README.md

---
title: LTX CrossView Camera Control
emoji: 🎥
colorFrom: red
colorTo: pink
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
pinned: false
hardware: zero-a10g
short_description: Re-render a video from a new camera angle via IC-LoRA
python_version: "3.12"
models:
  - Lightricks/LTX-2.3
  - Cseti/LTX2.3-22B_IC-LoRA-CrossView-Prompt
---

# LTX CrossView Camera Control

This Space demonstrates the [Cseti/LTX2.3-22B_IC-LoRA-CrossView-Prompt](https://huggingface.co/Cseti/LTX2.3-22B_IC-LoRA-CrossView-Prompt) IC-LoRA for LTX-Video 2.3.

## What it does

Upload a reference video and pick a new camera angle (azimuth, elevation, distance). The model re-renders the same scene from the requested viewpoint — like a virtual second camera on the same take.

## How it works

- **Base model**: [LTX-Video 2.3 (22B)](https://huggingface.co/Lightricks/LTX-2.3) via the diffusers-converted variant (`dg845/LTX-2.3-Diffusers`)
- **Pipeline**: `LTX2InContextPipeline` with `LTX2ReferenceCondition` for in-context video conditioning
- **LoRA**: applied via standard `pipe.load_lora_weights()` + `set_adapters()`

## Prompt vocabulary

The LoRA was trained on a fixed, discrete camera vocabulary. Prompts follow the template:

```
crossview. new camera angle: {azimuth}, {elevation}, {distance}.
```

Use the dropdowns in the UI to build a valid prompt — the app constructs it automatically.

## Tips

- **Small angle changes work best.** For larger viewpoint shifts, chain multiple small steps (feed the generated view back in as the new reference).
- **LoRA strength** of 1.0–1.5 is recommended. Higher values produce a stronger viewpoint shift.
- Generation takes ~2-4 minutes for a 3-second video at 30 steps.