This repository adapts the pDMD 2-NFE LoRA (rank 128, covering attention and feed-forward layers across all 50 blocks) released by upstream pdmd2026 for use with MiniMax H3 Image. It provides LoRAs with key names compatible with both diffusers and ComfyUI, as well as a single-file version where the LoRA is pre-merged into the Comfy-Org slim bf16 checkpoint (approx. 40 GB), eliminating the need for a LoRA node. The author states it has been verified on the latest ComfyUI, generating one 2048² image in about 20 seconds on H200, and includes an online demo space and two ready-to-use workflows.
Why it matters The official checkpoint requires a fixed 6 steps; switching to 2 steps significantly speeds up feedback when modifying prompts or changing seeds, making it ideal for quickly testing compositions.
Hugging Face Likes 30 · Hugging Face Trending Score 30 · Hugging Face Downloads 0
License: Conditional commercial use (base model restrictions) · ComfyUI: Mentioned in model card · VRAM: Not specified
Regional License: The MiniMax H3 open-weight license excludes the EU, UK, South Korea, and the US (Excluded Territories in the LICENSE); users in these four regions must obtain separate authorization from MiniMax.
Reasoning
- The author clearly outlines the trade-offs between the single-file and LoRA versions: quantized checkpoints cannot be pre-merged, so only the LoRA version can be used
Usage
- Place the single-file version in models/diffusion_models, and configure it according to the model card with MiniMaxH3SigmaShift (shift_video 12), euler + simple, and 2 steps; for those with the official checkpoint, use LoraLoaderModelOnly to load the LoRA, keeping the strength at 1.0
Risks
- The LoRA itself is Apache-2.0, but the base model is subject to the MiniMax H3 Image license, which includes revenue and regional restrictions
- The LoRA weights are byte-for-byte identical to the upstream pDMD release; this repository is for organization and merging; at 2 steps, the composition may deviate from the original 6-step version