H3 Prompt IDE

<p align="center"> <img src="assets/github-social-preview.svg" alt="H3 Prompt IDE — rich H3 prompt authoring for ComfyUI" width="100%"> </p>

H3 Prompt IDE for ComfyUI

A VS Code-inspired prompt editor for MiniMax H3. It provides rich editing, strict H3 sections, completions, reference previews, and a normal ComfyUI STRING output.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-H3-Prompt-IDE.git

Restart ComfyUI, then find the nodes under text → H3 Prompt IDE.

Use

1. Add H3 Reference Inputs and H3 Prompt IDE.
2. Connect images, video frame batches, or audio to the matching H3-labelled sockets.
3. Connect references to the Prompt IDE.
4. Connect the IDE's text output to your H3 prompt input.

H3 Reference Inputs handles labels and previews only. For Ref2VA generation, connect the same media loaders to the matching inputs on the native H3 conditioner. Video sockets use ComfyUI IMAGE frame batches, matching the native H3 ref_videos inputs.

MiniMax H3 Edit integration

Connect the IDE's text output directly to the prompt input on TextEncodeH3Edit. The schema selector automatically shows an Edit category and follows the encoder's current task: edit instruction, re-pose, character swap, new camera angle, character-sheet assignment, continuous frozen-scene coverage, or cinematic frozen-scene cuts. Changing the task/profile on the Edit encoder updates the IDE context live.

With MiniMax H3 Edit 0.10.0 or later, connect H3 Edit Options to the encoder and choose one canonical mode. Prompt IDE reads that Options node directly, so changing only its mode updates the visible instruction; the encoder's hidden legacy task/profile values cannot leave the IDE in the wrong category. The Room + object study preset requires MiniMax H3 Edit 0.11.0 or later.

For frozen-scene coverage, use the IDE instruction to identify the room or scene, declare the orbit center, and assign each optional <Picture N> as an alternate angle or room-design reference. The connected encoder writes the exact multi-view timing, world-freeze contract, and first/final keyframe alignment. Semantic generation mode can therefore describe a completely new room without turning the IDE text into a full Ref2VA prompt manually.

For cinematic cuts, name one exact visible person, object, architectural feature, or fixed point as the coverage target. The Edit encoder turns that short instruction into eight timed static shots with true hard cuts and no intermediate camera travel.

For Room + object study, treat the connected pictures as equal, complementary Qwen-only survey views of one physical room and name one exact target object. The legacy source_image socket only assigns the <Picture 1> ordinal in this preset; it does not make that picture a master, source frame, pixel/composition anchor, or VAE keyframe. The Edit encoder selects the 362-frame semantic reconstruction preset automatically and compiles six room-establishing shots plus 10 contextual object-study shots. Each object view retains the complete object, its floor/rug contact, and room anchors while locking its orientation; background parallax shows that only the camera changed position.

When a directed task or character-sheet profile is selected, the IDE loads a short task-specific starting instruction. Empty text and the previous generated template update automatically after a live task change; if the editor contains custom text, the IDE asks before replacing it. Loading or switching workflows only restores the detected task context and never opens a replacement confirmation. Task template can restore the current task's starting instruction at any time. Edit that short instruction directly—the downstream Edit encoder adds its full H3 task/timing wrapper. When the encoder uses use prompt verbatim, the IDE returns to the manually selected full H3 schema and does not offer a task template.

Nodes

• H3 Prompt IDE — rich editor with a Plain/Rich source toggle, T2VA, I2VA, FL2VA, L2VA, Ref2VA, and downstream Edit-instruction validation; H3-aware completion with automatic reference-label spacing; click-to-replace or delete reference and speaker tokens; contextual insertion and Ctrl/Cmd-click replacement for unstyled shot, language, task-directive, and retention_analysis markers; section insertion; diagnostics; and plain-text output.
• H3 Reference Inputs — independently autogrows <Picture 1>–<Picture 9>, <Video 1>–<Video 3>, and <Audio 1>–<Audio 6> while keeping media sockets off the editor node.

ComfyUI settings

The H3 Prompt IDE category in ComfyUI Settings controls the default Rich/Plain presentation for nodes without a saved per-node choice, automatic suggestions, a preferred dialogue language, trailing spaces after completed symbols, and marker-replacement interactions. When a preferred language is set, the paired <d>[Language] …</d> span becomes the first matching dialogue completion; leave it blank to keep the plain span first. Ctrl/Cmd+Space remains available when automatic suggestions are disabled. Native browser text undo and redo are preserved without allowing ComfyUI's workflow history to consume the same shortcut while either editor presentation is focused.

Reference associations use H3's native <Picture 1>, <Video 1>, and <Audio 1> naming. Audio labels must follow native presentation order: connected video soundtracks first, then standalone audio. This standalone editor intentionally does not include Motion Context @tag or @@@@tag compilation.

The editor also completes, highlights, and validates MiniMax H3's exact case-sensitive tokenizer tokens: <d>, </d>, <|cutoff|>, <|lyrics_start|>, <|lyrics_end|>, <|caption_start|>, and <|caption_end|>. Correct single-token encoding requires a ComfyUI build with PR #15808 or its merged equivalent. Existing <cutoff> text remains readable, but the IDE recommends the tokenizer-native <|cutoff|> form.

Requires a current ComfyUI version with the V3 node API and Autogrow support.

The editor includes a scoped workaround for the recurring LiteGraph widget-width bug, so it resizes correctly without a separate repair node.

Credits

H3 Prompt IDE is a standalone adaptation of the Rich Scene Prompt Editor originally developed for ComfyUI-MiniMaxH3-Contex-Loop. The Motion Context-specific plan, history, optimizer, @tag, and @@@@tag integration was removed here to provide a normal text-output node.

The rich reference presentation, media miniatures, quick-reference interaction, and compact authoring layout were inspired by ComfyUI-MiniMaxH3-Easy by nkxx188. The standalone H3 schema, graph-aware native reference labels, diagnostics, and plain STRING interface are implemented in this project.

The scoped widget-width compatibility repair is adapted with permission from ComfyUI-LegacyWidgetWidthFix by pekkAi-dev.

License

GNU GPL v3

Embedding the editor

Other local extensions can mount the same editor in a dialog or inspector. After ComfyUI loads extensions, detect the versioned API without importing and registering the main extension again:

const createEmbeddedPromptEditor = globalThis[Symbol.for("H3PromptIDE.embeddedEditor.v1")];
const editor = createEmbeddedPromptEditor(container, {
    value: promptText,
    onChange: text => markDirty(text),
    references: [{kind: "picture", ordinal: 1, token: "<Picture 1>", preview: imageUrl}],
});
// Read editor.getValue() when saving; call editor.destroy() when closing.

The host owns file persistence. setValue(text) and setReferences(records) refresh the editor without creating graph nodes. Consumers should detect this optional API and offer a basic text editor when it is unavailable.