MiniMax H3 Director Guide

The MiniMax H3 Director Guide is the node nobody gets excited about, and the one the whole Director setup falls apart without. The Director is a big authoring UI - media, trims, prompts, a timeline. But a timeline isn't a generation. Something has to take all that authored state, validate it, assemble the final prompt in the format MiniMax H3 actually expects, and call ComfyUI's native H3 nodes. That's the Guide.

Think of it as a thin adapter between the Director's guide socket and ComfyUI's built-in MiniMaxH3ImageToVideo / MiniMaxH3ReferenceToVideo nodes. You never wire those native nodes yourself when you use this pack - the Guide calls whichever one matches the mode, internally.

What it does, step by step

1. Validates the incoming guide - mode consistency, that required models/CLIP/VAEs are connected, and for REF2VA that an audio VAE exists. It refuses to run REF2VA without audio_vae; that's the error you'll see first if you're missing it.
2. Assembles the prompt. It reads the Director's builder_state, wraps your six REF2VA sections in their canonical headers (subject_definitions:, summary:, retention_analysis:…), injects alignment lines for the FL2VA endpoint modes, and produces the resolved prompt that goes to the model.
3. Routes to the native node - MiniMaxH3ImageToVideo for T2VA/I2VA/FL2VA/L2VA, MiniMaxH3ReferenceToVideo for REF2VA - and emits the standard ComfyUI outputs.

Inputs

Three required, one optional:

• guide - the MINIMAX_H3_DIRECTOR_GUIDE socket from the Director. This is the only input of its type; there's no other source for it.
• clip - the CLIP from your H3 text-encoder loader (Qwen3-VL).
• vae - the visual VAE (minimax_h3_video_vae_fp16.safetensors).
• audio_vae (optional, but only nominally) - required for REF2VA. The audio VAE is fp32 and only loads when you're in reference mode.

Outputs

Just two: positive (CONDITIONING) and latent (LATENT). They feed a completely standard sampler/decoder chain - KSampler, VAE decode, video combine - exactly like any other H3 workflow. There's no magic on the way out; the magic is all in prompt assembly.

Installing it

Same pack, same two routes - 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. This node has a hard dependency you can't install your way out of: a ComfyUI version with native MiniMax H3 support. It resolves comfy_extras.nodes_minimax_h3 at execution time, so if your ComfyUI is too old it throws a clear "update ComfyUI" error rather than failing silently.

Where people get burned

• "audio_vae is required for REF2VA." Not a bug - you forgot the second VAE. Connect the audio VAE and it proceeds.
• "requires ComfyUI's native MiniMax H3 node." Your ComfyUI predates H3 support. Update ComfyUI, not the pack.
• Wiring native H3 nodes in parallel. Don't. The Guide already calls them; adding your own MiniMaxH3ImageToVideo to the graph duplicates the work and usually doubles your VRAM pressure.

It's a two-output utility with no knobs, and that's the compliment: if the Guide is doing its job, you never think about it. Set the Director up right, and this node is just the quiet relay in the middle.