MiniMaxH3Director

MiniMax H3 was the release that finally gave open weights what Veo had in closed ones: video and stereo audio generated together in one pass, no separate "add a soundtrack" step. The catch for ComfyUI users is that the official H3 nodes are single-shot. One prompt, one clip, done. That's fine for a clip. It's wrong for a scene.

MiniMaxH3Director (also registered under the legacy id ComfyMiniMaxH3Director) is what you reach for when one clip isn't the job: a timeline editor packed into one node. You split, trim, and reorder segments, type a prompt per segment, queue - and it handles conditioning, sampling, AV decode and export for every segment in order, handing you one finished video with sound. No API, no key; it just drives the official H3 pipeline harder.

How it works

Each segment runs through ComfyUI's official MiniMaxH3ImageToVideo or MiniMaxH3ReferenceToVideo conditioning (tokenization happens internally, so you don't wire up text encoding yourself), then a single-stage KSampler at CFG 1.0 with MiniMaxH3SigmaShift applied, then an AV-latent decode that splits picture and stereo audio apart. The task_type menu decides which official node runs per segment: t2v, i2v, and fl2v (first-last frame) go down the fl2va path; r2v, v2v, and rv2v go down the reference path and need the audio_vae. There's a "mixed" mode too, if you want different segment types in one timeline.

The interesting trick is segment continuity. With inter-segment guidance on in the UI (default off), the tail motion and generated audio of one segment get pinned into the next and cropped off after sampling - which is what stops a multi-segment video from feeling like five videos glued together. Credit for the approach goes to ComfyUI-H3-Motion-Context.

Wiring it up

The pack needs ComfyUI 0.30.0+ - it calls ComfyUI's official MiniMax H3 nodes (PRs #15224/#15228) and refuses to load without them. Install:

cd ComfyUI/custom_nodes
git clone https://github.com/AIMixer/ComfyUI_MiniMaxH3_Director.git
pip install -r ComfyUI_MiniMaxH3_Director/requirements.txt

Restart, or use ComfyUI Manager → Install via Git URL with the same URL. requirements.txt adds opencv-python-headless, imageio-ffmpeg, and scenedetect for the v2v/rv2v source-video features and smart shot-splitting.

Models aren't bundled - that's the heavy part. You need the H3 video VAE, the H3 audio VAE (fp32), a Qwen3-VL CLIP loaded with CLIPLoader type = minimax, and two UNETs: fl2va for t2v/i2v/fl2v, ref2va for r2v/v2v/rv2v. The README points at Comfy-Org's MiniMax-H3 repo on Hugging Face or the author's Comfyit pack. This is a 33B model at roughly 42.5 GB full precision, so grab the pruned/int8 variants the README recommends and bring a real GPU. And check your region before downloading: the MiniMax H3 Community License excludes the US, EU, UK, and South Korea from running the local weights - a licensing fact, not a technical one.

The inputs that matter

The four model inputs are model, video_vae, audio_vae, and clip (minimax-type Qwen3-VL). Then the few you actually touch:

• task_type - t2v / i2v / fl2v / r2v / v2v / rv2v / mixed. Picks the official node and the UNET the workflow expects.
• global_prompt - sent straight to the H3 node. In r2v you reference a loaded image as <Picture 1>; in v2v the source segment is auto-bound as <Video 1>.
• frame_rate, width, height, total_frames - defaults 24 fps, 864×480 (0.4MP 16:9), 124 frames ≈ 5s; frame counts snap to MiniMax's 17k+5 grid.
• cfg (1.0 - not a typo, H3 wants it low), seed, plus the advanced group: steps 25, sampler res_multistep, scheduler simple, shift_video 12 / shift_audio 3. All mirror the official template.

The optional sockets are where this gets interesting: i2v_groups / r2v_groups take multi-group packs from this pack's Group nodes (external input overrides the UI cards), and refine takes the MiniMax H3 Director Refine node for a second-pass polish or upscale.

Outputs

images (one frame list per segment), audio (native AUDIO - wire it to your sound-saving node), fps, frame_count, and a report string summarizing each segment's plan. With a refine node attached, images_pre_refine carries the first-pass result for comparison. Leave source_images disconnected unless you explicitly want the original timeline frames decoded - it won't touch the main images output.

Where people get burned

Same three complaints every time: missing official H3 nodes (upgrade ComfyUI), a ref2va clip dropped on the fl2va UNET (wrong model for the task type), and a CLIP loaded without type=minimax (garbage conditioning). One more: the report output is where decode failures get explained - the node prints gray placeholders rather than fake frames, so read it before assuming the sampler broke.