MiniMax-H3-Image Two-Step Image Generation LoRA Packaged Version

A community author has adapted the 2-step LoRA (rank 128, covering attention and feed-forward layers across all 50 blocks) released in the pDMD paper (arXiv 2609.35768) for use with MiniMax-H3-Image: the original model was distilled using a fixed 6-step schedule, and images had not yet converged at 2 steps; after merging this LoRA, complete images can be generated with just 2 denoising steps at the native resolution of 4096×4096. The repository provides four types of files: a LoRA with diffusers key names, a LoRA with ComfyUI key names, a 40GB ComfyUI single-file checkpoint with the LoRA already merged, and the official transformer converted to diffusers format. The author states that on a single H200 GPU, the original 6-step generation at 4096×4096 takes approximately 25.1 seconds, while the 2-step version takes approximately 17.4 seconds.

Original title: akhaliq/minimax-h3-image-pdmd-2step

A community author has adapted the 2-step LoRA (rank 128, covering attention and feed-forward layers across all 50 blocks) released in the pDMD paper (arXiv 2609.35768) for use with MiniMax-H3-Image: the original model was distilled using a fixed 6-step schedule, and images had not yet converged at 2 steps; after merging this LoRA, complete images can be generated with just 2 denoising steps at the native resolution of 4096×4096. The repository provides four types of files: a LoRA with diffusers key names, a LoRA with ComfyUI key names, a 40GB ComfyUI single-file checkpoint with the LoRA already merged, and the official transformer converted to diffusers format. The author states that on a single H200 GPU, the original 6-step generation at 4096×4096 takes approximately 25.1 seconds, while the 2-step version takes approximately 17.4 seconds.

Why it matters The original model uses a fixed 6 steps and cannot be reduced; two-step generation allows for real-time previews while typing, faster seed changes and prompt edits, and includes three ComfyUI workflows

Upvotes 47 · Likes 47 · Downloads 17989

Reasons

  • Two paths for ComfyUI: directly use the single-file checkpoint (no LoRA node required), or add the 2GB LoRA to the official Comfy-Org package with a fixed strength of 1.0; the author states it has been verified on the latest ComfyUI with all key names matching
  • Includes a detail enhancement workflow: 2-step sketch at half size → pixel upscaling → re-encoding → 3-step resampling; starting noise between 0.4–0.55 preserves composition, while 0.7 yields more detail but may cause drift
  • The demo Space akhaliq/minimax-h3-image-2step allows viewing real-time previews of approximately 1MP while inputting

Usage

  • ComfyUI: Place minimax_h3_image_pdmd_2step_bf16.safetensors in models/diffusion_models; use the text encoder and VAE from Comfy-Org/MiniMax-H3-Image; set MiniMaxH3SigmaShift shift_video to 12, sampler to euler, scheduler to simple, and steps to 2; or directly drag in the ready-made workflows from the repository's workflows/ directory
  • diffusers: lora_model_0.safetensors can be merged into MiniMaxH3Transformer3DModel; use scheduler.set_shift(12.0) and set_timesteps(3); no CFG or negative prompts are needed

Risks

  • The LoRA itself is Apache-2.0, but generated content remains subject to the MiniMax-H3-Image community license (which includes regional and attribution clauses)
  • Example images were rendered using the P-H3-Image pipeline; the author states they have not yet tested whether the key names of the newly merged official MiniMax-H3-Image pipeline in diffusers match this LoRA
  • Quantized int8_convrot / w4a8 checkpoints cannot be pre-merged and must use a LoRA node; this item has not been run

Open

Hugging Face Trending

Explore the latest AI Radar updates