Cseti/MiniMax-H3_Ref2VA-LoRA-CrossView-Warp_v1

MiniMax-H3 — Ref2VA LoRA: CrossView Warp v1

Give it a video and a camera offset — azimuth, elevation — and it generates the same scene from that new viewpoint.

This is the CrossView-Warp idea ported to MiniMax-H3. The model reads two videos: a depth-warp of your clip, which carries the geometry, and the clip itself, which carries identity and appearance. The warp comes from the CrossViewWarp ComfyUI node.

Examples

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

Prompt

crossview

The compare_* clips show the warp, the original and the result side by side.

Usage (ComfyUI)

LoRA strength: 0.8 - 1.0 (0.8 is recommended)

Trigger word: crossview (note: You can also influence the generation by prompting what you want to see in the unseen (magenta) areas.)

Example ComfyUI workflow: crossview-warp-h3.json

The CrossViewWarp node has a walkthrough video that covers the camera controls; everything there applies unchanged:

The steps below are what it wires up.

1. Install the ComfyUI-CrossViewWarp node.
2. Feed your clip to Run MoGe Inference, and its geometry to CrossView Warp.
3. Load MiniMax-H3_Ref2VA-LoRA-CrossView-Warp_v1_3500.safetensors as a LoRA on the ref2va checkpoint.
4. Warp -> MiniMax H3 Add Guide, frame_idx = 0. Source -> MiniMax H3 Reference To Video
5. Resize warp and source to the same size before both nodes.
6. The trigger word: crossview. You can also influence the generation by prompting what you want to see in the unseen (magenta) areas.

Settings the released examples were generated with:

Frames	124
Resolution (1st pass)	0.5 MP, 16:9
Resolution (2st pass)	1.5 MP, 16:9
Sampler / steps	res_multistep, 8 steps with the DMD 8-step turbo LoRA
Sigma shift	12 video / 3 audio
LoRA strength	0.8

Dataset details

The same 719 Blender scenes as the LTX v2 release. If you want to know more about the dataset you can find more information in my LTX crossview-warp repo:

Third-party assets in the renders (CC-BY 3D models, CC0 HDRIs and textures, CMU motion capture) are listed in ATTRIBUTION.md.

Training

Trained on RunPod — NVIDIA RTX PRO 6000 Blackwell, 96 GB.

Base model	MiniMax-H3, ref2va
Framework	AkaneTendo25/musubi-tuner, branch minimax-h3
Strategy	Ref2VA, aligned_guide_indices = [0]
Released checkpoint	step 3,500 of 6,000
LoRA rank / alpha	32 / 32
Optimizer	adamw8bit 1e-4
Base precision	INT8 ConvRot
Resolution	512×512 × 73 frames
Schedule	6,000 steps, batch 1, no gradient accumulation
Measured	20.35 s/step, 33 h 55 min, 38.4 of 95.6 GB peak

License

The LoRA weights are a Model Derivative of MiniMax H3 and are distributed under the MiniMax H3 Community License Agreement

Support

Everything here is open, and the GPUs behind it are rented. If this was useful, please consider supporting my work: