CineStack H3 Director

MiniMax H3 is a 33B omni-modal video model: 4–15 second clips, 24fps, stereo audio generated jointly with the picture instead of bolted on afterwards. That's genuinely great, and also the whole problem. You don't want a ten-second clip, you want a scene - and the official H3 graph gives you one shot per chain, so five segments means five copies of the graph and a full re-render when segment 3 has a bad hand.

CineStack H3 Director is the answer to that specific annoyance. It's one node that holds a whole multi-segment timeline, generates it segment by segment, caches each segment to disk, and lets you redo just the one you hate.

What it actually is

This is a workbench node, not a sampler. Under the hood it does what the official H3 template does - encode the prompt with the minimax CLIP, build an audio+video latent, patch the model with MiniMaxH3SigmaShift, run a single-stage KSampler at CFG 1.0, then decode through LTXVSeparateAVLatent into frames and audio. The difference is that it repeats that per segment, writes each result to output/minimax_seg_cache/<cache_name>_<nodeid>/, and keeps a per-segment seed so a redo doesn't nuke the shots you approved.

The pack is an honest fork: AIMixer's MiniMax H3 Director, Apache-2.0, with segment seeds, strict cache reuse, selective rework and rolling export added on top. It's young - 0.6.11, Git-only, no Registry listing, almost no community thread trail yet - so you're trusting the README and the source.

The inputs that matter

The four socket inputs are non-negotiable: model (the H3 UNET), video_vae, audio_vae, and clip. audio_vae is a required socket even for pure t2v, where the tooltip says it's only for r2v/v2v/rv2v. Plug it in anyway - H3 decodes audio as part of the deal.

On the widget side, start here:

• task_type - t2v, i2v, fl2v, r2v, v2v, rv2v, or mixed if different segments need different modes. Reference modes use <Picture 1>, video-to-video uses <Video 1> in the prompt.
• global_prompt for the text, width/height at 864×480 for a first test, and total_frames for length. Frames snap to H3's 17k+5 grid, so 124 is your ~5 seconds at 24fps; asking for 120 quietly becomes something else.
• steps 25, sampler res_multistep, scheduler simple, cfg 1.0, shift_video 12, shift_audio 3 - those are the official template values. Leave them alone until something looks wrong.

timeline_data is marked internal and populated by the UI. Don't type in it.

The rest of the sockets are where the pack earns its keep: i2v_groups and r2v_groups take external shot packs from the Group nodes (external wiring wins over the UI cards), and semantic_bridge, selflift, refine, face_refine chain the four optional refinement nodes into the same pass.

Outputs, and where they go

Ten outputs, and the pair at the end is the important one. first_pass_video and final_video are VIDEO types meant to go straight into the official SaveVideo with format auto - that's the file-backed rolling path, where FFmpeg reads and merges per-segment files instead of rebuilding the whole film as one giant IMAGE tensor in Python.

If you prefer the classic route, images and audio go through CreateVideo → SaveVideo as usual, with the caveat that the whole timeline then lives in RAM at once. report is a STRING - stick a PreviewAny on it and read what actually happened. images_pre_refine is the first pass before any second sample; images_pre_face_refine stays blocked unless FaceRefine is connected and its "output pre-face-refine" switch is on.

Install

Standard custom-node install, in the same Python environment that runs ComfyUI:

cd ComfyUI/custom_nodes
git clone https://github.com/guomenghao/ComfyUI-CineStack-H3.git
python -m pip install -r ComfyUI-CineStack-H3/requirements.txt

That file pulls opencv-python-headless, imageio-ffmpeg, scenedetect and ultralytics - the last is a hefty install, and given the December 2024 compromised-Ultralytics incident that reached ComfyUI users through detector nodes, keep it current rather than pinned to something ancient. nvidia-vfx, for RTX VSR upscaling, is deliberately not a hard dependency: AMD and cloud boxes would break.

You also need ComfyUI with H3 support, plus the weights: minimax_h3_fl2va_pruned_int8_convrot.safetensors for t2v/i2v/fl2v, minimax_h3_ref2va_pruned_int8_convrot.safetensors for the reference modes (both in models/diffusion_models/), qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors in models/text_encoders/, and the fp16 video VAE plus fp32 audio VAE in models/vae/. In CLIPLoader, the type must be minimax. Nodes live under CineStack/H3.

Where people get burned

Turning the file outputs on without rolling mode. Wire final_video into SaveVideo in normal mode and the pack tells you to enable rolling generation, because the file ports only exist on that path.

Assuming "fixed" means reproducible. Fixed seeds manage attempts; they don't guarantee pixel-identical frames across models, versions or cards. Cache reuse is stricter but pickier - change the prompt, seed or sampling settings and the cached first pass no longer matches, and "second pass only" stops and names the segment and the field instead of silently re-running.

Expecting rolling generation to be a memory ceiling. It removes decoded-frame accumulation; it doesn't make model residency, current-canvas size or VAE decode free. On a 24GB card keep clear_vram_between_segments on (it's the default), and switch on clear_vram_before_refine when your two passes use different UNETs.

Missing nodes after an update. Restart the server and hard-refresh (Ctrl+F5). 0.6.11 specifically patched panel remount and load retries on the Nodes 2.0 frontend, so if the Director panel is blank, update before you debug.

And one non-technical trap worth knowing before you invest a weekend: the MiniMax H3 community licence excludes the US, EU, UK and South Korea from its applicable territory. That has nothing to do with this pack, and everything to do with whether you're licensed to run the weights at all.