 'Comfy Labs Online'`.|
|Cloudflare **error 1033**|The tunnel is down. In an admin PowerShell: `Restart-Service cloudflared`.|
|**ComfyUI offline** banner on the workflow page|Jobs wait in the queue; nothing is lost. The admin starts ComfyUI with **Restart ComfyUI** on System health, which goes through the 5090 queue.|
|**Worker down**|The worker stopped. The scheduled task normally restarts it; if not, restart the task.|
|The site is down **after a reboot**|The task starts at Chipp's logon. It stays down until Chipp logs in to Windows.|
|A friend's invite code fails|Check **Invites** for used or expired codes, and create a new one.|
|A friend cannot get past the Cloudflare page|Their email is not in **Friends Only**. Add it (section 8).|
