MiniMax H3 GGUF Loader + Spectrum

MiniMax H3 is a genuinely big model - diffusion transformer plus a Qwen3-VL text encoder plus two VAEs (one for video, one for audio). If you're not on a 24GB card, the realistic way in is GGUF quantization, and this node is the pack's one-stop shop for loading all four pieces in a single step. It also has a party trick most loaders don't: an optional Spectrum acceleration stage that can make the whole render faster without you touching a single sampler setting.

What it loads

The node reads from standard ComfyUI folders - GGUF UNet files from models/unet, GGUF CLIP files from models/text_encoders, and the video/audio VAEs from models/vae (picked by filename keyword, so name them with video and audio in the filename or the dropdowns will be confusing). If nothing's found, the dropdowns show NO_H3_GGUF_UNET_FOUND and friends, and the node refuses to run - that's the signal that a download hasn't finished.

Under the hood it calls ComfyUI-GGUF's UnetLoaderGGUFAdvanced and CLIPLoaderGGUF (type minimax) plus the standard VAELoader, then hands you model, clip, video_vae, and audio_vae out the other side - exactly the four inputs the Shotboard Backend wants.

The inputs that matter

• model_task - fl2va or ref2va. The node checks your selection against the filename and warns if they disagree (e.g. you picked ref2va but the model file doesn't say ref2va). Load the wrong family and H3 silently produces junk - the warning is there to save you an hour of that.
• acceleration - native (no Spectrum), spectrum_conservative (default), or spectrum_aggressive. This wraps the model with SpectrumApplyMiniMaxH3 from the ComfyUI-Spectrum-MiniMax-H3 pack. Spectrum is a history-based patch that speeds up sampling; conservative uses a 0.50 blend weight and a small flex window, aggressive goes 0.75 and wider. Start conservative.
• spectrum_history - whether Spectrum's history lives in system_ram or vram. System RAM is the safe default on 12GB cards; VRAM is faster if you have headroom.
• text_encoder_device - auto (default) or cpu_safe_12gb. The Qwen3-VL encoder is the VRAM hog on smaller cards; the safe option forces a GPU-first-then-CPU fallback strategy for 12GB systems.

Where it fits

Wire its four outputs into IAMCCS_MiniMaxH3Backend (or the FLF Parity Conditioning), and feed model_task + report to whatever node wants to know what you loaded. If you're not using GGUF - you have the full fp16/pruned files - this node isn't for you; use a normal loader and skip the quantization overhead.

Installing it

The pack installs from ComfyUI Manager (search "IAMCCS") or git clone https://github.com/IAMCCS/IAMCCS-nodes.git into custom_nodes. But this node is useless without two more packs: ComfyUI-GGUF (city96) for the GGUF loaders, and ComfyUI-Spectrum-MiniMax-H3 (xmarre) if you want the acceleration stage. Install all three and restart.

Gotchas

The acceleration options only work if the Spectrum pack is installed - a node present in the graph doesn't accelerate a path that's bypassed. And the GGUF quality ladder you know from Flux applies here: Q8 is basically indistinguishable from fp16 at half the size, Q5 is the last stop before visible loss. If your model fits at Q8, don't chase Q4 for the VRAM savings you don't actually need.