A Two-Step LoRA Cuts MiniMax-H3-Image Generation Latency by 31%

A new LoRA adapter reduces MiniMax-H3-Image generation from six denoising evaluations to two. At the model’s native 4096×4096 resolution, the project reports latency falling 31%, from 25.1 seconds to 17.4 seconds on an NVIDIA H200. One-megapixel previews take roughly 1 to 1.5 seconds, although fine textures remain softer than those from longer or refined runs.

The release checkpoint packages the source LoRA for the 33-billion-parameter base checkpoint. The rank-128 update covers attention and feed-forward layers across all 50 transformer blocks. It targets 4096×4096 output and runs without classifier-free guidance, an extra conditional and unconditional comparison that guidance-distilled models no longer require.

Two evaluations require a trained endpoint

A network-function evaluation, or NFE, is one transformer pass at a selected noise level. The stock MiniMax-H3-Image checkpoint uses a fixed six-step PDD schedule. Running that checkpoint for only two evaluations leaves it midway through denoising, with recognizable composition but soft contrast and unfinished detail.

The adapter trains the model to reach the image endpoint after two passes on a simple flow-matching grid, which selects the two noise levels where the transformer runs. Adapters trained for other step counts target different endpoints and are not interchangeable.

Projected Distribution Matching Distillation modifies Distribution Matching Distillation, a method that teaches a low-step student to reproduce a teacher model’s output distribution. According to the