Introduction

MiniMax's open-weight H3-Image only landed this week, and it already has a fast mode. A new Hugging Face repo, akhaliq/minimax-h3-image-pdmd-2step, merges the pDMD 2-NFE LoRA into the official 33.1B image transformer so it finishes a picture in 2 denoising evaluations instead of 6, still at the model's native 4096×4096 (16 MP), with no CFG.

It's an image-gen drop with open weights, and you can run it today in three ways: in ComfyUI (a one-file checkpoint or a 2 GB LoRA, plus three ready-made workflows), in diffusers, or in the free demo Space, which renders roughly 1 MP previews while you type.

What shipped

The stock H3-Image checkpoint is step-distilled for a fixed 6-step schedule, and you can't usefully push it lower. At 2 steps it's still mid-denoise, so you get soft, unfinished frames. The pDMD LoRA (from the paper PDMD: Projected Distribution Matching Distillation for Video Diffusion Models, arXiv 2609.35768) is what makes 2 steps land with real texture and detail.

The repo packages it a few ways:

  • minimax_h3_image_pdmd_2step_bf16.safetensors is a 40 GB one-file ComfyUI checkpoint: Comfy-Org's pruned bf16 H3-Image with the LoRA already merged in. No LoRA node needed.
  • minimax_h3_pdmd_2nfe_comfyui.safetensors is the LoRA alone in ComfyUI layout, for people who already have the Comfy-Org checkpoint. It's also the only route for the int8_convrot and w4a8 quantized checkpoints, since those can't take a pre-merge.
  • lora_model_0.safetensors is the original pDMD release with diffusers keys.
  • transformer/ holds the official H3-Image transformer pre-converted to diffusers layout.

The pDMD LoRA itself is Apache-2.0. Anything you render with H3-Image still falls under MiniMax's H3-Image Community License, which has regional and attribution terms, so read it before commercial use.

How fast is it?

The repo's measurements are refreshingly honest:

  • On one H200 at 4096×4096 in bf16, stock H3-Image takes about 25.1 s per image on its 6-step schedule, and the 2-step LoRA takes about 17.4 s, roughly 1.4× faster.
  • In ComfyUI, the author reports about 20 s per 2048² image on an H200 with the bf16 model and an int8 text encoder, with zero unmatched keys against the Comfy-Org repack.
  • On the demo Space (an FP8 build on a ZeroGPU slice), a 2048×2048 render takes about 4.6 s with the detail pass on, versus about 7 s for the official 6-step Space at the same size.
  • In realtime edit mode, the Space lands a 2-step frame in about 1 to 1.5 s, which is why previews can keep up with your typing.

The big win is iteration speed. When you're hunting for a composition or a seed, 2 steps feels close to interactive at 2 to 4 MP.

Run it in ComfyUI

Update ComfyUI to the latest master first. For the one-file route, put the merged checkpoint in ComfyUI/models/diffusion_models/, and grab the text encoder (qwen3vl_32b, int8 or nvfp4) and the VAE from Comfy-Org/MiniMax-H3-Image.

hf download akhaliq/minimax-h3-image-pdmd-2step minimax_h3_image_pdmd_2step_bf16.safetensors --local-dir ComfyUI/models/diffusion_models

If you'd rather keep your existing checkpoint, download just the LoRA:

hf download akhaliq/minimax-h3-image-pdmd-2step minimax_h3_pdmd_2nfe_comfyui.safetensors --local-dir ComfyUI/models/loras

The graph settings that matter:

  • Load Diffusion Model, then MiniMaxH3SigmaShift with shift_video 12 and shift_audio 3, then Basic Guider.
  • Load CLIP with type minimax, and a positive prompt only, since the model is guidance-distilled.
  • EmptyHunyuanLatentVideo at your size. The H3 image latent is a one-frame video latent, and 4096×4096 is native.
  • SamplerCustomAdvanced with euler, scheduler simple, steps 2, denoise 1.0.
  • On the LoRA route, add LoraLoaderModelOnly at strength 1.0. Any other value scales the distillation itself, not a style.

You don't have to wire any of that by hand. The repo's workflows/ folder has three drop-on-canvas graphs: plain 2-step text-to-image, the LoRA version, and a detail-pass version.

The detail pass

A 2-step draft is fast but can look a little soft up close. The repo's fix is a light second pass in stock ComfyUI with no extra files: render a half-size draft at 2 steps, decode it, upscale in pixel space with lanczos, re-encode, then sample again for 3 steps with denoise set between 0.4 and 0.55. Going up to 0.7 adds more detail but can drift or speckle flat areas. For edits, stay at 0.55 or below so the face and layout hold.

For maximum micro-detail there's also a heavyweight option using the Open-H3-Regenerate-2K-IC refiner, but it adds about 33 s per 16 MP image on an H200. It's a quality play, not a speed play.

Prompting tips

H3-Image likes long, complete scene descriptions, and every sample grid in the repo renders prompts as written. MiniMax's own prompt-writing guide splits it into two shapes: for text-to-image, describe the full scene; for edits, say what to preserve, then what to change, then describe the target. A paste-ready starter in that style:

A rain-soaked night market street seen from shoulder height, warm paper lanterns strung overhead, a young woman in a mustard raincoat holding a steaming paper cup of noodles, neon signs reflected in deep puddles, soft mist drifting between food stalls, shallow depth of field, cinematic 35mm photograph, rich teal and amber color grade, crisp detail on rain droplets and fabric texture

Also new: H3-Image in diffusers

The model card also notes that native H3-Image support landed in diffusers main on November 10 via PR #15230, as MiniMaxH3ImageModularPipeline. The LoRA's fit with that new 2D transformer class is untested so far, so stick with the repo's documented route if you want the 2-step results shown on the card.

Conclusion

Days after H3-Image's open-weights launch, the community has already turned its 16 MP model into something you can sketch with. If you've got the VRAM, grab the one-file checkpoint and the ready-made workflow. If you don't, the demo Space is the quickest way to feel how different 2-step previews are. Use 2 steps to find the idea, then let the detail pass finish it.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=minimax-h3-image-2-step-pdmd-lora-comfyui

—Aurelia ♡