Skip to main content
Back

ChetiArt's CrossView-Warp Lets MiniMax H3 Reshoot Video From Any Angle

A community LoRA for MiniMax H3 rebuilds any input clip from a new camera angle using depth warping and a Blender-trained novel view synthesis pipeline.

  • Cseti released CrossView-Warp LoRA for MiniMax H3 Ref2VA, enabling camera angle changes on existing clips.
  • Pipeline: MoGe depth estimation, then reprojection into a new viewpoint with magenta holes.
  • The ComfyUI-CrossViewWarp node ships a 3D orbit picker and keyframe-based camera interpolation.
  • Trained on 719 synthetic Blender scenes using an RTX PRO 6000 Blackwell 96 GB on RunPod.
  • 596 MB safetensors, trigger word crossview, distributed under MiniMax H3 Community License.
  • Ports Cseti's earlier LTX-Video 2.3 CrossView-Warp technique to MiniMax's open video model.

Community developer Tamas Cseh, also known as ChetiArt, has released CrossView-Warp v1, a LoRA that conditions MiniMax H3 video generation on a new camera position or movement. It can regenerate an existing clip from another viewpoint without reshooting the footage or rerunning the original simulation.

A LoRA is a compact set of learned weight changes loaded alongside a base model. The 596 MB adapter does not contain MiniMax H3 itself. It ports Cseh’s earlier CrossView-Warp work for LTX-Video 2.3 to MiniMax H3 Ref2VA, the model variant that conditions generation on reference video.

MiniMax H3 also offers FL2VA, which conditions generation on first and last frames. Adapters for FL2VA and Ref2VA are incompatible, according to the project’s compatibility notes, so this release requires the Ref2VA checkpoint.

A geometry pass sets the camera

The workflow combines estimated depth, deterministic reprojection, and generative completion. Its ComfyUI extension uses MoGe, a monocular geometry model, to estimate scene depth from the source footage.

  1. Estimate depth: MoGe predicts the distance of visible surfaces in each source frame.
  2. Choose a camera pose: The CrossViewWarp node accepts an offset or an animated series of camera positions.
  3. Reproject the footage: The node moves visible pixels into the requested view using the estimated geometry. Areas hidden from the original camera remain magenta.
  4. Route both inputs: