H3 • Models

What this node actually is

MiniMax H3 is a 33B video model - omni-modal, native stereo audio, open weights from August 2026 behind a community licence that geofences the US, EU, UK and Korea out of the local-weights path. It is not an image model. But the community clocked the workaround early: run a video model as an image model by generating one frame, and it holds up better than you'd expect. That comment thread is the entire premise of this pack.

H3 • Models is the pack's single loading point. One diffusion model, one Qwen encoder, one video VAE, shared by both text-to-image and image-edit branches. Everything downstream - generation, editing, LoRA chains - takes the models bundle this node outputs.

Worth knowing: it is a loader, not a pipeline. It loads models; it renders nothing.

How it works (and why that matters)

The node doesn't reimplement anything. Under the hood it calls ComfyUI's own UNETLoader, CLIPLoader (with the minimax type) and VAELoader. That's deliberate, and it's good news: you get ComfyUI's native quantization handling, offloading and model cache for free instead of fighting a bespoke loader that keeps a second copy of a 30GB checkpoint in VRAM.

The three dropdowns aren't free-text - they're filtered lists scraped from your model folders:

• diffusion_model needs a filename containing both h3 and minimax
• text_encoder needs qwen and minimax
• video_vae needs minimax_h3_video_vae

That filter is the single most common "the node is broken" post. Your file is almost certainly fine; it's just not in the folder the combo scans, or it's named something the filter skips.

The inputs that matter

You set three things, and for the standard setup you set them once:

diffusion_model - the H3 checkpoint. The default the pack looks for is minimax_h3_fl2va_pruned_int8_convrot.safetensors. FL2VA is the starting model for both workflows, including editing; REF2VA is an optional comparison, not a requirement.

text_encoder - the Qwen3-VL text/vision encoder, qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors. Note it's a vision encoder too: that's what lets the Image Edit node feed reference pictures through it.

video_vae - the official H3 video VAE, INT8 ConvRot or FP16. Yes, video VAE for stills. The pack ships its own single-frame decoder and the VAE is what that decoder drives. You do not need an image VAE, a hybrid checkpoint, an audio VAE, or the old Qwen detail refiner. People burn hours downloading those because older guides told them to.

There's one optional input, turbo_lora, tucked in the advanced section. It's a compatibility shortcut that applies Larry's v4 adapter at a fixed 0.38 - it only ever lists files matching minimax_h3_turbo_v4_step600_ema. If you're applying that adapter through an H3 • LoRA node instead, leave this on None. Setting both applies it twice, and doubled adapter strength does not look like a slightly-too-strong LoRA. It looks like a broken image.

Install

Through ComfyUI Manager (search "MiniMax H3 Image Studio"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/astropuzzo/ComfyUI-MiniMax-H3-Image-Studio

Restart ComfyUI, then hard-refresh the browser. You need ComfyUI 0.37.0 or newer - that's not a soft suggestion, the pack's dynamic picture sockets depend on it. 0.38.0 is what the author validated against.

Dependencies: none. The repo's requirements.txt says so explicitly - "No extra runtime dependencies. Uses PyTorch and ComfyUI's native MiniMax H3 support." No PyAV, no weird wheels, no dependency roulette. That is genuinely rare for a pack this ambitious and it's the main reason to install it rather than hand-build the graph.

Then download the three models from Comfy-Org/MiniMax-H3 and drop them in models/diffusion_models/, models/text_encoders/ and models/vae/. Pick the quantization your CUDA/PyTorch combo actually supports - INT8 ConvRot and NVFP4 want reasonably current hardware. Reuse anything already on disk; there's no download script to run.

Where people get burned

Empty dropdowns. Files aren't there, or names don't match the filters above. Restart and refresh before you start editing filenames.

The licence. One thing worth knowing before you pull 40GB over your connection: the H3 Community License excludes the United States, EU, UK and South Korea, and the exclusion covers outputs, not just weights. MiniMax's hosted Hailuo API stays globally available; the local file path does not. That's a real constraint to decide on, not a footnote.

Load errors mentioning VAE decoder methods. That's the pack telling you it found the wrong VAE. Use the official H3 video VAE and update ComfyUI.

Wire the output into H3 • Text to Image or H3 • Image Edit, or into the first of your H3 • LoRA nodes.