Workflow Config MiniMaxH3

MiniMax H3 is MiniMax's next-generation video model - a multimodal, audio-and-video-in-one model that landed as an open-weights release around mid-2026 (the "open" part drew some skeptical side-eye in the community, since MiniMax's Hailuo family had been API-only). If you're building workflows around it, you want the same thing every video config wants: a way to store named scene setups and flip between them without rewiring the graph. This node is that - the pack's scene system, retargeted at H3.

It works exactly like its siblings in comfyui-daz-tools. Scenes are saved as takes inside dx_*.json "movie" files in ComfyUI/user/default/workflows/.dx_mgr/, and you switch between them with three dropdown inputs: movie, scene, take. Pick a take and the node loads the UNet, VAEs, CLIP, reference image and audio, dimensions, steps, seed, prompts, and all 8 LoRA slots, then pushes every value out as an output. You wire the graph once and iterate by changing dropdowns.

Where it differs from the pack's other config nodes is the model layout. H3 has a single UNet - no high/low pass split like WAN2.2 - but keeps the dual-VAE design: video_vae and audio_vae, because H3 is another audio-and-video-in-one model. One clip, one cfg, and 8 LoRA slots where, again, the first slot is named distillation_lora - H3's distilled workflows run few steps at low CFG, so that's your speed slot.

The outputs that matter:

• unet_stack (MODEL) - the UNet with all enabled LoRAs applied. Wire this into your sampler; it's the node's reason to exist.
• unet_only (MODEL) - the bare model if you want to apply LoRAs yourself.
• video_vae, audio_vae (VAE), clip (CLIP) - the decode and text-encode stack.
• image (IMAGE), audio (AUDIO) - reference inputs for I2V and audio conditioning.
• width, height, steps, seed, cfg, total_frames, fps - straight into sampler and latent setup.
• pos_prompt, neg_prompt, master_prmt, is_relay_prompt - the prompts; the relay flag flips on for Smart (segmented) prompts.
• filename, flag_1/2/3, custom_1/2 - output path and routing helpers.

Note what's not here compared to the LTX2.5 sibling: no latent_upscaler output. H3 keeps its scaling inside the pipeline, so that output simply doesn't exist on this node.

The editor is the same full-screen panel: prompts with a visual segment editor (including an H3-specific prompt type where segments are written At X.Ys, … in decimal seconds), a sizing dialog with /32 divisibility checks, and duration quick-buttons that compute frames as duration × fps + 1. Save or "+ Take" writes back to disk.

Install:

cd ComfyUI/custom_nodes
git clone https://github.com/denyazzolin/comfyui-daz-tools

Restart ComfyUI, or use ComfyUI Manager and search comfyui-daz-tools. No Python dependencies beyond what ComfyUI already ships. If you want GGUF-quantized UNets - the dropdown lists .gguf files alongside .safetensors and loads them via ComfyUI-GGUF automatically - install ComfyUI-GGUF first, or GGUF files won't appear and a GGUF-configured scene errors at runtime.

Practical notes. Since H3 is young, you're probably assembling the model files yourself; double-check which VAE file is the video VAE versus the audio VAE, because the node has separate slots for each and swapping them is the classic silent failure. Your scenes live in the .dx_mgr folder - back it up. And as with the whole pack, scenes are interoperable across the config nodes: any movie file can hold scenes for any class, so you can organize by project and jump between H3 and LTX takes from the same folder.