Turn a description into a MiniMax H3 video without leaving ComfyUI. Build a prompt you can inspect, generate clips with the official H3 workflows, plan a longer video, then cut and render the result in a built-in editor.

Everything lives in one ComfyUI sidebar, backed by 28 canvas nodes that share the same prompt pipeline. This is an independent custom-node project, not MiniMax's official Context-IR implementation.

<img src="assets/overview.jpg" alt="Conceptual overview of the MiniMax H3 Studio workflow" width="80%"> <details><summary><h2>Latest Updates - Click to expand</h2></summary> <details> <summary>1.4.0 — Directional wipes, image cropping and easier App Mode setup</summary>

The clip editor adds four directional wipes, Wipe from left, Wipe from right, Wipe from top and Wipe from bottom. They appear beside Cross dissolve in the Inspector's Transition tab, and the final render closely matches the preview.
A project or saved draft that uses a directional wipe needs 1.4.0 or later; earlier versions refuse it rather than dropping the wipe.
The final render now encodes H.264 at a fixed, higher quality setting, so flat colours match the preview more closely. Rendered files can be larger than before.
The clip editor now opens with ComfyUI frontend 1.55.11 and later, which current ComfyUI versions use. Versions of this extension up to 1.3.2 could not open it there.
The clip editor now refuses a title that uses characters the bundled fonts cannot draw, such as Chinese, Japanese, Korean or emoji, and says why as soon as you add or edit it, instead of failing at the final render. The text you typed stays in place.
App Mode can crop a first frame, last frame or reference image: use Crop image beside the image, drag the rectangle or pick an aspect ratio, then Apply. Starting or replacing the canvas adds ComfyUI's native Crop Image node, which you can still edit on the canvas.
Start H3 App Mode and Replace canvas and start H3 App Mode now work for I2VA, FL2VA, L2VA and Ref2VA before you choose a source. They place the template's Load Image nodes, and the sidebar lists which nodes still need your images before you queue.
Empty source lists explain which node to add to the main canvas, in English, Traditional Chinese and Simplified Chinese.
App Mode recognizes the installed official H3 weight variants that ComfyUI's templates select, such as their video VAE and model precision variants, and prefers the template's own choice when you have it installed.
Refine prompt now follows instructions that ask for camera moves or cuts between shots, while dialogue, on-screen text, reference labels and timing stay exact. If a suggestion cannot keep them, it says why instead of offering a changed prompt.
Optimize prompt and Refine prompt keep the required content that is already in your prompt and never add forbidden content, as set on the H3 Hard Constraint Producer node.
The GitHub repository is now named MiniMaxH3-Studio. GitHub redirects the old address, so existing clones and links keep working, and an existing installation folder can keep its name. The Comfy Registry name ComfyUI-MiniMaxH3-Studio is unchanged.

</details> <details> <summary>1.3.2 — App Mode finishes on current ComfyUI</summary>

App Mode runs finish again on current ComfyUI versions. ComfyUI now adds extra details to each saved video, which 1.3.0 did not recognize, so a finished run could stay at "Verifying the exact saved output" indefinitely.
If ComfyUI's saved-video message is missing or not recognized when the run finishes, the sidebar looks the video up in that run's ComfyUI history for up to about 16 seconds. If it still cannot be matched, the run ends with a message and Continue with native nodes instead of waiting.
The H3 Reference Registry node refuses more than three audio tracks in total, including video soundtracks, before generation starts. This matches the sidebar and the native H3 limit.

</details> <details> <summary>1.3.0 — Ordered references, original-size assembly and clearer cut reports</summary>

Ref2VA now executes linked references from the sidebar and bundled reference graphs.
Select and reorder up to 9 images, 3 videos and 3 audio references, within a total of 12 files and 3 audio tracks including video soundtracks. Move or remove individual references and see their final labels before applying the graph. At least one image or video is required.
Assemble at the segments' original size, 24 fps and 32 kHz stereo by default, with one final video encode. The previous 512 × 512, 30 fps, 48 kHz format remains selectable.
Review a report for each cut: smooth, visible change or not measured. These bounded measurements describe the cut; they do not guarantee visual or audio continuity.
A finished sequence can offer assembly when the media tools become ready later, without restarting ComfyUI. English assembly inputs now show separate segment and cut counts.

</details> <details> <summary>1.2.1 — Manual media tools setup and sequence assembly fix</summary>

The extension no longer downloads or installs FFmpeg. Settings → Media tools now opens the exact download page, shows the archive name and SHA-256, and admits the extracted bin folder after verifying both programs. Retry the action you started once the tools are selected.
A ? button beside Media tools explains setup, checksum verification and the supported build in English, Traditional Chinese and Simplified Chinese.
Tools that 1.2.0 or earlier already set up keep working without changes. Leftover files from an earlier interrupted setup are removed only when you choose the cleanup control.
Sequences built from your own ComfyUI video generations now assemble; earlier versions refused those segments because of their color tag and their millisecond-rounded length.

</details> <details> <summary>1.2.0 — Title decorations, editor updates and local perception setup</summary>

Style titles with strokes, sharp shadows and background padding. Preserve literal line breaks in preview and final rendering, and enable or reset each decoration separately.
Save accepted title styles in project files and acknowledged draft recovery. Older 1.1.0 readers refuse the new optional decoration settings rather than silently discarding them.
Configure qualified local visual or audio analysis through an operator-owned startup file, then inspect readiness in Settings → Local perception. Execution requires its supported native Windows setup; refreshing status does not launch a model or change configuration.
The timeline toggle is now called Ripple edits and keeps its existing edited-track scope.
Reduce repeated redraws of unchanged timeline, monitor and media controls during edits. Keep seeking and toolbar actions synchronized with the current editor state.

</details> <details> <summary>1.1.0 — Project files, recovery and editor improvements</summary>

Save and open editable .h3proj projects, with explicit relinking when media is missing.
Optionally retain verified generated videos and recover acknowledged drafts with data-only undo/redo after a restart. Each storage option is off by default.
Adjust clip volume, mute and fades in the Inspector, preview and final render; enjoy steadier playback across edits and cuts, reliable timeline shortcuts and clearer title-insertion feedback.
Generated-video closure now accepts Linux host paths. Editor media and durable recovery still require their supported Windows setup.
Source and Registry distributions include complete public developer inputs and exclude private records, dependency installations and caches.

</details> <details> <summary>1.0.2 — Assisted prompt writing and model setup</summary>

Set up Ollama or a remote provider, reload available models and choose an exact model; readiness is checked automatically, and reloading keeps the selection when the model is still available.
Browse provider-reported model dates, limits and lifecycle labels, with a local filter for large model lists and requests adapted to each provider's API.
Use Refine prompt to direct a revision while preserving exact dialogue, visible text, reference roles and timing. Every suggestion remains yours to review before applying.
Read and compare prompts without losing local edits, focus, text selection or scroll position.
Review typed semantic suggestions with a bounded repair attempt for invalid responses. The canvas H3 Semantic Proposal Producer can use an exact installed Ollama model while keeping historical workflows compatible.

</details> </details>

Features
Before you start
- Requirements
- Compatibility
Install and update
Quick start
The sidebar at a glance
Create a clip
Production
Clip editor
Longer videos
Canvas nodes
Assisted authoring (optional)
Privacy and storage
Known limitations
Troubleshooting
More documentation
License

Features

Inspectable prompts. Turn your intent, references, duration and exact dialogue or on-screen text into a structured plan, a prompt and a validation report. Missing information stays visible instead of being guessed.
Five task modes. Text to video, opening image, first and last frames, ending image, or a reference set of image, video and audio with explicit roles.
One-click official workflows. App Mode writes the official MiniMax H3 workflow for your mode into the active workflow tab using the H3 weights you have installed, or connects your prompt to an H3 graph you already have. Use a separate empty tab to try the official template while keeping your current workflow. Crop a first frame, last frame or reference image before the canvas is prepared.
Production workbench. Collect generated clips as segments of a project; reorder, replace or regenerate them, preview them and pick which ones to edit.
Longer videos. Plan a 4 to 60 second video as a storyboard, generate its segments one after another and join them.
Built-in editor. Arrange clips on up to eight tracks, trim and split, add text, picture and video overlays, adjust position, crop, colour, transitions and each clip's volume and fades, then render and download an MP4.
Optional assisted authoring. Ask Ollama or Gemini to suggest prompt improvements; you accept or reject every suggestion. OpenAI and Anthropic connections support model discovery and readiness checks.
Project files and optional recovery. Save accepted editable work to .h3proj files. On a qualified local host, separately opt in to metadata, verified-video retention or draft recovery. Exported projects do not embed media or execution permissions.
Three languages. English, Traditional Chinese and Simplified Chinese.

Public source and Registry packages include the test suites and their support files. See development and testing for environment setup and validation commands.

Before you start

Requirements

ComfyUI with its native MiniMax H3 nodes (MiniMaxH3ImageToVideo, MiniMaxH3ReferenceToVideo), the official video_minimax_h3_* workflow templates and a frontend that supports sidebar extensions. Python 3.10 or newer.
MiniMax H3 model weights and suitable hardware to generate video. Building and checking prompts needs no weights.
Windows x64 with 64-bit Python for the clip editor, Production import, longer-video assembly and final rendering, plus the media tools, which you provide and the extension verifies. Linux support for these features is coming in an upcoming update.

The package has no required Python dependencies and never installs or changes ComfyUI, its frontend or your models.

Compatibility

The sidebar checks what your host actually provides (the native H3 nodes, their inputs, the templates and the frontend functions it uses), not a version number. If something is missing, the message names it. The canvas nodes keep working even when the sidebar cannot load.

The prompt pipeline runs on Windows, Linux and WSL. The media features need Windows x64 for now; Linux support for them is coming in an upcoming update.
The clip editor and image crop have been tested on ComfyUI 0.41.0 with frontend 1.55.18. This is a tested combination, not a minimum version.
Longer-video generation checks the installed H3 implementation separately. If it reports an incompatibility, use a complete, compatible ComfyUI installation; installing FFmpeg or waiting does not fix it.

If you open ComfyUI through a reverse proxy, a port mapping or a host name other than localhost, see opening ComfyUI from another address.

Install and update

Install

To install manually:

Save any open ComfyUI workflows and download the videos you want to keep, then stop ComfyUI.
Clone or copy the complete repository into ComfyUI/custom_nodes/MiniMaxH3-Studio:
```
cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/MiniMaxH3-Studio.git
```
Keep both the root __init__.py and the comfyui_h3_context/ folder.
Start ComfyUI and check that the H3 Context nodes and the MiniMax H3 Studio sidebar appear.

Installation does not download models, contact a provider or run a frontend build; the browser extension comes prebuilt.

If the package is available through your Comfy Registry or ComfyUI Manager, search for ComfyUI-MiniMaxH3-Studio, or run:

comfy node install minimax-h3-studio

Update an existing installation

Save your workflows and download the videos you want to keep.
Replace the whole package with the new version. Do not mix files from two versions, and do not keep a second copy of this pack beside the first.
Restart ComfyUI and refresh the browser. If the old interface still shows, hard-refresh after checking that the host loaded the intended version. The sidebar header shows the loaded version and, when the host provides them, short sources and bundle build identifiers. A ! beside them means the loaded browser file does not match its build record, which can happen when files from two versions are mixed.

Keep the previous version for rollback. Save accepted Production/editor work as a .h3proj file and preserve its media separately before updating. A workflow save or downloaded video does not back up the editable project. Optional recovery covers only acknowledged server saves, not edits still waiting in the browser.

Media tools

The clip editor, Production import, longer-video assembly and final rendering use one exact FFmpeg build on Windows x64 with 64-bit Python: the gyan.dev full build 2026-02-26-git-6695528af6. Linux support is coming in an upcoming update.

Provide the supported tools. In Settings → Media tools, choose Open download page (the release page), download ffmpeg-2026-02-26-git-6695528af6-full_build.zip (about 235 MB), compare the archive SHA-256 shown on the card, and extract it yourself. Enter the bin folder containing both ffmpeg.exe and ffprobe.exe, then choose Use folder. The extension verifies both programs before saving the selection. Retry your original action after selecting them. The ? button beside Media tools shows these steps, including checksum commands, in your language.
Automatic discovery. A supported pair in your Python environment or on PATH is found on first use. Choose Check again after changing PATH; Advanced → Search automatically clears a saved local selection and restores discovery. An installation made by an earlier version is still discovered and reused. Earlier setup leftovers are removed only by the explicit cleanup control, which preserves an older copy until its managed replacement is verified.
Offline or managed computers. Select a locally provided supported pair with Use folder. The extension performs no media-tool download, archive extraction or executable installation.
Administrators can set H3_CONTEXT_AUTHORIZED_FFMPEG_PATH and H3_CONTEXT_AUTHORIZED_FFPROBE_PATH to the two programs in one folder; this overrides everything else. H3_CONTEXT_AUTHORIZED_MEDIA_SCRATCH_ROOT can also name the temporary media folder; it is accepted only together with both program paths.

Any other FFmpeg build, a modified file, Windows on ARM, 32-bit Python or a non-Windows host is not used, and the card says why. To check a copy yourself, compare these SHA-256 fingerprints:

ffmpeg.exe: abf5e1652dfdd3f8d7cfbb4900a4b590b7e8d192876164b7b7e9efaebe26d67f
ffprobe.exe: fb81e32ea05d77049291d9cffb0ec2677cfbe5e1ce2021bbcb9eb5abdc6ac576

FFmpeg is licensed separately under GPL-3.0-or-later and is not shipped with this package.

Quick start

Save your current ComfyUI workflow. To try the official H3 template separately, open a new empty workflow tab and make it active.
Open MiniMax H3 Studio in the ComfyUI sidebar.
On the Context page, choose a task mode, describe the video and set a duration of 4 to 15 seconds.
If the canvas already has image, video or audio nodes you want to use, choose them for that mode. In an empty tab there is nothing to choose yet; leave the sources unchosen.
Press Start H3 App Mode. The official workflow appears in the active tab. For I2VA, FL2VA and L2VA, each frame you left unchosen becomes its Load Image node from the template. For Ref2VA, leaving every reference unchosen keeps the template's Load Image nodes. They hold sample images until you set your own, and the sidebar lists the nodes that still need them. For Ref2VA, then choose the references in the pickers.
Check the workflow's settings and queue it with Apply and queue current H3 graph or ComfyUI's own Queue button.
When the video is ready it appears on the Production page. Each new run from the same workflow adds a segment to the same project.

To use your own H3 workflow, keep its tab active and follow Connect and queue current canvas. If you want to leave the current canvas untouched, choose Keep canvas and exit H3 App Mode when offered.

The sidebar at a glance

Open MiniMax H3 Studio from the ComfyUI sidebar. It has three pages:

| Page | What it is for | | ---------- | --------------------------------------------------------------------------------------------- | | Context | Choose a task mode, describe the video, pick references, review the prompt and start App Mode | | Production | Your generated clips as a project, longer-video planning, and the entry to the clip editor | | Settings | Language, media tools, recovery metadata, retained media, local-perception status and the optional assisted-authoring provider |

Settings → Language offers Automatic (ComfyUI locale), English, 繁體中文 and 简体中文.

Settings → Local perception shows whether the host's existing visual and audio setup is available. Refresh status reads its readiness; it does not install dependencies, launch a model or reread startup configuration. Local perception runs only on its supported native Windows setup; Linux support is coming in an upcoming update. Ask the host operator to configure it using the local-perception setup instructions, then restart ComfyUI after changing the startup file. These host settings are separate from assisted-authoring provider connections and cannot be edited on this page.

The prompt is built on the ComfyUI host by the same nodes you can use on the canvas, so the sidebar and the canvas never disagree.

Create a clip

The Context page walks through five stages, shown as numbered tabs: Intent / Mode, Media / Roles, Understand / Plan, Audit / Validate and Execute / Export. You can also simply fill in the request and choose Start H3 App Mode. Describe the video in Intent, which takes up to 4,096 characters.

Task modes and references

| Task mode | You provide | | --------- | --------------------------------------------------------------------- | | T2VA | Text to video: a description only | | I2VA | Image to video: a First frame source in a Load Image node | | FL2VA | First and last frame: two different images in two Load Image nodes | | L2VA | Last frame: a Last frame source in a Load Image node | | Ref2VA | Reference set: an image, a video and/or audio, each with a clear role |

Sources are Load Image, Load Video and Load Audio nodes on the main canvas, outside subgraphs. You can choose existing sources, or leave the sources unchosen and use Start H3 App Mode or Replace canvas and start H3 App Mode to place the template's loaders. A frame you leave unchosen uses the template's own Load Image node, which holds a sample image until you set yours there. If you choose both frames for FL2VA, they must be two different Load Image nodes. Set the images in those Load Image nodes before choosing Apply and queue current H3 graph. For Ref2VA, also choose the references in the pickers before queueing. The sidebar does not upload files.

For a reference set, choose up to nine images, three videos and three standalone audio files under Reference images, Reference videos and Reference audio. Add sources in order; the arrow and remove controls update their labels as you edit. The total limit is 12 files and three audio files, including video soundtracks. Audio cannot be the only reference in the sidebar; add an image or a video. Once a reference video is chosen, Reference video soundtrack appears with three options: Submit all selected videos’ own soundtracks (the default), Do not submit soundtracks and Soundtracks unavailable. The last two both leave the soundtracks out, and the first is disabled when including them would exceed the file or audio limits. Each video or audio clip should be 2–15 seconds, with at most 15 seconds per kind in total. The live counter includes submitted soundtracks.

Roles and order are never guessed: a missing or conflicting role is reported so you can fix it.

Crop frame and reference images

Choose an image source, then use Crop image beside a first frame, last frame or reference image. Drag the rectangle or its eight handles. The arrow keys move a focused rectangle by one source pixel, or ten with Shift. On a focused handle they resize by the same amount in Free mode, and in whole ratio steps with Original or a fixed aspect ratio, always within the image. Choose Free, Original or a fixed aspect ratio, then Apply. The source row shows the cropped width and height. To remove a crop, choose Reset, which returns the rectangle to the full image, then Apply.

In I2VA, the output follows the crop's aspect ratio; in FL2VA and L2VA, the output size comes from the canvas resolution setting, so choose a crop ratio that matches it.

Starting or replacing a canvas inserts one native Crop Image node per cropped source, before all its consumers. On a canvas that is already prepared, Crop image is unavailable; edit that visible node directly. Videos and audio have no crop action. Changing modes or removing an image clears its sidebar crop. A host without the native Crop Image node cannot prepare a cropped canvas.

With a crop applied, prepare the canvas with Start H3 App Mode or Replace canvas and start H3 App Mode in Intent / Mode, or reset the crop, before using Execute / Export.

Dialogue and exact text

The sidebar has no fields for dialogue or exact text. Enter them on the canvas, in the H3 Hard Constraint Producer node, and connect its hard_constraints output to the optional hard_constraints input of H3 Context Request (see Canvas nodes). Its dialogue, visible_text, required_content and forbidden_content inputs hold dialogue, on-screen text and anything that must or must not appear. They are kept word for word.

For dialogue the node also has:

dialogue_language: auto or one of Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. auto recognizes Korean, Japanese and Chinese from their scripts; other scripts stay untagged with a warning. The spoken words never change;
dialogue_speaker, as a short identity phrase such as the guide;
dialogue_delivery: on_screen for spoken on screen, or voiceover for an off-screen voice (the speaker's lips stay closed).

Duration

Enter Clip duration (seconds) as a whole number from 4 to 15 (default 5). The extension converts it to a length H3 can produce and shows the delivered seconds and frame count. Start H3 App Mode stays disabled until that is resolved; if it fails, choose Retry duration resolution. The workflow uses this one duration everywhere, so there are never two conflicting length settings.

Review the prompt

Once App Mode has prepared the graph and its Context, the numbered tabs fill in. Understand / Plan shows the plan behind the prompt. Audit / Validate holds the prompt and the validation diagnostics. Execute / Export shows the validation status and guide readiness. Fix errors before queueing; warnings point out uncertainties you may accept. You can also:

type @ in the prompt to insert a reference;
edit the prompt, give a Revision reason and choose Stage revision, then choose Validate revision to check it;
Copy prompt, Export JSON or Import JSON;
choose Optimize prompt when assisted authoring is set up, then edit, accept or reject the proposal;
open Revision instruction to request wording, style, camera moves or cuts between shots and choose Refine prompt;
use Reader for the complete prompt, or Compare with current report for an active AI proposal.

Guide readiness is a separate check against the official H3 prompt guide: Ready means every known requirement is covered, Incomplete means something the guide expects is missing, even if the prompt is valid. Neither judges visual quality.

Start H3 App Mode

Start H3 App Mode writes the official ComfyUI MiniMax H3 workflow for your task mode onto the active workflow tab, without queueing it. Save any unsaved workflow before starting. To try the official template separately, open a new empty workflow tab and make it active first. To connect your own H3 workflow instead, keep its tab active and use the Connect option below.

App Mode uses these official templates:

| Task mode | Official template | | ----------------- | --------------------------------- | | T2VA | video_minimax_h3_t2v | | I2VA, FL2VA, L2VA | the video_minimax_h3_i2v family | | Ref2VA | video_minimax_h3_r2v |

Model loaders are filled from your installed official H3 weights. When the template's own choice, such as a particular video VAE or model precision variant, is installed, it is kept; otherwise an installed official variant for that role is used. If a model cannot be matched, it is named as a warning and left for you to pick on the canvas; ComfyUI checks the models when the workflow is queued. The workflow is a normal ComfyUI graph that you can inspect and change. Queue it with Apply and queue current H3 graph or ComfyUI's own Queue button.

If the canvas already has nodes, App Mode asks first:

Replace canvas and start H3 App Mode replaces the graph in the active tab with the official workflow. Save the current workflow before choosing it;
Connect and queue current canvas keeps your graph, connects the prompt to the H3 generation node you choose and queues it once. Your models, LoRAs, sampler and other settings stay as they are;
Keep canvas and exit H3 App Mode leaves the current canvas untouched and exits App Mode, without queueing a run.

Connect adds the prompt connection and queues a run; Keep makes no canvas changes. Neither choice saves a backup of your workflow. Use ComfyUI's own workflow Save action to preserve unsaved canvas work. Saving a ComfyUI workflow does not back up the separate Production or editor project.

Connect works with your own H3 workflows, not only the official ones. It looks for the native H3 generation nodes on the canvas and one subgraph level inside it; when there are several, choose one under H3 generation node to connect. The task mode must match how that node is wired: a connected first frame means I2VA, a last frame L2VA, both FL2VA, and neither T2VA; a native reference-to-video node means Ref2VA. Connect is not offered when the canvas already contains H3 Context nodes; use Apply and queue current H3 graph instead.

If an existing canvas is replaced without a choice being shown, report the loaded sidebar version and the actions you took. Include Copy diagnostics when available so the unexpected change can be investigated.

Cancel App Mode is offered only while App Mode is still preparing the graph and the Context. Once generation is queued, use ComfyUI's own queue controls. Continue with native nodes leaves App Mode so you can work on the canvas yourself.

During and after a run

A run is queued once through the normal ComfyUI queue. It is reported as finished only after the saved video has been checked against that run. If ComfyUI's saved-video message is missing or not recognized when the run finishes, the sidebar looks the video up in that run's ComfyUI history for up to about 16 seconds. A video that still cannot be matched is reported as not verified, the saved file is left unchanged and the run is never queued again automatically. If the connection to ComfyUI drops, the sidebar picks the run up again when it reconnects.

The finished video is added to the Production project shown above the App Mode buttons ("Adds to Project …" or "Creates Project …"):

every new run from the same workflow adds one segment to the same project;
choose New project before starting when the next result should go elsewhere;
Edit App Mode setup returns to the request for the next run; Cancel edit leaves the setup and keeps the previous result unchanged.

If App Mode declines a run, it shows the reason and a remedy, with Retry H3 App Mode or Retry output verification where they apply. Copy diagnostics copies a short report for support; it never contains your prompt, workflow, file paths or credentials.

Production

The Production page has two views, switched at the top: Production for your project and Clip editor for the editor.

Project and segments

The status band names the project and its segment count, with New project and a progress meter while segments generate.
Segment overview shows each segment's share of the video. Its seconds are marked "snapped" when H3 delivers a slightly different length from the one requested.
The segment list shows 32 segments per page. Tick segments to select them; with exactly one selected, its row expands to show its state, output, continuity and boundary.
Project target duration (seconds), from 4 to 60, is the intended length of the whole video. It is used by longer-video planning, is separate from the next App Mode clip's 4–15 seconds, and never starts a generation when changed.

Arrange and generate segments

Reorder by dragging a tile in the segment overview, with the arrows on a selected segment, or by typing its Position and choosing Apply.
Add current Context adds a segment from the current Context request. With one segment selected, Replace segment N with current Context or Delete segment N (the only remaining segment cannot be deleted). A project holds up to 64 segments.
Relationship for segment N sets how a segment joins the previous one: Independent, Predecessor, Adjacent pair, Cut or Reset; the second and third need an earlier segment. Choose Apply relationship to segment N to keep it. Changing a segment only reruns what depends on it.
Generate segment N queues one segment when it is ready; Blockers says why it is not.
Release workspace closes this Production workspace after Confirm release. The Context and earlier outputs stay unchanged.

Previews and intent proposals

Preview segment N (the play button on a tile) and Preview aggregate output play a copy on the host. The aggregate preview marks where each segment begins, and its sample thumbnails jump to that point. A missing button means that output cannot be previewed yet.
Understand segment N and Understand selected open a proposal review when the segment kept an intent proposal from its original Context run. A segment without one is still valid.

To edit outputs, open the clip editor and use its Import button (see Media bin).

Clip editor

In short: in Production, tick up to three finished segments, switch to Clip editor and choose Open full editor. Bring the segments in with Import in the media bin, place them on the timeline, trim and arrange them, add titles and overlays, then choose Export → Render final video. The result is a 1920 × 1080, 24 fps MP4 that you download with Download original.

Open the editor

Switch the Production page to Clip editor and choose Open full editor. If there is no editor project yet, choose Start authoring from this context first. The editor opens as one window over ComfyUI; all editing, selection, undo and redo happen there. Close it with its close button or Escape.

The Clip editor view in the sidebar summarizes the editor project and offers Refresh workspace and Release workspace (not while the editor is open).

Layout

The window has a top bar and four regions: the Media bin on the left, the Preview monitor in the middle, the Inspector on the right and the Timeline across the bottom. The media bin has three tabs: Media, Text and Sequence (for longer videos).

Drag the dividers between regions to resize them, or focus one and use the arrow keys (Shift for bigger steps). Enter or a double-click resets a divider.
Drag the window's corner grip to resize the window, down to 720 × 480 pixels. In a small window the regions scroll.
The top bar shows the edit status: Saving…, Saved, Edit refused, Save status unknown and similar. Saved means ComfyUI accepted the edit; it is not a project file. If the status is unknown, let the editor re-read the host and follow its message before editing again.
Use Project file to save or open an editable .h3proj document. It appears in Production, in the Clip editor project summary and in the editor's top bar, with Open project, Save project and, where the browser supports a save-file picker, Download project. See Project files for media relinking and the difference between a written file and a prepared download. A project that uses a directional wipe needs 1.4.0 or later to open.

Media bin

The Media tab holds the clips available to this editor project. It starts with the reference images and videos of the Context the project was started from; pictures in the bin come from those reference images.

Import brings in the outputs selected in Production: up to three finished segments. If the selection also contains unfinished segments, the button reads Import ready (2 of 3) and imports only the finished ones. Importing adds sources to the bin; it does not place them on the timeline or start a generation. The bin holds up to 128 items.
Search media, Sort and filter (source or reverse order; all media, videos, pictures, or those already added to the timeline) and Grid view / List view help in a full bin.
Click a card to add it: a video goes to the end of the Main track, a picture goes to a Picture overlay track at the playhead. A new project has only Main, so add a picture overlay track from a track menu first (see Timeline). You can also drag a card onto a track. Clips on one track cannot overlap, so a picture added at the playhead, or a card dropped on a track, needs free space there.
Right-click a card or press Shift+F10 on it (touch screens show a More clip actions button) for Add to timeline, Insert at playhead and Overwrite at playhead. Insert at playhead moves the later clips on that track along and does not cut through a clip, so put the playhead on a cut or in a gap, or split the clip first. Overwrite at playhead replaces whatever the new clip covers.

The Text tab adds titles: enter Title text, pick a Font and choose Add title. A title holds 1 to 2,048 characters on at most 32 lines. The title is placed at the playhead on the first unlocked text track that is free there; if there is no such track, the editor creates one within the eight-track limit. A new title lasts one second; trim it like any clip to change its length. One Undo removes both the new track and its title. You can add the first title to an empty project. If insertion is unavailable, the disabled button's explanation says what to change, such as freeing a track or choosing a font.

Timeline

The timeline has the Main video track plus video, picture and text overlay tracks: up to eight tracks, 128 clips and 2.5 minutes in total.

Track headers show or hide a track (eye button) and lock or unlock it. The track name opens its menu. Pick the Kind of a new track (Video overlay, Picture overlay or Text overlay) and its Order (the position), then choose Add track; Reorder track moves the current track to the Order you set. Remove track works only on an empty track, and Main cannot be removed.
Selecting: click a clip; Shift-click to extend the selection, Ctrl-click to add or remove one, or drag across an empty lane to box-select.
Moving: drag a clip to another time or a compatible track (hold Alt to suspend snapping). With a clip focused, Enter starts a keyboard move: arrow keys move it, Alt+Up/Down changes track, Enter applies and Escape cancels.
Trimming: drag a selected clip's edge, or focus the edge and use the arrow keys, Enter and Escape. Turn on Ripple edits to make trims and deletes shift the clips that follow on the same track.
Roll edits move a shared cut between two neighbouring clips with the handle on the cut.
Clip menu (right-click a clip, or press Shift+F10 on a focused clip): one-frame trims, Replace clip asset, Slip clip, Slide clip, Merge clips with the next one, and Clip enabled. Menus close when you press elsewhere.
Toolbar: Undo, Redo, Split at playhead, Trim start to playhead, Trim end to playhead, Delete (or Ripple delete), Snapping, Ripple edits, zoom and Fit timeline. More timeline tools adds roll edits, numeric moves, selection, clip navigation and panning.
Ruler: click or drag to move the playhead. Over the tracks, Ctrl + mouse wheel zooms; Shift or Alt + wheel pans.

Each completed edit is one undo step. One edit or group move covers at most 32 clips. An edit can be refused, for example by a locked track or the source's length; read the reason before trying again. Refused edits are never replayed automatically. If the timeline changed under you, a banner offers Rebase rejected edit.

The workspace summary shows Content extent (up to the end of the last placed clip, counting every clip, including disabled clips and clips on hidden tracks) and Edit capacity (the space available). The final video covers the content extent, gaps included. An empty timeline cannot be rendered.

Keyboard shortcuts

| Action | Keys | | ---------------------------- | -------------------------- | | Undo / Redo | Ctrl+Z / Ctrl+Shift+Z | | Split at playhead | Ctrl+B | | Trim start / end to playhead | Q / W | | Delete / Ripple delete | Delete / Shift+Delete | | Select all clips | Ctrl+A | | Zoom in / out | Ctrl+= / Ctrl+- (or = / -) | | Fit timeline | Shift+Z | | Play or pause | Space | | Step one frame | , and . |

Cmd works in place of Ctrl. Edit and view shortcuts also work after clicking a clip, a timeline tool, a trim grip or a roll handle. Space and Enter keep the focused button's own action. A keyboard move, trim or roll draft keeps its keys until you commit or cancel it. Fields, value controls, menus and buttons outside the timeline keep their own keys, so typing in the Inspector never edits the timeline.

Keys used inside the editor stay within its window, including undo and redo. If the host cannot keep the editor's keys separate from ComfyUI's own shortcuts, the full editor reports that it is unavailable.

Inspector

Select a clip to edit its properties. The tabs depend on the clip:

| Tab | What it changes | | ---------- | ------------------------------------------------------------------------------- | | Basic | Scale, position, rotation, anchor and alignment; opacity and blend mode | | Crop | The four edges | | Colour | Brightness, contrast and saturation (choose Color adjustment as the effect) | | Text | Text, size (8–512 px), weight, italic, alignment, line height, colour, stroke, shadow and background | | Transition | None, Cross dissolve, Wipe from left, Wipe from right, Wipe from top or Wipe from bottom, and its length in frames (1 up to the smallest of 300, the clip's length and the room on the clip below) | | Audio | Volume (−60 to +12 dB), mute, fade in and fade out (up to 10 seconds each, and no longer than the clip together) |

The Text tab appears for titles and the Audio tab for videos with sound.

A transition belongs to the start of the selected clip and plays it in over an enabled clip on a visible lower track that is already playing there and covers the whole transition. A wipe reveals the clip across its own cropped and scaled frame. A transition changes the picture only: sound follows the Main track and is not blended. Two clips on the Main track cannot transition into each other; place the incoming clip on an overlay track above instead.

Titles preserve literal line breaks, including blank and trailing lines. Text alignment places each line inside the title canvas; clip alignment in Basic places the whole canvas. Stroke width is 1–32 px, shadow offsets are −64 to +64 px, and background padding is 0–64 px. Zero padding removes the extra padding. Each decoration has its own enable switch and Reset; changing one group preserves the other title settings. Decoration opacity is 1–255, while fill and background opacity also allow zero. Shadows use a sharp edge and follow the clip's scale and rotation.

Sliders apply when you release them. Number fields apply on Enter or the group's apply button; moving focus away does not apply them. Escape cancels a change, and Reset restores the whole group as one undo step.
Audio settings are heard only while the clip is on the Main track, because overlay videos are silent. Raising the volume of already loud audio can distort it.
With no clip selected, the Inspector shows the read-only project settings: resolution, frame rate, duration, audio and export format.

Preview monitor

The monitor previews the timeline in your browser with the main track's sound. Its transport has Previous frame, Play / Pause, Next frame, Fit picture / Actual size and Full screen. With the monitor focused, Space plays or pauses and Left/Right (or , and .) steps one frame.

On-picture handles move, scale and rotate the selected clip when one enabled clip on an enabled, unlocked track is selected and the playhead is inside it. One drag is one undo step; Escape cancels it.
Thumbnails, filmstrips and waveforms load from the source media. Waveforms show the main track's embedded sound; they do not reflect volume, mute or fades.
If the monitor shows Preview unavailable, read the reason and choose Retry.

The monitor is a preview. Render the final video to check the exact result.

Render the final video

Open Export in the top bar and choose Render final video. The host renders a 1920 × 1080, 24 fps MP4 (H.264 at a fixed high-quality setting, with 48 kHz mono AAC sound when the main track has audio) using each clip's audio settings. You can follow progress, Cancel render, Preview output and Download original. Preview output plays a smaller 640 × 360 copy, while Download original gives the full-size file. A render made before later edits is labelled Output from an earlier revision.

Titles use the bundled Noto Sans fonts, which cover Latin, Greek and Cyrillic characters only. Add title and the Inspector refuse title text or a style change that needs other characters, such as Chinese, Japanese, Korean or emoji, and say why; the text you typed stays in place. A title like this from an older project still opens and can be moved, trimmed or deleted. It cannot be rendered, and its style cannot be changed, until you edit its text to characters the fonts cover.

Rendered files are kept for about an hour, at most 16 at a time. Download the ones you want to keep.

Keeping your work

Closing the editor or switching sidebar pages does not undo accepted edits. While the page stays open, the editor also remembers its layout, selected tab, playhead and zoom.
Unapplied Inspector changes stay with their clip; if the clip or timeline changes, the old draft is cleared with a notice instead of being applied elsewhere.
A page reload can reconnect while the current editor workspace is still available; restarting ComfyUI ends that live workspace. Active, visible work renews its lease; hidden or disconnected work can expire after about 15 minutes without activity. Opt-in recovery protects tracked unsaved server drafts, but does not prove that browser-only edits reached the host.
Use Project file to save or open accepted editable data. With Save draft recovery enabled, Recovery saved confirms an exact server revision; restore it explicitly as a new project after a restart. Neither recovery nor opening a file starts generation or rendering. See Project files for limits, save acknowledgements and media relinking.
Imported clips depend on their Production project. Releasing that project or replacing a source makes the clip unavailable; select a current output and import it again.

Longer videos

A longer video, from 4 to 60 seconds, is planned in Production and then generated and joined in the clip editor's Sequence tab.

Plan in Production

Set Project target duration and a Segmentation policy: Automatic storyboard, or fixed 5, 10, 12 or 15 second segments. To keep the plan apart from your current project, Create planned project starts a new one from the current Context.
Choose Prepare planning context.
Give the plan its shots: choose Review storyboard rows, write a storyboard script, choose Split script into shots, check the rows (you can edit them or Add shot) and choose Use reviewed rows. Use generated storyboard is enough when the target equals the Context's own clip length.
Choose Propose segments. With Automatic storyboard, cuts fall on shot boundaries where 4 to 15 second segments allow it; a fixed policy cuts at its length, and the last segment must be 4 to 15 seconds long. If the proposal cannot be used, the reason is shown, for example a cut inside a line of dialogue.
Choose Approve and import to Production. This creates the segments in the project; nothing is generated yet.

Write a storyboard script

The Context describes one clip of 4 to 15 seconds: the style, subjects, sound and hard constraints every segment keeps. The storyboard script describes what happens across the whole video:

Start each shot with [Shot 1], [Shot 2] and so on, or separate shots with a blank line.
To set the cuts yourself, give every shot after the first a start time, such as [Shot 2] At 00:08.000, the courier crosses the bridge. (At 0:08, also works). Times must increase and stay within the target. Give times to all of those shots or to none.
Without times, the shots share the target duration evenly.

Example for a 60 second target:

[Shot 1] The courier leaves the depot at dawn.
[Shot 2] At 00:08.000, the courier crosses the bridge.
[Shot 3] At 00:20.000, rain starts over the market.
[Shot 4] At 00:33.000, the courier shelters under an awning.
[Shot 5] At 00:47.000, the parcel is delivered.

This gives five segments of 8, 12, 13, 14 and 13 seconds. Shots shorter than 4 seconds are grouped, a shot longer than 15 seconds is split, and a shot marked as a Hard boundary is never split (so it must fit in 15 seconds). A script holds up to 32 shots. Load shots from Context copies the shots the Context already describes as a starting point.

Exact dialogue, on-screen text and timing instructions from the Context stay in the opening part of the video, and no cut is placed inside a shot that holds exact text. Plans that use reference media can be proposed and reviewed, but generation currently accepts text-to-video plans only.

Generate and assemble in the editor

Open the clip editor's Sequence tab and choose Check readiness. It confirms that your host, models and extension can generate the plan; the result is valid for about a minute.
Choose Generate approved sequence. Segments are generated one after another. A failed segment can be retried with Retry segment N, and Cancel sequence stops the rest.
To leave while it runs, choose Continue current segment while I leave: the current segment finishes and the rest pause. Later choose Reconnect to running sequence and Resume remaining segments before the deadline shown.
Choose Output format before assembly: Original size (24 fps) is the default, with 32 kHz stereo sound and the segments' common original size. 512 × 512 (30 fps) keeps the previous format with 48 kHz stereo sound. Then choose Assemble sequence. While an assembly is running, Cancel assembly stops it; after a failed assembly, choose Retry assembly.

The original-size format requires all segments to share one supported size; incompatible sizes are refused rather than resized. It keeps the native rates and encodes the final video once. After assembly, Cuts between segments labels each cut smooth, visible change or not measured. A label is a bounded measurement, not a guarantee of continuity or generation quality. A sequence that finishes before the media tools are ready can offer assembly on a later Production refresh once the tools become available.

Assembly does not apply editor overlays or effects. To edit the result, import the segments into the media bin, place them on the timeline and render the final video there.

Canvas nodes

Every sidebar feature is built on canvas nodes you can also use directly. The smallest route is:

Request -> Plan -> Compiler -> Validator -> Preview

H3 Context Request holds your intent, task mode, duration and hard constraints. H3 Context Plan and H3 Context Compiler turn it into a plan and the final prompt. H3 Context Validator reports problems without rewriting anything, and H3 Context Preview shows the result. Add H3 Reference Registry for reference media, and H3 Native MiniMax H3 Adapter to connect the prompt to ComfyUI's H3 nodes. Media always stays on its own connections.

All 28 nodes, grouped by purpose:

| Group | Nodes | | ---------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Core prompt pipeline | Context Request, Reference Registry, Context Plan, Full-Reference Plan, Context Compiler, Context Validator, Context Preview, Context Audit Override | | Detailed planning | Hard Constraint Producer, Intent Graph Producer, Evidence Fusion Producer, Cross Reference Producer, Directive Authority Producer, Full Reference Timeline Producer, Feasible AV Timeline, Hierarchical Evidence Reduction, Constrained Semantic Planning, Source-Profiled Renderer | | Media and perception | Media Admission Producer, Visual Perception Producer, Audio Perception Producer | | Generation and sidebar | Native MiniMax H3 Adapter, Product Shell Boundary, Local Reconstruction Acceptance | | Providers and status | Provider Transparency, Reliability Status, Semantic Proposal Producer, Official Context-IR |

A few nodes worth knowing:

H3 Context Audit Override records a deliberate decision to proceed with a prompt edit while keeping the original diagnostics visible.
H3 Intent Graph Producer has a complete_silence option. Turn it on only when silence is the intended sound; an empty soundscape otherwise counts as unspecified.
H3 Provider Transparency and H3 Reliability Status only display information; they never call a provider or start, retry or cancel work.
H3 Visual Perception Producer and H3 Audio Perception Producer describe short media locally after a host operator sets them up (see host integration). Without that setup, perception reports unavailable; describe the fact yourself or enter it as a hard constraint.
H3 Official Context-IR needs a connection supplied by the host and otherwise refuses to run.

Bundled files:

subgraphs/: the reusable H3 Context Assistant (base and reference) and H3 Product Shell Boundary graphs.
workflows/: example graphs that need no models, in ComfyUI's API format (each file's prompt section), covering the base, reference and full-reference routes and the other node groups. Use the current bundled reference examples when upgrading. In an older saved API prompt, connect each reference to its own numbered Registry input instead of placing several links inside one input value.
examples/minimal_core_pipeline.py: a pure-Python example that needs neither ComfyUI nor a provider.

None of these load model weights or promise a particular result.

Assisted authoring (optional)

Assisted authoring is off by default, and the normal path needs no provider; None selected is the default. It powers Optimize prompt and Refine prompt on the Context page. Set it up in Settings → Assisted authoring provider:

| Profile | Runs on | | ------------------------------------ | ----------------------------------- | | ollama.local | Your own Ollama on the ComfyUI host | | anthropic.remote | Anthropic | | openai.remote | OpenAI | | gemini.remote | Google Gemini |

Prompt generation is enabled for Ollama and Gemini in this release. OpenAI and Anthropic support connection setup, model discovery and readiness checks; prompt generation is currently unavailable for those connections.

For Ollama:

Choose Ollama, then Reload available models.
Pick the Exact model. Its readiness check runs automatically.

For Anthropic, OpenAI or Gemini:

Choose the provider and read What leaves this computer.
Enter your API key and choose Allow text-only requests and reload models.
Pick the Exact model. Its readiness check runs automatically; switching models within the same connection keeps the permission you granted.

Use Optimize prompt or Refine prompt when the selected connection is ready and assisted execution is available. A reachable provider may still show that execution is unavailable. Withdraw consent removes the permission and Discard removes the API key.

The model list uses the provider's exact identifiers. Large lists offer a local filter; model dates, limits and lifecycle labels appear only when the provider reports them. Reload available models keeps the selected model and refreshes its readiness if it is still present. If the model disappears, the selection and readiness are cleared; choose a model from the current list.

Refine prompt takes a revision instruction of up to 2,048 characters and 8,192 UTF-8 bytes. It can revise wording and style, or add camera moves and cuts between shots. Protected dialogue, visible text, reference labels, reference roles and timing stay exact. Optimize and Refine retain required content already present and never introduce forbidden content that was absent. An explicit static-camera, no-cuts or single-shot restriction still wins, and added cuts must fall within the clip's duration. Change typed facts through their existing controls. Stage an unstaged prompt edit first, and resolve any active proposal before asking for another suggestion. Your instruction stays in browser memory for the current workspace.

Reader retains your text, selection and scrolling when you return to the editor. Compare with current report shows the report and the active proposal separately and retains newer local edits. These views do not send requests or accept a proposal. Copy and export continue to use the current report; viewing a candidate does not make it the accepted prompt.

Only prompt text, revision instructions and derived text are sent, never media. A suggestion is shown for you to accept, edit or reject; nothing is applied automatically, and a failure never switches to another provider.
Switching providers or replacing credentials requires permission for the new setup. Sessions expire after 30 minutes without use or eight hours in total, and a ComfyUI restart clears them.
Billing, quota and data retention are between you and your provider.

The canvas H3 Semantic Proposal Producer is configured separately. Choose ollama.local and enter an exact installed Ollama model identifier in ollama_model. It checks the model's current native identity and completion capability before generating, and finishes cleanup before a proposal can be reviewed. The previous exact-model profile remains available for historical workflows.

Semantic suggestions have separate subject, scene, action, camera, style and audio fields, grounded in the facts you supplied. Exact dialogue and text bindings remain protected. An invalid semantic response can receive one bounded repair attempt; a repaired suggestion still needs your review and is never applied automatically.

See Security and providers for what is sent and checked.

Privacy and storage

Prompt building, previews and final rendering happen on the ComfyUI host and in your browser. If ComfyUI runs on another computer, your prompts and media are processed there too, and the sidebar adds no sign-in of its own. No feature sends media to an assisted-authoring provider, and credentials are never saved into workflows, reports or diagnostics. The optional H3 Official Context-IR node sends its prompt, and any connected media as public HTTPS URLs, to MiniMax only when the host supplies a connection.

| What | Where and for how long | | ------------------------------------ | ------------------------------------------------------------------------------------- | | Copies of verified generated videos | ComfyUI's private user data; volatile process-scoped storage, not saved projects | | Rendered final videos | The media tools' temporary folder; about one hour, at most 16 | | App Mode diagnostics | Browser storage, 32 KiB across the four latest runs; codes only, no prompts or media | | Sequence reconnect pointer | Browser storage until it expires | | Production and editor workspace contents | Host memory; cleared on restart or after about 15 minutes without workspace activity | | Project and provider session pointers | Tab session storage, cleared when the tab closes; these are not saved projects | | Provider credentials and permission | Host memory only; 30 minutes idle or eight hours, cleared on restart | | Optional workspace metadata | Private host storage; enabled separately with Keep recovery metadata in Settings; no prompts or media | | Optional retained videos | Verified copies in private host storage; enabled with Keep verified media in Settings, then created with Retain output on the Production page for a finished segment; up to 16 videos and 256 MiB, and copies that nothing references expire after seven days | | Optional draft recovery | Private host snapshots of acknowledged editable data and bounded undo/redo; enabled with Save draft recovery in the Project file controls; closed recoveries are kept for seven days | | Exported .h3proj files | The file you save or download; includes editable text, not media bytes | | Image crop | The crop dialog reads the chosen image from the ComfyUI host into your browser; App Mode writes no image file, and ComfyUI's Crop Image node may save a preview of the cropped picture in its temporary folder |

Verified generated-video copies use a random process folder under private user data, outside ComfyUI's served input, output and temporary directories. The original output stays unchanged. The live store allows at most 64 copies and 1 GiB, with a seven-day expiry. After a restart, the next generated-copy operation performs bounded cleanup of recognizable folders whose owner has exited; busy or unknown folders are retained. This storage is not durable recovery, and old preview handles do not survive a restart. Download originals you need to keep, or explicitly retain a verified copy with Retain output after turning on Keep verified media under Settings → Retained media on a qualified recovery host.

Metadata, video retention and draft recovery are three separate, default-off choices. Durable storage requires a trusted single-user Windows host on loopback, fixed local NTFS storage, no permissive CORS and no configured public-origin override. It is not available through a shared or proxied host; Linux support is coming in an upcoming update. Recovery can contain private prompts and titles; restoring it is always explicit. Only a saved acknowledgement covers the named revision. It is not a zero-loss, power-loss or cloud-backup guarantee. See Project files.

Thumbnails, filmstrips and waveforms are made from your media and can reveal private content. Review workflows and screenshots before sharing them. More detail is in Security and providers.

Known limitations

Longer videos generate text-to-video plans only. Readiness is valid for about a minute. After a page reload, a running sequence can be viewed but not resumed: cancel it and start again.
Audio as the only reference is not offered.
Image crop works on images from 32 to 16,384 pixels per side, and a crop is at least 32 pixels wide and high.
Media features (clip editor, import, assembly, rendering) need Windows x64 and the media tools. Linux support is coming in an upcoming update.
Audio in the editor comes from the main track's video only, with each clip's volume, mute and fades. Overlay videos are silent and there are no separate audio tracks.
Frame rates: assembly defaults to the original-size 24 fps / 32 kHz stereo format; the selectable 512 × 512 format uses 30 fps / 48 kHz stereo. The editor's final render remains 1920 × 1080 at 24 fps, up to 2.5 minutes.
Project media and recovery: .h3proj files do not embed media. Durable metadata, retained videos and draft recovery are opt-in and require the qualified local Windows setup (Linux support is coming in an upcoming update); unacknowledged edits can be lost. Rendered files remain temporary and are kept for about an hour.
Previews in the monitor are approximate; check the rendered file.
Title characters: the bundled Noto Sans fonts cover Latin, Greek and Cyrillic characters only. The editor refuses titles with other characters, such as Chinese, Japanese, Korean or emoji, and such a title from an older project cannot be rendered into the final video.
Quality: validation and guide readiness check the prompt, not the visual result, and do not make this equivalent to MiniMax's own implementation.
Local perception requires the qualified native Windows setup described in the host integration guide; Linux support is coming in an upcoming update. An unavailable status does not install a model or fall back to a provider.
Visual timeline: the H3 Full Reference Timeline Producer needs a connected, qualified result from the Visual Perception Producer and an operator-configured local visual profile. Its timeline remains partial; without that setup, write prompts with the Plan and Compiler nodes.
Official Context-IR needs a host-supplied connection.

Troubleshooting

Setup and App Mode

| Problem | What to check | | ---------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | H3 Context nodes are missing | The complete repository must be directly under custom_nodes. Restart ComfyUI once. | | The sidebar is missing | The host must provide the native H3 nodes and a frontend that supports sidebar extensions. The canvas nodes still work. | | An old interface remains after updating | Make sure only one complete version is installed, restart ComfyUI and hard-refresh the browser. | | I want to try App Mode without replacing my workflow | Save the workflow, then make a new empty workflow tab active. To leave the existing canvas untouched, choose Keep canvas and exit H3 App Mode. | | I want to use the prompt with my existing H3 workflow | Keep its tab active and choose Connect and queue current canvas when offered. This connects the prompt and queues once; your models, LoRAs and sampler settings remain. | | Start H3 App Mode stays disabled | The duration has not resolved yet, or a reference selection exceeds its limits. Queueing Ref2VA still needs a chosen image or video reference. A notice above the button names any other blocker, such as an unsupported host or template. | | Crop image is unavailable | Crop works when preparing a new canvas, on a host with ComfyUI's native Crop Image node. On a canvas that is already prepared, edit its Crop Image node. | | Execute / Export asks you to prepare the canvas | A crop is applied. In Intent / Mode, choose Start H3 App Mode or Replace canvas and start H3 App Mode, or reset the crop. | | A model role is reported missing | Install an official MiniMax H3 weight for that role (any subfolder works) and pick it on the canvas loader. | | App Mode refuses a run | Follow the reason shown. Copy diagnostics gives a report for support with private details left out. | | A request is rejected | Read the first diagnostic and supply the named input, or fix the mode or duration conflict. | | Validation passes but Guide readiness is Incomplete | Add the missing information named in the reason. For intended silence, enable complete_silence on H3 Intent Graph Producer. | | Perception reports unavailable | No perception was set up for that fact; enter it yourself or leave it unknown. | | Actions fail when ComfyUI is opened through a proxy or host name | The ComfyUI log shows H3 Context refused a browser request. Add the page's exact origin to H3_CONTEXT_PUBLIC_ORIGINS and restart, or open the page at the address ComfyUI serves. |

Media, editor and longer videos

| Problem | What to check | | --------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | | A media feature says it is unavailable | Open Settings → Media tools and follow the card: Use folder, Check again, or the reason shown. | | Import is missing or does nothing | Select up to three finished segments in Production first. | | An imported clip is in the bin but not on the timeline | Import only adds sources; click its card. | | An imported clip became unavailable | Keep its Production project available, or open a saved project and explicitly relink compatible retained media. A released source is never silently replaced. | | Add title is unavailable | Read the reason beside the button. Choose an available font, shorten the text, or free a text track at the playhead when all eight track slots are in use. | | The final render says the title font cannot render some characters | Edit the title text to Latin, Greek or Cyrillic characters, or delete the title, then render again. | | Add title says some characters are not supported | Use Latin, Greek or Cyrillic characters. The bundled title font does not include others, such as Chinese, Japanese, Korean or emoji. | | A number changed but the picture did not | Number fields apply on Enter or the group's apply button. | | On-picture handles are missing | Select one clip, put the playhead inside it, and make sure the clip and track are enabled and unlocked. | | The Audio tab is missing or has no effect | It appears only for a video with sound, and is heard only on the Main track. | | A keyboard shortcut does nothing | Finish or cancel a keyboard edit draft. Fields, menus and buttons outside the timeline keep their own keys. | | Thumbnails or waveforms stay on Loading | Check the media tools and that the source is still available. Importing the same source again does not help. | | The monitor shows Preview unavailable | Read the reason and choose Retry. | | An edit is refused | Read the reason (source length, locked track or placement) before trying again. | | The final video is longer than expected | Check Content extent, including disabled clips and clips on hidden tracks at the end. | | The render is labelled as an earlier revision | You edited after rendering; render again. | | A sequence cannot be resumed after a reload | Cancel it and start again. | | Readiness expired before generating | Choose Check readiness again. | | Readiness says the host does not provide every native MiniMax H3 workflow | The loaded native H3 implementation is incomplete or not supported; use a complete, compatible ComfyUI installation. | | A remote assisted-authoring profile does not run | Check the message, profile, permission, API key and your provider quota. |

When reporting a problem, include the node name, the diagnostic code and the profile or contract version. Remove credentials, URLs, private paths, media and provider responses from logs and screenshots.

License

Apache-2.0; see LICENSE and NOTICE. The bundled Noto Sans fonts use the SIL Open Font License 1.1. The compiled sidebar includes React, ReactDOM and Scheduler, which are licensed under the MIT License; their notice is in NOTICE. FFmpeg is not included: you provide the supported build yourself, and it is licensed separately under GPL-3.0-or-later. The official workflow templates are served by the ComfyUI frontend (Comfy-Org workflow_templates, MIT) and are not redistributed here. Models and provider services have their own terms. This project is not affiliated with or endorsed by MiniMax.