CoolFace
Apppublic

Cseti/ltx2-crossview-warp-demo

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes
App README

LTX CrossView Warp

This Space demonstrates the Cseti/LTX2.3-22B_IC-LoRA-CrossView-Warp IC-LoRA for LTX-Video 2.3 (22B).

What it does

Upload a source video and set a camera offset (azimuth / elevation / distance). The model re-imagines the same scene from the new viewpoint. Unlike the CrossView Prompt LoRA which takes camera angles as text, this one takes them as precise numbers — the depth-warp conditioning tells the model exactly where to move the camera.

How it works

The pipeline reproduces the ComfyUI-CrossViewWarp workflow in diffusers:

  1. 1.Depth estimation: Depth Anything V2 Large computes per-frame depth maps.
  2. 2.Depth-warp: The source video is reprojected into the target camera viewpoint using monocular depth, with magenta disocclusion holes where the model must hallucinate content. This is a faithful port of the CrossViewWarp ComfyUI node's warp math.
  3. 3.IC-LoRA generation: Both the warp video and the original source video are passed as LTX2ReferenceCondition references to the LTX2InContextPipeline, which conditions the generation on both signals.

Camera angle ranges

ZoneAzimuthElevationWhat to expect
Greenup to ±45°up to +30° / −15°Reliable
Yellow±45–65°+30–40° / −15–25°Usable
Beyond>±65°>+40° / <−25°Outside training range

The zones combine — a large azimuth plus a large elevation is harder than either alone.

Tips

  • —Small angle changes work best. The model steers rather than fully reprojects.
  • —LoRA strength of 1.3 is recommended (the default).
  • —Distance > 1 often needs a prompt describing the newly revealed content.
  • —Generation takes ~2-4 minutes for a 2-second video at 30 steps.

Credits