MiniMax H3 LoRAs for ComfyUI

Experimental LoRAs for MiniMax H3, packaged for ComfyUI and tested mainly with the ref2va weights.

These are research models, not finished tools. Results vary substantially with the source clip, resolution, prompt and reference images.

Models

Motion Transfer and TARA

Four downloadable workflows: Guide Latents / native, Guide Latents / Planner, TARA / native, TARA / Planner.

References are 1..N pictures with roles assigned by the prompt, not a required person/background pair. Several images can describe one identity, multiple characters, outfits, objects, scenery or style. A separate background image is optional. See the Motion Transfer guide for model setup, multiple-reference examples and the modified temporal Comfy Kitchen kernel required for TARA's fused one-pass Kitchen path. The TARA workflows default to Flash partitions, which does not require that extension.

Training adapter V2. The Motion Transfer LoRAs were trained on top of ostris' training adapter V2 (minimax_h3_ref2va_training_adapter_v2.safetensors) kept frozen. Since 2026-11-02 the two files above carry it: LoRA x1 + adapter x1 merged by exact rank concat, then per-module SVD keeping 99% of the Frobenius energy (185 MB and 126 MB instead of 620 MB + 310 MB). Load only turbo + motion LoRA; do not load the adapter again. In same-seed tests they matched the full LoRA + separate adapter, including TARA's frame sync across camera cuts. The previous files (without the adapter) are kept in loras/base_loras_ignore/.

TARA-H3 — Temporally Aligned Reference Attention. Native reference conditioning alone did not reliably reproduce the timing of a source performance in H3, while aligned guide latents (Add Guide) copy it frame by frame. TARA keeps the motion video as native reference tokens, translates its temporal RoPE coordinates onto the generated video's clock, and adds a bounded temporal penalty only to generated-video queries attending to motion-reference keys; image references keep their native coordinates and the attention change adds no learned layers. Motion Transfer TARA v1.0 was trained for this method. Node: BFSMiniMaxH3NativeTemporal in ComfyUI-BFSNodes. Technical report (v0.1.2, experimental, not peer reviewed): TARA-H3: Temporally Aligned Reference Attention for Native Video Conditioning.

All of these run in Henge, the open-source studio built on them.

Latent Upscaler

h3_upscaler_lms_v0.1 is the recommended FP16 release of an experimental latent-upscaler checkpoint. It was trained for 2,000 steps with the same sharpness dataset used for the LMS LoRA and initialized from the LBH-123-AI/Minimax_h3_latent_Upscaler project and its minimax_h3_latent_upscaler_3d_fp16.safetensors checkpoint.

See the Latent Upscaler guide for setup, scope and limitations.

Still images with Pruna P-H3-Image (guide-latent LoRAs)

PrunaImageH3WorkaroundGuideLatents.json is a workaround to use PrunaAI's P-H3-Image with MiniMaxH3AddGuide, so LoRAs trained on guide latents can edit a single image. PH3ImageConditioning with no images gives the one-frame H3 latent, MiniMaxH3AddGuide adds the image to edit as the guide at frame_idx = 0, MiniMaxH3ReferenceToVideo (length 5) carries the prompt and the reference picture, and the P-H3-Image VAE decodes the single frame. The model stack is the H3 ref2va weights, the Pika turbo LoRA, the P-H3-Image LoRA and the task LoRA, at 8 steps with BasicGuider.

Tested with Character Swap v1.1; Head Swap and the other guide-latent LoRAs load the same way. Keep the image size a multiple of 32 (the workflow resizes to a 2048 px long side).

Common setup

Place LoRAs in ComfyUI/models/loras/. The included workflows list their required models and custom nodes inside the graph. Keep the output aspect ratio matched to the source clip and use an H3-supported frame count (17n + 5: 5, 22, 39, 56, 73, 90, 107, 124, ...).

The LMS, Style Transfer, VFX Edit and Head Swap models use an aligned source-video guide through MiniMaxH3AddGuide at frame_idx = 0. This is different from a native image/video reference: the guide shares the generated video's spatial and temporal grid.

License and credits

Licenses are specified by asset. The separate Character Swap license and Motion Transfer license apply only to the Character Swap and Motion Transfer LoRA weight files listed in LICENSES.md; they do not apply to the other LoRAs, workflows, docs, or examples. Use of these LoRAs, including commercial use in videos and films, requires no prior permission. Organizations with over US $1,000,000 in annual gross revenue must credit its origin for commercial use. Offering the LoRA as a hosted API, SaaS, or other remote model service is not allowed. See the full Character Swap and Motion Transfer licenses and check each other asset's applicable terms before use.

Base model credit to MiniMaxAI, ComfyUI-ready weights to Comfy-Org, and guide-latent training support to ostris/ai-toolkit and ComfyUI-AIToolkit-MiniMaxH3.