Download
1 variant available
2
8
27
MiniMax H3 is licensed by MiniMax under the MiniMax H3 Community License Agreement. That agreement’s Applicable Territory excludes the European Union, the United Kingdom, the Republic of Korea and the United States of America. Your use of H3 and of any H3 derivative is subject to that agreement and its Acceptable Use Policy.
MiniMax H3
# MiniMax H3 Motion Director: A whole video project in one node (v1.8.4) Updated September 18th 2026
One ComfyUI node that turns MiniMax H3 from "generate a clip" into "produce a video."
Build a timeline of shots, mix generation methods segment by segment, carry visual and generated-audio context across the cuts, rerun only the shots that need fixing, refine the result, and export the finished video — without your graph turning into a wall of duplicated nodes.
Required for this workflow Minimax H3 Motion Director Fork
This fork: https://github.com/The5toryofthe5ecret5tar5ystem/ComfyUI-MiniMax-H3-Motion-Director
Full Motion Director node and workflow with ref2va/t2v/i2v/v2v, Character Replace with sam3 negative image masking, RefMod character safetensors, prompt enhance using local or API LLM with prompt guides to help the model make more usable minimax h3 prompts, many different post process options for upscale, face refine, and audio room, per segment resume and new seed regenerate allowing you to restart at the segment you want.
## Where this came from
This is a maintained fork of [j955229's MiniMax H3 Motion Director](https://github.com/j955229/ComfyUI-MiniMax-H3-Motion-Director). Full credit for the original architecture and the upstream feature set goes to the upstream author — everything below is built on top of their work, not instead of it.
The project is deliberately integrated rather than pretending everything was invented from scratch. Upstream contains and adapts work from:
- [AIMixer / ComfyUI_MiniMaxH3_Director](https://github.com/AIMixer/ComfyUI_MiniMaxH3_Director) — Apache-2.0
- [NikoDemon80 / ComfyUI-H3-Motion-Context](https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context) — GPL-3.0
- [Carasibana / ComfyUI-H3-FaceRefine](https://github.com/Carasibana/ComfyUI-H3-FaceRefine) — MIT
- [Kijai / ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes) — GPL-3.0
Thanks to all of them. Full attribution lives in the repo's NOTICE, LICENSE and LICENSES/.
License: GPL-3.0
---
## Install
Currently not on comfyui manager. Will add later on once further work is done.
Manual:
cd ComfyUI/custom_nodes
git clone https://github.com/The5toryofthe5ecret5tar5ystem/ComfyUI-MiniMax-H3-Motion-Director.git
cd ComfyUI-MiniMax-H3-Motion-Director
python -m pip install -r requirements.txt
Masking requires yolo and sams
This requires face_yolov8m.pt in ComfyUI/models/ultralytics/bbox/
https://huggingface.co/Bingsu/adetailer/blob/main/face_yolov8m.pt
For face refine's mask mode: SAMS model in ComfyUI/models/sams
https://github.com/ultralytics/assets/releases/download/v8.4.0/sam2.1_b.pt
----------------------------------------------
And the optional Ultralytics full package
https://github.com/ultralytics/ultralytics
# Windows portable
python_embeded\python.exe -m pip install ultralytics
# Linux/macOS
python -m pip install ultralytics
```
Restart ComfyUI completely afterwards. If a frontend file changed, do a hard refresh in the browser too.
Requirements: a recent ComfyUI build with official MiniMax H3 support. This fork calls the H3 nodes with keyword arguments, which matches the ComfyUI core node API from the io.Schema / ComfyNode rewrite (v0.34.x-era) that broke the older positional calls — and it still works on earlier builds.
---
## Quick start: the four workflows people actually use
The Director exposes native T2V / I2V / FL2V / R2V / V2V / RV2V modes plus Mixed Mode, where every segment picks its own method. Here are the four most useful patterns. Full walkthroughs are in docs/USER_GUIDE.md.
### 1. T2V text to video
The simplest path, and the best way to test the node.
1. Generate → select T2V — Text to Video.
2. Set aspect ratio, megapixels and FPS.
3. Add Prompt Group once per shot. Give each group its own prompt and duration.
4. Write each shot so it carries character, environment and action state forward — that is what makes separate segments read as one scene.
5. Run, watch Live Preview, then Results → Final Result to save.
For a 30-second piece, use three 10-second groups rather than one long one. If a single shot comes out wrong, turn on Selective Run and reroll just that shot.
### 2. I2V animate an image
Use this when you already have the frame you want to start from.
1. Select I2V.
2. Upload one start image to the segment.
3. Describe what happens next. Don't re-describe what the image already makes unambiguous — spend the prompt on motion, camera and performance.
4. Set duration and generate.
Works well for animating a character portrait, starting from a designed scene, or continuing from a previous segment's last frame (**Segment Result** lets you pull a decoded frame from an earlier Mixed segment straight into an I2V input).
### 3. Ref2V / ref2va one character, many shots
This is the mode for a recurring character. The node's R2V mode drives the REF2VA model path.
1. Select R2V.
2. Put the identity references in Common References so every group inherits them:
- Picture 1 = headshot (face identity)
- Picture 2 = character sheet (front / side / back — body shape and wardrobe)
3. Add one Assets Group per shot. Each group takes up to Picture 1–9, Video 1–3, Audio 1–3.
4. Keep only shot-specific material local to each group — a scene plate, a motion reference, a voice clip.
5. Write each shot's prompt, including how each reference should be used (and what it should not be used for).
6. Generate. Reroll only the shots that miss.
Because references live at project level, the character stays consistent across the whole sequence without re-uploading anything.
### 4. Character Replace swap the performer in real footage
Re-renders chosen windows of a source video with your referenced character. Each window gets an automatic subject mask (SAM3, by text prompt) and is rendered as an independent masked clip: the person is regenerated from your references, the background stays the real source background.
1. Load a Source Video and add your replacement character to Common References (Picture 1 = face, Picture 2 = character sheet) — set it up as you would for RV2V.
2. Flip Replace: ON on the timeline.
3. Add windows with + Add after / + Add at playhead (default 5 s, up to 20 s; lengths snap to the H3 frame grid). Windows may overlap or be out of order — list order is render order.
4. Per window, set:
- kind = sam3 (auto mask from a text prompt, no mask files needed)
- render = anchor (recommended — strongest identity) or inpaint (pixel-exact background, weaker identity)
- SAM3 prompt describing the source person, e.g. the woman, full body from head to toe, including every strand of her hair
- grow / feather for seam control
5. Generate. During the prepare phase, Live Preview shows a Mask check card (source | red subject overlay | motion reference) so you can abort before sampling if the mask is wrong.
If SAM3 misses, OOMs, or can't run, that window falls back to a plain RV2V re-render — it never hard-fails the run. Resume reuses finished windows from cache and does not re-run SAM3 on them.
Validated on an RTX 3090 (24 GB): 0.4 MP working resolution, 5–10 s windows, in-run SAM3 auto-masking. 16 GB users should prefer the pre-made frames mask kind.
---
## What's new in this fork
Everything below is additive on top of upstream. Upstream's full feature set is untouched.
v1.7.0 2026-09-14 Render once, then iterate on post-process without re-rendering
Reuse the first pass. Turn on Reuse cached first pass and the raw first-pass AV latent is written to disk keyed on everything that produced it except post-processing. Re-run with the same seed, prompt, references and resolution but different Global Refine, Face Refine or Audio Room settings and the expensive H3 sample is skipped, a 107-frame r2v render measured 363 s down to 33 s on reuse. The key only invalidates when the seed, prompt, references, resolution, sampler or model change, because those genuinely change the first pass. Every hit and miss is logged, and a settings snapshot plus a field-level drift diff explain any surprise miss.
Tiled refine. The Global Refine second pass can now run as overlapping spatial tiles (tile_size, tile_overlap), so a large upscale fits in less VRAM. Each tile is sampled alone with its own keyframe crop, brightness-matched back to the input, and stitched with a linear cross-fade on the interior edges; audio is taken from the first tile.
Compare raw vs processed. Export comparison writes <prefix>_raw_firstpass and <prefix>_postprocessed so the unprocessed and processed videos can be A/B'd side by side.
Refine opt-out on an external patch. allow_refine_on_external_patch lets the second sampling pass be skipped when a patch already supplied the output.
Latent continuation (early test). An opt-in, experimental latent_continuation_enabled port of the community "continue in latent space" joiner: the previous segment's final latent is injected into the next segment's sampling stream and locked with a nested H3 noise mask. Flagged EARLY TEST, not for production yet.
v1.6.0 — 2026-09-13
RefMod identity references, a Results player that streams instead of shipping base64, and
Audio Refine made cache-free.
currently testing in the RefMod workflow only but will take over ref2va as an option with more testing and updates.
Added
RefMod identity references. A connected
Apply H3 RefModchain now reaches every
segment: the Director harvests its reference blocks and appends them to each segment's
own conditioning, so a character mod can carry identity across a whole chain with no
headshot and no character sheet. The mod is attached to the DiT only — it is never
presented to the text encoder, and no<Picture n>label is created for it.Minimax h3 Director - ref2va + RefMod example workflow 1x4s.json— a single 4 s shot
for fast identity iteration, shipping an in-workflow usage guide covering identity,
wardrobe, the knobs and their reasons, the failure modes, and a bisect order.A wardrobe channel. A RefMod latent is roughly a 96 x 54 thumbnail per frame, so it
carries face and body and physically cannot carry a garment's cut, seams or trim. Clothing
now comes from awardrobe:prompt section, or from a full-resolution picture reference,
with the prompt assigning each source its job so the two do not fight.Streamed preview clips in Results. Finished segments and final results are encoded to
a small all-intra H.264 clip in ComfyUI's temp directory and streamed over/view, instead
of pushing every frame as base64 JPEG through the websocket. One 243-frame segment measured
~11.9 MB of base64 parsed synchronously on the browser main thread; the clip is ~2.5 MB and
streams with ordinary HTTP range requests, at zero websocket cost. All-intra is deliberate:
every frame is a keyframe, so a seek lands on the exact frame rather than the nearest
earlier one.
### v1.5.0 — Audio Room, per scene
Generated audio no longer has to sound like a dry camera mic.
- Pick a named space — dry, bedroom, bathroom, bar, office, car, hall, cathedral, outdoor — and all six reverberation parameters are set for you. Or choose Custom and dial reverberance, HF damping, room size, stereo depth, pre-delay and wet gain yourself. A separate Level section adds normalise and gain.
- Per scene, not per project. A scene can declare its own room, so a bathroom scene and a bedroom scene in the same render stop sharing one acoustic setting.
- It cannot invalidate a cache. The chain runs at output assembly, downstream of the segment audio cache, and model audio is processed before the merge so a merged export still gets one space per scene. Turning a room on or changing it invalidates no segment, no context cache, no finished render.
- Stereo is preserved, failures are loud. Built on SoX. Unlike the third-party effects node it replaces — which read waveform[0, 0] and returned mono from stereo input, silently returned the original on every failure, and round-tripped through 16-bit PCM — channel count and sample count are preserved, and a track that can't be processed is left dry and named in the report.
- Never-spoken guard. H3 generates audio from the same text it renders, so a bare token or quoted phrase in shared reference material is a shape it can read aloud. A checkbox appends an explicit control line marking the block as reference-only.
### Re-ground — stop drift on long chains
Long multi-segment chains drift in colour, contrast and identity after several hops. Re-ground re-anchors a segment's continuity context at the chain root (the clean start of the job) instead of the immediately-previous segment, resetting accumulated error without breaking the flow.
Every segment boundary shows two small circles: the top ↔ / × is the existing context link, the bottom *R** is Re-ground. Left-click to toggle (it turns amber); right-click either circle for the boundary menu. Use it every 3–5 shots on long jobs, alongside Latent Scale Lock and Color Re-anchor.
### Performance
- ~525× faster prompt truncation on the draw path. Prompt fitting used an O(n²) character-by-character measureText loop and could freeze the Director modal for up to ~30 s on projects with large per-segment prompts; it now uses binary search.
- Live preview decode cap. With shipped defaults the live preview decoded an 8-frame clip on every diffusion step of every stage, on the sampling thread, on the same GPU doing the diffusion. Decodes are now capped at one per 400 ms per segment/stage, while always decoding the first and final step so the preview stays a faithful progression. Output-neutral.
- Idle timeline loop. The Replace-windows loop ran at display rate as long as a node sat on the canvas. It now drops to a 250 ms poll when idle and returns to full rate on hover — and the frame handle is finally stored, so it can actually be cancelled (previously cancelAnimationFrame was a no-op).
- Cheap rendering wins content-visibility on batch group cards).
### Resume, Stop and correctness
- Resume was being ignored by four separate plan builders — prompt_batch/gen, fl2v, mixed and external_groups each dropped the resume flag, so Resume silently restarted from segment 1 on those timeline shapes. Fixed, and a structural test now scans every builder for the resume fields (which is how the fourth instance was found).
- The engine now decides the resume start point, not the dialog. The preview and its audio check no longer have blind spots that could present a partially-cached run as complete, and a stopped run no longer stays marked running.
- Resume was greyed out whenever the manifest had no done marks, even when caches were present and reusable. Fixed.
- A CUDA out-of-memory was reported as a sampler incompatibility — an OOM in the external-sampler branch got wrapped into "the sampler does not support MiniMax H3 inputs," sending you after the wrong problem entirely. OOM now surfaces as OOM.
### v1.4.0 — Stop, undo/redo, presets, sweeps
- Partial export on Stop. Stopping no longer throws the run away: the current segment finishes, everything completed is assembled into a real partial video, and the run is recorded as stopped with the resume manifest intact — so Resume continues from the first unfinished segment.
- Timeline undo / redo — Ctrl+Z / Ctrl+Shift+Z Ctrl+Y too), bounded at 50 steps, recorded from the single commit funnel so every edit is covered. Deliberately inert while you're typing in a text field.
- Named presets storing settings only — never your segments, prompts, task type or seed, so a preset can't quietly rewrite a project. Server-side atomic storage; a corrupt index fails loudly rather than being replaced by an empty one.
- Multi-seed sweep — render N takes with N seeds in one click. Take seeds are hashed rather than seed + 1, because consecutive seeds can produce visibly correlated results.
- Fewer surprises before you queue — Validate runs a pre-flight check (empty timeline, H3 frame-grid violations, missing references, Character Replace setup mistakes), Preview prompt shows the exact text the model receives including what the engine appends, and References… cross-checks which slots your prompts mention against the files actually attached.
### Safety guards
- Global Refine is skipped automatically when the diffusion model carries an external attention patch such as SLA or Spectrum, which a second guidance-distilled pass would otherwise corrupt. An opt-out exists for advanced users.
- Per-segment audio in Results — each segment's audio reaches the Results player as soon as that segment finishes, so the Segment view has real aligned sound while the job is still running.
### Tests + CI
- Python unit and contract tests run from any directory with no live ComfyUI python -m pytest), plus standalone frontend tests. A CI workflow runs both on every push and PR.
- 410 Python tests, 21 frontend tests.
---
## Models used by the bundled example workflows
Drop these under your ComfyUI models/ folder using the subfolders shown. The paths match the workflows' model subgraphs — swap the loader entries if your filenames differ.
| Role | Place in ComfyUI/models/… | Source |
|---|---|---|
| REF2VA diffusion (beta4) | diffusion_models/Minimax/10Eros_Max_h3_TURBO-hybrid_beta4_int8_convrot.safetensors | [TenStrip/10Eros-Max](https://huggingface.co/TenStrip/10Eros-Max) |
| REF2VA diffusion (beta5) | diffusion_models/Minimax/10Eros_Max_h3_TURBO-hybrid_beta5_int8.safetensors | [TenStrip/10Eros-Max](https://huggingface.co/TenStrip/10Eros-Max) |
| FL2VA base | diffusion_models/Minimax/minimax_h3_fl2va_pruned_int8_convrot.safetensors | [Kijai/MiniMax-H3-experimental](https://huggingface.co/Kijai/MiniMax-H3-experimental) |
| Fast-video path | diffusion_models/minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors | [Kijai/MiniMax-H3-experimental](https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors) |
| Text encoder (uncensored, nvfp4) | text_encoders/qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors | [sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4](https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4/blob/main/qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors) |
| Video VAE (int8 convrot) | vae/minimax_h3_video_vae_int8_convrot.safetensors | [Kijai/MiniMax-H3-experimental](https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/minimax_h3_video_vae_int8_convrot.safetensors) |
| Audio VAE (fp32) | vae/minimax_h3_audio_vae_fp32.safetensors | [Kijai/MiniMax-H3-experimental](https://huggingface.co/Kijai/MiniMax-H3-experimental) |
| REF2V turbo LoRA (4-step) | loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors | [LINK NEEDED] |
| FL2V turbo LoRA (4-step, 768p) | loras/minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors | [LINK NEEDED] |
Base model: [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) — the official release, including the FL2VA, Ref2VA, vae, text_encoder and transformer_ref folders.
Notes
- The int8-convrot VAE needs ComfyUI 0.31.0 or newer or you get black outputs. It also speeds up VAE decode by roughly 1.5×.
- The nvfp4 text encoder is ~15.7 GB and fits a single 16 GB card.
- The bundled workflows also wire an empty Power Lora Loader — it is deliberately left empty as a free slot for your own LoRA.
- Kijai's repo also carries an alternative set of accelerator LoRAs MiniMax-H3-Ref2VA-Acc-8Step, MiniMax-H3-FL2VA-Acc-8Step, pruned and unpruned) if you prefer 8-step acceleration over the 4-step turbo LoRAs.
---
## Example workflows included
Three ready-to-run workflows ship in example_workflows/:
| File | What it shows |
|---|---|
| Minimax h3 Director - ref2va example workflow 3x7s.json | 3 × 7 s continuous golden-hour forest run. One character, dynamic camera, three segments, with bundled AI-generated placeholder headshot + character sheet. |
| Minimax h3 Director - t2v example workflow 5x7s - elf vs giant orc.json | 5 × 7 s pure-prompt story (35 s @ 24 fps), 8-step hybrid path, Audio Refine on, all cross-segment context links on. |
| Minimax h3 Director - character replace example workflow 3x7s - elf vs giant orc.json | 3 × 7 s ref2va character replacement — your referenced character replaces the lead of the source action. |
Each one opens routed to the right model path via the workflow's ImpactSwitch — select = 1 for the FL2VA/T2V-I2V path, = 2 for REF2VA, = 3 for the fast-video path. Nothing else needs wiring.
The bundled ref2va_example_assets/headshot.png and charsheet.png are AI-generated placeholder characters, not real people — swap in your own.
---
## Things worth knowing
- Do not load the standalone ComfyUI-H3-Motion-Context alongside this pack — Motion Context compatibility is already integrated.
- Continuity improves cross-segment handoff but does not guarantee an invisible boundary every time. H3 can still introduce visual, motion, lighting or identity drift — that is what Re-ground, Latent Scale Lock and Color Re-anchor are for.
- Clear VRAM Between Segments is the biggest non-output speed lever on multi-segment runs. Turn it OFF if your GPU fits the model resident; turn it ON when SAM3 is in use so SAM3 and the DiT never peak together.
- Cache rules are precise and documented: resolution/megapixels, prompts, references, shot ranges and continuity invalidate; seed, steps, sampler, CFG, bitrate/CRF and the refine toggles do not.
- Optional post-processing (NVIDIA RTX VSR / Deblur, Face Refine detectors, upscale models) degrades gracefully — if a stage can't run, the Director keeps the usable earlier result instead of discarding the generation.
---
## Credits
Upstream: [j955229 / ComfyUI-MiniMax-H3-Motion-Director](https://github.com/j955229/ComfyUI-MiniMax-H3-Motion-Director)
Model and tooling credits: MiniMax (H3), TenStrip (10Eros-Max), Kijai (experimental H3 quantisations and VAEs), sakamakismile (uncensored nvfp4 text encoder), SoX (audio room DSP).
This fork: [The5toryofthe5ecret5tar5ystem / ComfyUI-MiniMax-H3-Motion-Director](https://github.com/The5toryofthe5ecret5tar5ystem/ComfyUI-MiniMax-H3-Motion-Director) — GPL-3.0. Issues and PRs welcome.
---
<small>MiniMax H3 Motion Director is a ComfyUI custom node pack. Bring your own MiniMax H3 weights — none are redistributed here. Check the MiniMax H3 community licence before commercial use.</small>