CrossView IC-LoRA: Change Camera Viewpoints in LTX-2.3 Video with Prompt Control
Cseti releases a proof-of-concept IC-LoRA for LTX-2.3 that acts as a virtual second camera, letting you change the camera angle of existing video footage using a discrete prompt vocabulary.
Cseti has released CrossView IC-LoRA, a proof-of-concept adapter for LTX-Video 2.3 that acts as a virtual second camera. Give it a reference video and a camera-angle prompt, and it re-renders the same scene from the requested new viewpoint — keeping the subject and content while changing where the camera stands.
The LoRA is trained on synthetic multi-view data (SynCamVideo dataset rendered in Unreal Engine 5), but generalizes to real footage with impressive results. As a v0.9 proof-of-concept, it has clear limits but already produces "quite good" output according to early community tests.
Example results from the model page show scenes re-rendered from different angles — a warrior figure from lower-right, an underwater scene from right-further, a car from right-closer, and more:
| Reference | Generated (prompt: right, lower, closer) |
|---|---|
| Example 1: Warrior | Example 2: Underwater |
| Example 3: Car | Example 4: Rollerskate |
See the full showcase montage on the Hugging Face model page.
How It Works
Unlike traditional camera control methods that require depth maps or 3D geometry, CrossView uses an In-Context LoRA (IC-LoRA) approach. The reference video is conditioned through the model's attention mechanism, and a discrete prompt vocabulary tells the model where to move the virtual camera.
Prompt Vocabulary
The model uses a fixed, discrete camera vocabulary — every prompt must start with the trigger crossview. followed by the template:
crossview. new camera angle: {azimuth}, {elevation}, {distance}.| Axis | Allowed phrases |
|---|---|
| Azimuth (orbit around subject) | same angle · slightly to the left · slightly to the right · to the left · to the right · far to the left · far to the right |
| Elevation (camera height) | lower · same height · higher |
| Distance (to subject) | closer · same distance · further |
Example prompts:
crossview. new camera angle: to the right, lower, closer.
crossview. new camera angle: to the left, higher, further.
crossview. new camera angle: same angle, same height, closer.Usage Tips
- Small steps work best: For large viewpoint shifts, chain several small angle changes — feed the generated view back in as the new reference and apply another small angle
- LoRA strength on distilled models: The LoRA was trained on the full (non-distilled) LTX-2.3. On distilled few-step workflows, try a LoRA strength of 1.2–1.5 or run it in the first non-distilled pass
- Use exact phrases from the vocabulary — the model learned this specific vocabulary; synonyms like "45 degrees left" work less reliably
Training Details
| Parameter | Value |
|---|---|
| Base model | LTX-Video 2.3 (22B) |
| Training framework | ltx-trainer (Lightricks) |
| LoRA rank / alpha | 16 / 16 |
| Target modules | attn1, attn2 (to_k/q/v/out) — attention only |
| Training dataset | 294 curated camera pairs from SynCamVideo |
| Resolution | 768×768×81 @ 15fps |
| License | Apache 2.0 |
Availability
The LoRA is available under the Apache 2.0 license on Hugging Face:
Model: Cseti/LTX2.3-22B_IC-LoRA-CrossView-Prompt
Example workflow: ComfyUI Workflows
Base model: Lightricks/LTX-2.3
Comments
Sign in with GitHub to join the discussion.