Summary
Veteran AI introduces CrossView Warp, a ComfyUI technique for repositioning the virtual camera in an already-generated video without needing to recreate the original shot from scratch. The tutorial covers the full setup: cloning the CrossView Warp extension into ComfyUI’s custom_nodes directory, installing Depth Anything V2 for spatial understanding, downloading the corresponding IC-LoRA weights, and optionally running the workflow on RunningHub for faster iteration.
The core workflow runs on the distilled LTX 2.3 model and uses a two-branch IC-LoRA structure. The first branch feeds the original video directly to preserve character identity and motion; the second processes the video through Depth Anything V2 to generate a depth map, which then passes through the CrossView Warp node to produce a spatially-warped version. This warped output — which looks distorted, with pink areas representing regions the original camera never captured — is what guides LTX 2.3 to reconstruct the missing visual information from the new viewpoint.
Camera positioning is controlled via a sphere interface color-coded by confidence: green zones produce reliable results, yellow zones introduce creative latitude but less consistency, and gray zones fall outside the training distribution entirely. The tutorial walks through left/right/overhead/low-angle tests across multiple source videos, including two-person scenes and shots with pre-existing camera motion, honestly documenting where character consistency degrades and where the technique holds up cleanly. An online workflow is available on RunningHub for viewers who want to test without a local ComfyUI installation.
📺 Source: Veteran AI · Published August 03, 2026
🏷️ Format: Tutorial Demo
