--- license: apache-2.0 base_model: - MiniMaxAI/MiniMax-H3 tags: - lora - video - comfyui - minimax-h3 --- # Minimax H3 LoRAs for ComfyUI LoRAs for [MiniMax H3](https://huggingface.co/MiniMaxAI/MiniMax-H3), built to run in ComfyUI with the [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) weights. Trained and tested mainly against the `ref2va` base model; `fl2va` should also work as a base but is less tested. ## Repository structure ``` loras/ minimax_h3_lms_v1.0_r64.safetensors # the LMS LoRA workflows/ minimax_h3_lms_workflow.json # ComfyUI workflow set up for the LoRA above examples/ comparison-1.mp4 # before/after, no audio comparison-2-audio.mp4 # before/after, with audio comparison-3-audio.mp4 # before/after, with audio comparison-4.mp4 # before/after, no audio comparison-5.mp4 # before/after, no audio comparison-6.mp4 # before/after, no audio comparison-7.mp4 # before/after, no audio comparison-8.mp4 # before/after, no audio ``` ## `minimax_h3_lms_v1.0_r64` — "a little more sharpness" Rank-64 LoRA for MiniMax H3 `ref2va` that sharpens a source video while keeping it photorealistic. It conditions on the source through guide latents rather than through the model's native reference-video node. **Trigger / caption:** ``` Enhance this video with sharp, crisp details while preserving a natural photorealistic appearance. ``` ### How the guide works This LoRA conditions on your source video as a latent guide, not as a text description of it. The source clip is encoded by the video VAE and packed into the transformer's sequence as a conditioning block that is aligned to the target timeline. Three properties define the arrangement: - **Same temporal origin.** The guide's clock starts where the target's starts, so guide latent frame `i` sits at the same position as target latent frame `i`. - **Same spatial grid.** The guide is encoded at the target's resolution, so guide token `(t, y, x)` lands on target token `(t, y, x)`. - **It does not advance the reference clock.** Unlike an ordinary reference block, an aligned guide occupies the target's own timeline rather than being appended before it. The result is that attention between guide and target costs nothing positionally — the correspondence is handed to the model instead of being something it has to search for. This is the same arrangement used by in-context video LoRAs on LTX. The guide is held near-clean during training (about 0.1% noise augmentation) while the target is noised normally, so the model learns to map one to the other rather than to denoise both. The text encoder never sees the guide. Only the caption reaches it; the video reaches the transformer purely as latents. The model therefore learns a pixel-level correspondence with the source, not a paraphrase of it. ### Using it Feed the source clip into the guide input, anchored at frame 0, at the same resolution as the output. Both the guide and the target must land on the model's valid clip lengths — `17k + 5` frames (5, 22, 39, 56, 73, 90, 107, 124, …). A guide that is shorter than the target, or at a different resolution, breaks the alignment. In ComfyUI this is the native `MiniMaxH3AddGuide` node with `frame_idx = 0` — it ships with ComfyUI's own MiniMax H3 support, nothing extra to install for it. This is a different mechanism from the also-native `MiniMaxH3ReferenceToVideo` node, which passes a reference block with its own clock instead of an aligned guide. This LoRA was trained specifically for `ref2va`, not `fl2va`, and primarily to use guide latents alone, so treat `MiniMaxH3ReferenceToVideo` as optional/experimental here: combining it with `Add Guide` can be worth trying, but expect different — not necessarily better — results, since you'd be feeding two different kinds of conditioning at once. In principle this LoRA is meant for a second pass (sharpening an already-generated clip) rather than direct generation from scratch, though it can be used for the latter too — just expect it to behave differently. The included workflow ([`workflows/minimax_h3_lms_workflow.json`](workflows/minimax_h3_lms_workflow.json)) additionally wires in `AIToolkitMiniMaxH3RefVideo`, a node from [ostris/ComfyUI-AIToolkit-MiniMaxH3](https://github.com/ostris/ComfyUI-AIToolkit-MiniMaxH3) — clone that repo into your ComfyUI `custom_nodes/` folder for the workflow to load as-is. This LoRA was trained with [ai-toolkit](https://github.com/ostris/ai-toolkit)'s custom fork that added guide-latent support, which is where this node comes from. As far as I can tell it just resizes the clip to the target resolution before feeding it in, so if you'd rather not add the dependency, swap it for any other resize step feeding into `Add Guide`. ### Examples ## License Apache 2.0. Base model credit to [MiniMaxAI](https://huggingface.co/MiniMaxAI/MiniMax-H3) and the ComfyUI-ready weights from [Comfy-Org](https://huggingface.co/Comfy-Org/MiniMax-H3). Guide-latent training support and the ComfyUI node credit to [ostris](https://github.com/ostris) ([ai-toolkit](https://github.com/ostris/ai-toolkit), [ComfyUI-AIToolkit-MiniMaxH3](https://github.com/ostris/ComfyUI-AIToolkit-MiniMaxH3)).