Summary
Veteran AI demonstrates CrossView, an IC-LoRA built on top of LTX Video 2.3, which allows filmmakers and video creators to virtually reposition the camera on footage that has already been captured—without 3D reconstruction and without reshooting. The technique uses the entire original video as a reference clip, then follows a structured three-axis camera prompt to regenerate the same performance from a new angle. The three axes control lateral orbit (left/right around the subject), height (higher/lower), and distance (closer/farther), yielding 63 distinct combinations.
The tutorial walks through a full ComfyUI workflow hosted on RunningHub, covering model loading (LTX 2.3 v1.1 main model plus CrossView LoRA at strength 1.5), dual-prompt design (the first prompt drives camera repositioning; the second acts as a content consistency anchor for the tiled upscaling stage, which processes the output in tiles regardless of output resolution), and latent upscaling by 2x before a Tiled Latent Sampler reconstructs fine detail.
Test results span four scene types: a sword-fighting female warrior in a rainy courtyard, a walking figure in an ancient corridor, a robotic arm in a laboratory, and a flying dancer with ribbon effects. Fixed-camera single-subject scenes show the most reliable camera control, while complex multi-character or high-motion scenes show greater scene reinterpretation at larger angle changes. The creator notes left/right orbiting and height adjustments are more stable than extreme close-up generation, and recommends using the official prompt vocabulary rather than paraphrasing.
📺 Source: Veteran AI · Published July 15, 2026
🏷️ Format: Tutorial Demo
