LTX 2.3 CrossView Prompt IC-LoRA: Virtual Camera Control for Video Generation

A new community-trained In-Context LoRA (IC-LoRA) for LTX-Video 2.3 brings virtual camera control to AI video generation. Created by community developer Cseti, the CrossView Prompt IC-LoRA acts as a "virtual second camera" — it takes a reference video and a short camera-angle prompt, then re-renders the same scene from the requested new viewpoint.

How It Works

The LoRA uses a fixed, discrete prompt vocabulary based on three camera axes:

• Azimuth — orbit around the subject (same angle, slightly to the left/right, to the left/right, far to the left/right)
• Elevation — camera height (lower, same height, higher)
• Distance — proximity to subject (closer, same distance, further)

Every prompt must start with the trigger word crossview. followed by the template:

crossview. new camera angle: to the right, lower, closer.

This gives access to 63 valid camera-angle combinations, enabling precise viewpoint shifts in one step.

Example Outputs

The LoRA produces side-by-side comparisons showing the original reference video (top) alongside the generated new camera view (bottom). Example output videos demonstrate camera shifts including:

• Moving the camera to the right, lower, and closer to the subject
• Shifting to the left, higher, and further away
• Changing elevation from above to below the subject

ComfyUI Integration

The LoRA is designed for ComfyUI video-to-video IC-LoRA workflows. Usage is straightforward:

1. Load LTX2.3-22B_IC-LoRA-CrossView-Prompt_v0.9_13700.safetensors as a LoRA in ComfyUI
2. Provide a reference video (the scene you want to re-shoot)
3. Write a camera-angle prompt using the fixed vocabulary
4. Generate

An example workflow is available on the companion dataset page.

Tips for Best Results

• Small angle changes work best — the model is most reliable on single-step shifts. For a larger viewpoint change, chain several small steps: feed the generated view back in as the new reference and apply another small angle.
• Distilled model users — on few-step distilled workflows, try a higher LoRA strength (1.2–1.5) or run it in the first non-distilled pass.
• The model was trained on LTX-Video 2.3 (full, non-distilled) with 294 synthetic camera pairs from the KlingTeam SynCamVideo dataset.

Training Details

Download

The LoRA is available on Hugging Face under the Apache 2.0 license:

Cseti/LTX2.3-22B_IC-LoRA-CrossView-Prompt