Download
1 variant available
bf16 SafeTensor
ltx-face-prior-f1-profile-correction-step11019.safetensors
BF16, good balance • 96.1 MB
Verified: a day ago
First public release.
Profile-correction pass on top of the base face-prior run. This checkpoint
specifically improves side and three-quarter angles, where earlier steps still
let the jawline and nose bridge drift during camera movement.
Recommended strength: 1.0.
Show more

450 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
1100 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
License:
LTXV2
## East Asian Facial Fidelity LoRA for LTX-2.3
This LoRA is designed to improve the natural appearance and identity consistency of East Asian female faces in LTX-2.3 Image-to-Video generation.
It focuses on reducing facial drift during large camera movements and correcting exaggerated eye size, visible sclera, iris proportions, and double-eyelid geometry in full side-profile views. Its effects are especially noticeable in front-to-profile transitions and Camera Orbit shots.
### Key Features
- Improved facial consistency during large camera-angle changes
- More natural eye geometry in left and right full side profiles
- Better stability of the nose, lips, chin, and facial silhouette
- Helps preserve the source subject, hairstyle, and clothing
- A general East Asian female face prior, not a character-specific LoRA
- No Trigger Word required
### Usage
- Base Model: LTXV 2.3
- Recommended Task: Image-to-Video
- Recommended LoRA Strength: 1.0
- Trigger Word: None
Place the file in ComfyUI/models/loras and apply it through a LoRA Loader in your existing LTX-2.3 workflow. You can use ordinary subject phrases such as woman or young woman without a special activation token.
The published LORA OFF / LORA ON comparisons use identical source images, prompts, seeds, sampling settings, Image Guide Strength, resolution, and camera movement. The application of this LoRA is the only intended difference between each pair.
### Training Information
The LoRA was trained with the official LTX Trainer and the LTX-2.3 22B Dev model. Following the initial general face-prior stage, an additional correction stage was completed using 1,127 manually selected full side-profile images.
- Final Step: 11,019
- Rank / Alpha: 16 / 16
- Learning Rate: 5e-5
- Precision: BF16
The most thoroughly verified use case is LTX-2.3 Image-to-Video generation featuring East Asian female subjects.
This is a standard LoRA, not an IC-LoRA, ControlNet, full checkpoint, or character-specific identity model.
### License
This is a derivative model based on LTX-2.3 and is subject to the LTX-2 Community License Agreement.
https://huggingface.co/Lightricks/LTX-2.3/blob/main/LICENSE
ㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡ

