MiniMax H3 Sharpness LoRA: Detail Recovery for Video

A new sharpness LoRA for MiniMax H3 takes a different route to video detail enhancement: instead of upscaling, it runs a same-resolution v2v pass that adds crisp, photorealistic detail back into soft or over-smoothed H3 output. The LoRA, minimax_h3_lms_v1.0_r64, was trained and published by community trainer Alissonerdx on Hugging Face, where it has collected over 70 likes and 1,300 downloads in its first days and become one of the most discussed tools in the H3 community.

How it works: guide latents, not reference video

The rank-64 LoRA conditions on the source video through guide latents rather than the model's native reference-video node. The source clip is VAE-encoded and packed into the transformer's sequence as a conditioning block aligned to the target timeline: same temporal origin, same spatial grid, and it does not advance the reference clock. This is the same in-context arrangement used by video LoRAs on LTX, and it hands the pixel-level correspondence to the model directly instead of making it search for one.

Practical consequences:

• The guide is trained near-clean (about 0.1% noise augmentation) while the target is noised normally, so the model learns a source-to-target mapping rather than denoising both
• The text encoder never sees the guide, only the caption
• Guide and target must sit on valid clip lengths (17k + 5 frames: 5, 22, 39, 56...) at the same resolution
• A shorter or differently-sized guide breaks the alignment

In ComfyUI this uses the native MiniMaxH3AddGuide node (frame_idx = 0) that ships with ComfyUI's own H3 support, so there is nothing extra to install for the conditioning path. The LoRA was trained specifically for the ref2va variant and is primarily intended as a second pass on already-generated clips, though it can be used for direct generation too.

Community results

The LoRA circulated on the Banodoco Discord before and after its release, where testers converged on a few usage notes:

• It is not an upscaler: the sharpening works at the same resolution as the source, so pair it with an upscale workflow if you need more pixels
• Strength around 0.8 keeps skin and textures natural; pushing to 1.0 in both passes goes extreme
• It requires the Add Guide node to work, since the training relied on guide latents
• It also runs on H3 as a detail pass with LTX-style conditioning, per cross-model tests from other trainers

A second experimental sharpness LoRA from trainer NRDX followed a similar build and showed comparable extreme sharpening at strength 1, confirming the guide-latent approach works for this task class.

Download H3 sharpness LoRA workflow

The official workflow wires in AIToolkitMiniMaxH3RefVideo from ostris/ComfyUI-AIToolkit-MiniMaxH3 for the guide input. If you would rather not add that custom node, any resize step feeding into Add Guide works, since the ostris node just resizes the clip before the guide encoding.

Availability

• LoRA: Alissonerdx/Minimax-H3-ComfyUI on Hugging Face, loras/minimax_h3_lms_v1.0_r64.safetensors (1.2 GB, Apache-2.0)
• Workflow: bundled in the same repo under workflows/
• Base model: Comfy-Org/MiniMax-H3 ComfyUI weights