MiniMaxH3Director

MiniMax H3 is the video model that dropped in late July 2026 and immediately had the local community clearing hard-drive space: open weights, up to 15 seconds at 2K, native stereo audio, and reference-to-video input that ComfyUI supports natively. Native support, though, means native nodes - and the native MiniMaxH3ImageToVideo / MiniMaxH3ReferenceToVideo nodes are bare function calls. If you want to compose a scene from multiple references, trim clips, order shots, and manage prompts without a spreadsheet, you need something on top.

MiniMax H3 Director is that something. It's a timeline-based authoring node - a direct descendant of the LTX Director idea that WhatDreamsCost made famous - rebuilt for H3. One node holds your images, videos, audio, trims, ordering, and prompts, and hands the whole thing to ComfyUI's native H3 implementation.

The mechanism

The Director doesn't reimplement MiniMax H3 - the README is explicit that it routes to ComfyUI's built-in nodes. It does the authoring: you drop media into two timeline lanes (Image/Video and Audio) via drag-and-drop, paste, or file picker, set per-clip trim ranges, reorder tiles by dragging, and write your prompt in a mode-specific builder. On queue it decodes everything (video cropped to your trim, audio extracted and cropped the same way), validates H3's hard limits, and emits a structured guide dictionary plus a resolved prompt to the sibling MiniMax H3 Director Guide node, which makes the actual native call.

Two modes:

• FL2VA - text-to-video, first-frame, or first+last-frame interpolation. Up to 2 image slots; audio/video blocked. The builder has guided fields (integrated_multimodal_description, overall_soundscape, non_diegetic_music) and auto-inserts the alignment lines that H3 wants.
• REF2VA - the reference mode that makes H3 interesting. Up to 9 images, 3 videos, 3 audio clips (12 files total), each video switchable V / A / V+A. The builder here is six free-text sections following H3's official full-reference format, with helper buttons: Insert [Shot N], Prefill Labels & Summary, and Preview Prompt so you see the exact assembled prompt before you commit.

Inputs that matter

• mode - T2VA / I2VA / FL2VA / L2VA / REF2VA. Everything about the node's behavior follows this switch.
• duration (default 5s) and frame_rate (default 24) - set your clip length and FPS. frame_rate is also emitted as an output so downstream nodes read the effective value.
• The Resolution panel (Aspect / Resolution / Input scaling, all default Auto) drives the output canvas on H3's 16px grid. Auto resolution sets a 768px short side; Auto aspect follows your first visual reference.
• The optional fl2va_model / ref2va_model sockets enable lazy loading - only the model for your active mode gets requested.
• external_prompt_overwrite (STRING) and the paired external_width_overwrite / external_height_overwrite (INT) let you bypass the builder and canvas entirely when you already know exactly what you want.

The outputs

The one that matters is guide - it feeds the Director Guide node. Also useful: duration, frame_rate, width, height, positive_prompt (the assembled text, handy for debugging), and model / fl2va_requested / ref2va_requested for routing.

Installing it

ComfyUI Manager (search DaSiWa-Nodes), or:

cd ComfyUI/custom_nodes
git clone https://github.com/darksidewalker/ComfyUI-DaSiWa-Nodes
pip install -r requirements.txt

then restart. You also need a ComfyUI build with native MiniMax H3 support and the H3 model files (checkpoint, Qwen3-VL text encoder, visual VAE, and the audio VAE for REF2VA) via ComfyUI's model manager. This is where the pack's RTX requirement gets fuzzy - H3 itself runs on consumer cards (it reportedly runs on a 3060, slowly), so this node needs no NVIDIA SDK, just the model and patience.

Where people get burned

• Forgetting the audio VAE in REF2VA. The Guide refuses to run without audio_vae connected. Wire vae/minimax_h3_audio_vae_fp32.safetensors in.
• Hitting the limit walls. References must be 2–15 seconds, combined visual ≤15s, combined audio ≤15s. The node validates this and shows red status messages - believe them.
• 16px vs 32px grid trap. H3's VAE works on 16px latent cells, but its transformer groups them in 2×2 patches - so a width/height that's a clean multiple of 16 but not 32 can fail at sampling. Use the resolution panel's presets rather than raw CUSTOM values unless you know what you're doing.
• Media that stays on your disk. Files you upload get stored under ComfyUI's input/ directory, referenced by relative path. Delete them there and old timelines break.

This is a big, opinionated node with a real learning curve, and it's the reason to install the pack if you do H3 video. The one thing it can't do is make a 15-second 2K generation fast - that's H3's problem, not the Director's.