| # Archived example workflows |
|
|
| The complete maintained 0.5 catalog is preserved unchanged in [`0.5/`](0.5/). |
| Those files retain their historical node types, positional widget values, |
| frontend metadata, prompts, and asset bindings for exact reproduction; they |
| are not recommended for a new 0.6 project. |
|
|
| The additional files directly in this Archive are still older mixed and |
| experimental examples retained for compatibility and reference. The clean 0.6 |
| workflow catalog lives one level above this folder, beginning with |
| [`T2V Normal - MiniMax H3 0.6.json`](<../T2V Normal - MiniMax H3 0.6.json>) and |
| [`T2V Studio - MiniMax H3 0.6.json`](<../T2V Studio - MiniMax H3 0.6.json>). |
|
|
| `Ref2V Legacy Scheduled - MiniMax H3.json` and |
| `Ref2V Studio Legacy Scheduled - MiniMax H3.json` preserve the former active |
| Ref2V examples with explicit numeric scene selectors. The recommended Ref2V |
| examples now use prompt-driven Tagged nodes one level above this folder. |
|
|
| ## V2 demos: choose single-image I2VA, core FL2VA, or Scheduled Ref2VA |
|
|
| [`Looping Single Image I2VA 20s V2 - MiniMax H3.json`](<Looping Single Image I2VA 20s V2 - MiniMax H3.json>) |
| is the simple long-form image-to-video starting point. It uses one opening |
| image, no last-frame input, and two requested 10-second scenes. The included |
| **First-Scene Image Gate** passes `<Picture 1>` to ComfyUI's stock |
| `MiniMaxH3ImageToVideo` only for scene 1; later recursive scenes receive no |
| first-frame keyframe and continue from saved H3 Motion Context. The prompt |
| editor marks Picture 1 active for scene 1 and inactive for scene 2. |
|
|
| H3 rounds each 10-second request up to 243 raw frames. With the example's |
| five-frame head overlap, the final delivery is `243 + (243 - 5) = 481` frames, |
| or 20.04 seconds at 24 fps. Duplicate the second Plan scene to extend the |
| chain; each additional requested 10-second continuation contributes 238 frames |
| (about 9.92 seconds). Replace the opening image and both generic motion prompts |
| before queueing. |
|
|
| [`Looping Core FL2VA V2 - MiniMax H3.json`](<Looping Core FL2VA V2 - MiniMax H3.json>) |
| is the scheduler-free starting point. It uses ComfyUI's stock |
| `MiniMaxH3ImageToVideo` with a deliberately one-scene, 124-frame plan so one |
| first/last keyframe pair is applied exactly once. The prompt follows H3's |
| FL2VA format and the Scene Prompt Editor previews both connected native Picture |
| labels. Disconnect the last image for a one-scene I2VA render, the first for |
| L2VA, or both for T2VA. For a multi-scene I2VA chain use the dedicated example |
| above; a globally connected keyframe would otherwise constrain every recursive |
| scene. |
|
|
| [`Looping Seamless Chain V2 - Scheduled Refs - MiniMax H3.json`](<Looping Seamless Chain V2 - Scheduled Refs - MiniMax H3.json>) |
| is the full fourteen-scene Ref2VA demonstration. It includes the large Scene |
| Prompt Editor, reference hover previews, Review Gate, muted recovery assembly, |
| and date/version-safe final filenames. Its schedule exercises every media |
| route: |
|
|
| - `@hero_face` is active only in scenes 1–7; |
| - `@hero_look` is always active and therefore renumbers from `<Picture 2>` to |
| `<Picture 1>` beginning in scene 8; |
| - `@performance` and paired `@performance_audio` are active in scenes 4–6; |
| - the frame-exact `@song` slice is always active, moving from `<Audio 1>` to |
| `<Audio 2>` while the paired soundtrack is present and back afterward. |
|
|
| Both workflows contain in-canvas operating notes. Replace their placeholder |
| media filenames and model selections before queueing. The older global-reference |
| workflow remains available unchanged for compatibility and comparison. |
|
|
| ## Scheduled Ref2VA wiring |
|
|
| Version 0.3.10 adds a typed reference chain which can replace the stock |
| **MiniMax H3 Reference to Video** node inside the primary loop workflow: |
|
|
| ```text |
| Picture Ref (scenes 1,3,5:8) ─→ Video Ref (scenes 1:4) |
| ─→ Audio Ref (scene 3) ─→ Scheduled Ref2VA |
| |
| Current Shot ─ prompt, clip_index, clip_count, width, height, length ─────↗ |
| ``` |
|
|
| You may write stable aliases such as `@hero_face` in Plan scene prompts. They |
| are optional authoring conveniences rather than required H3 syntax. Scheduled |
| Ref2VA compiles them to the active scene's exact `<Picture N>`, `<Video N>`, |
| and `<Audio N>` numbering before expanding to ComfyUI's stock Ref2VA node. |
| Write all subject, video, and audio definitions directly in the Plan/Prompt |
| Editor; schedule nodes never inject hidden prompt text. |
| Video-paired soundtracks remain paired on the same dynamic index and receive a |
| separate tag (blank `audio_tag` derives `@<video_tag>_audio`). A scene may have |
| no active references; it still expands through the stock node with no dynamic |
| reference sockets. |
|
|
| To adapt the bundled global-reference loop without rebuilding its links by |
| hand, right-click its core **MiniMax H3 Reference to Video** node and choose |
| **Convert to MiniMax H3 Scheduled Ref2VA**. Existing connected reference |
| sockets become all-scene schedule entries; narrow their `scenes` fields and |
| replace fixed native labels in Plan prompts with the generated `@tags`. |
|
|
| ## Three-Angle Guitar Ref2VA - Experimental - MiniMax H3 |
|
|
| A one-pass performance re-filming experiment, rather than a recursive loop. |
| Core **Load Video** opens `3ClbaJYWVO4_000030.mp4`; **Reference Video Prep** |
| samples it at H3's 24 fps, selects 209 valid frames (8.708 seconds), and copies |
| the matching source-audio samples without padding or time-stretching. Stock |
| Ref2VA receives the synchronized picture/audio pair and a source-specific |
| three-angle prompt. At export, the original waveform replaces generated audio. |
|
|
| The prompt preserves the visible performer, plaid shirt, cream Telecaster-style |
| guitar, hand choreography, and musical timing while deliberately removing the |
| source product card, website watermark, text, and split-screen layout. Treat |
| this as experimental and select model paths available in your installation. |
|
|
| ## Extend Existing Video Model Workflow - Experimental - MiniMax H3 |
|
|
| A compact two-scene model for extending an existing MP4. Core **Load Video** |
| connects its native `VIDEO` directly to **MiniMax H3 Existing Video Context**. |
| VHS and other loaders can instead use the adapter's separate `IMAGE`, `AUDIO`, |
| and `source_fps` inputs. Scene 1 continues from the imported tail, generated |
| audio can inherit its ending, and `prepend_original` places the normalized |
| source before the generated extension. |
|
|
| The Review Gate is fully wired between **Segment Save** and **Loop End**, with |
| frame-locked preview audio from Loop Trim. Its recovery branch is muted by |
| default. Select your own source video and model files before queueing. This is a |
| new standalone example; none of the earlier workflow JSON files were changed. |
| Treat it as experimental until the imported video/audio continuation path has |
| received broader testing across source codecs, frame rates, and H3 setups. |
|
|
| ## Looping Seamless Chain Global Refs Example - MiniMax H3 |
|
|
| Disk-backed recursive Ref2VA chain using the visual H3 Chain Plan editor, |
| global character references, a frame-exact source-song timeline, per-segment |
| checkpointing, interruption resume, and final assembly. The recovery branch is |
| muted by default and can assemble an already completed chain without sampling |
| the last scene again. This is the primary workflow for |
| `ComfyUI-MiniMaxH3-Contex-Loop` and uses the uniquely named |
| `MiniMaxH3LoopTrim`, so it can run while NikoDemon80's upstream Motion Context |
| pack is installed. |
|
|
| Replace the supplied image/audio filenames and model selections with files |
| available in your ComfyUI installation. Scene-count and duration labels are |
| intentionally generic because both are controlled by the editable plan. |
|
|
| ## Seamless Chain Global Refs 6 Clips - MiniMax H3 |
|
|
| This is also a historical manual workflow rather than the recursive loop demo. |
| It is a six-clip Ref2VA chain with global character-reference images, 39-frame video |
| and timeline-audio context, optional full previous-clip audio references, and |
| sequential clip bypass controls. |
|
|
| Workflow and the underlying Ref2VA multi-reference/audio compatibility patch |
| were contributed by **seitanism** in the Banodoco MiniMax H3 |
| seamless-extension thread: [original patch](https://discord.com/channels/1076117621407223829/1535700117452226560/1535771676158206032) |
| and [original workflow](https://discord.com/channels/1076117621407223829/1535700117452226560/1535771814452793474), |
| shared on 2026-08-08. Its original Motion Context node ids resolve through |
| NikoDemon80's upstream pack. Do not run the separately posted global patch |
| script alongside either marker-gated custom-node implementation. |
|
|
| Extra custom nodes used by the demo: |
|
|
| - [rgthree-comfy](https://github.com/rgthree/rgthree-comfy) for group controls |
| and `Any Switch`. |
| - [ComfyUI-VideoHelperSuite](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite) |
| for preview/final video combining. |
|
|
| The workflow's optional full-audio-reference section is off by default. Keep |
| it off for the baseline test because a full Ref2VA audio reference can make |
| music restart or replay; Motion Context's 39-frame timeline-audio path remains |
| enabled independently. |
|
|
