⭐ Star Minimax All In One

MiniMax H3's reference-to-video node is powerful and genuinely intimidating: model loader, text encoder, video VAE, audio VAE, reference conditioning, sampler, two decoders - a wall of nodes that new users copy-paste and pray over. Star Minimax All In One collapses that entire pipeline into a single node. It loads the diffusion model, the minimax text encoder and both VAEs internally, builds the <Picture i> / <Video k> / <Audio j> reference conditioning for you, samples, and decodes video and audio. One node, in and out.

This is the pack's flagship new addition, and it's aimed squarely at people who've fought the H3 pipeline before: the reference slots grow automatically as you connect more images, videos or audio clips (the COMFY_AUTOGROW_V3 type), exactly like the core node. It also ships a 2:1 panorama preset and an image mode that renders 9 frames and hands you only the best one (frame index 8) as a still - useful when you want a high-quality frame without decoding the whole clip.

What you're really wiring

• mode - video (full clip with audio) or image (single best still frame, skips duration and audio decoding).
• prompt - reference inputs by tag in connection order (<Picture 1>, <Video 1>, <Audio 1>), then describe scene, motion and audio. The ordering matters - it matches your socket order.
• aspect_ratio / megapixels - output dimensions; megapixels also offers an audio only option with a tiny 32x32 canvas if you just want sound out.
• match_ratio_from_image - pulls the closest aspect ratio from your first reference image.
• duration - seconds at 24 fps, internally snapped to the model's 17k+5 frame grid (5s → 124 frames).
• ref_image_size - match scales refs down to the generation's pixel area; max uses a 2048px short edge for best identity fidelity but runs much slower.
• steps / sampler / scheduler / denoise / seed - standard sampling dials. The tooltip notes beta or normal tends to beat simple for reference-heavy prompts.

model_override is the interesting optional input: plug in an external MODEL (say a sage-attention patched H3) and it ignores the internal diffusion_model dropdown entirely.

The model files it expects

This node is not fire-and-forget on dependencies. It looks for the MiniMax H3 files in your standard ComfyUI folders, with these defaults (all overridable in the dropdowns):

• diffusion_model: minimax_h3_ref2va_pruned_int8_convrot.safetensors
• clip_name: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (text encoder)
• vae_name: minimax_h3_video_vae_fp16.safetensors
• audio_vae_name: minimax_h3_audio_vae_fp32.safetensors (fp32 recommended)

If those names don't show up in the dropdowns, the files aren't where ComfyUI can find them. This is a heavy local video pipeline - expect real VRAM requirements and long runs, not a casual test node.

Outputs

IMAGE, AUDIO, FPS (fixed 24.0 - feed it straight into a video combine/save node), LATENT (the combined video+audio NestedTensor pre-decode), plus MODEL, CLIP, VAE and AUDIO_VAE if you want to reuse the internals elsewhere.

Install

Same pack, same routine:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

Restart and search star. The node itself is pure ComfyUI + the models you download; the requests/imageio-ffmpeg dependencies in requirements.txt cover the pack's video siblings, not H3 specifically.

Where people get burned

Refusing to read the prompt-tag ordering is the classic - references silently do nothing if your prompt doesn't reference them by tag. Second: running ref_image_size = max on a big reference and wondering why the run took forever; match is the sane default. And the image mode is a still-from-video trick, not an image-edit mode - if you wanted an edit pipeline, you grabbed the wrong flagship.