MiniMax H3 Director 2.0

If you've ever tried to drive MiniMax H3 in raw ComfyUI, you know the pain: reference clips and trims live in scattered loaders, the prompt has to be hand-formatted with alignment lines and section headers, and the native MiniMaxH3ReferenceToVideo node expects everything pre-shaped. MiniMax H3 Director 2.0 is DaSiWa's answer - a timeline node where you drag in references, trim them, write the prompt in a guided builder, and it handles conditioning, sampling, and output itself. It's the LTX Director idea (the pack credits whatdreamscost's node as the inspiration) ported to MiniMax's 33B omni-modal model. One node, one job: turn your media into an H3 clip without building a graph around it.

Before you get excited, the model itself is the gatekeeper: H3 is a ~42.5 GB 33B open-weights video model with native audio, and the MiniMax H3 Community License geofences out the US, EU, UK and South Korea - if you're in one of those, running the local weights isn't licensed, period. And it wants serious VRAM; there's no verified consumer-hardware floor yet, so don't plan around a midrange card. This node is for people who already run H3, or are willing to pay for the hardware to get there.

How it works

Under the hood the Director doesn't reimplement MiniMax's backend. It validates your timeline, assembles the prompt (adding the alignment headers and section labels you'd otherwise type by hand), then calls ComfyUI's own native nodes - MiniMaxH3ImageToVideo for the T2VA/I2VA/FL2VA/L2VA family, MiniMaxH3ReferenceToVideo for REF2VA - samples with the standard custom-sampler chain (defaults res_multistep / simple / 25 steps, all overridable), decodes, and writes the result to ComfyUI Assets. The model sockets are lazy-loaded: only the model matching your active mode is requested, so you can wire both an fl2va_model and a ref2va_model without loading the one you're not using.

The inputs that matter

• mode - the one you'll touch most: T2VA (pure text, no image), I2VA (first frame), FL2VA (first+last frame interpolation), L2VA, REF2VA (multi-reference), and Image Inpaint. Each mode changes the timeline lanes and prompt builder.
• fl2va_model / ref2va_model (MODEL sockets) - your H3 UNet, plus clip and vae. In REF2VA you also need audio_vae; the node refuses to run without it.
• duration (seconds), width / height, frame_rate (default 24). The node re-emits frame_rate and duration as outputs so downstream video nodes read the effective values instead of guessing.
• prompt and the hidden timeline_data / builder_state JSON - don't hand-edit those two; they're the node's serialized state. If you'd rather not use the visual builder, external_prompt_overwrite replaces the assembled prompt, and a complete external_width_overwrite + external_height_overwrite pair bypasses the Director's canvas sizing entirely.

Its outputs are frame_rate (FLOAT), duration (INT), and images (IMAGE). Feed images into Enhanced Video Combine or a standard saver.

Modes at a glance

REF2VA is the star: up to 9 images, 3 videos, 3 audio clips (12 files total), a per-video V / A / V+A switch for using embedded audio, and draggable crop markers. Clips must be 2–15 seconds each with ≤15s combined visual and audio - the node validates all of it and shows red status messages instead of letting H3 fail on you. FL2VA is stricter: up to 2 images, no video or audio. Keep at least one image or video when using audio in REF2VA.

Install

Via ComfyUI Manager, search DaSiWa-Nodes and install. Manual:

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

then restart ComfyUI. The pack's requirements are heavy (torch, transformers, accelerate, PyAV, imageio-ffmpeg) and the README asks for an NVIDIA RTX GPU with drivers 530+ - that's mostly the RTX upscaler in the same pack, but it's the pack's stated floor. More importantly, your ComfyUI build needs native MiniMax H3 support plus the H3 model files (UNet, the Qwen3VL text encoder, video VAE, and audio VAE for REF2VA) or the Director has nothing to hand off to.

Common issues

The author is the DaSiWa behind the well-regarded Wan 2.2 all-in-one workflows - superb, feature-dense, but that style is also the trap: these packs lean on the latest frontend features, and the community has hit workflows breaking after ComfyUI updates. Same expectations apply here. If REF2VA errors instantly, you almost certainly forgot the audio_vae. If sampling fails at odd resolutions, remember H3's VAE has 16-px latent cells but the transformer groups them in 2×2 patches, so a 16-px-only edge can fail - the resolution panel snaps to the 32-px grid for you, and defaults are Auto for a reason. Keep the node updated alongside ComfyUI, and don't chase the newest sampler controls in external_sampler until the defaults stop doing what you need.