Alissonerdx/BFS-Best-Face-Swap-Video

⚠️ Ethical Use & Disclaimer

This model is a technical tool for Digital Identity Research, Professional VFX Workflows, and Cinematic Prototyping.

By downloading or using this LoRA, you acknowledge and agree to the following terms:

• Intended Use: This tool is designed for filmmakers, VFX artists, and researchers exploring high-fidelity video transformations.
• Consent & Rights: You are strictly required to possess all necessary legal rights and explicit consent from any person whose likeness is being processed.
• Legal Compliance: You must comply with all local and international laws regarding digital identity and synthetic media.
• Liability Waiver: This model is provided "as is." As the creator (Alissonerdx), I bear no responsibility for how this tool is used by third parties. Any legal, ethical, or social consequences resulting from the content you generate are your sole responsibility.

---

📺 Video Examples

Generated using the Frame 0 Anchoring technique. All examples follow the guide video motion while preserving the identity provided in the first frame.

Example 1	Example 2

Example 3	Example 4

Example 5

---

🛠 Technical Background & Script Modifications

To achieve this level of identity transfer, I heavily modified the official LTX-2 training scripts. Key improvements include:

• Novel Conditioning Methods: New ways to inject reference data into the latent space.
• Noise Overhaul: A complete rebuild of the noise scheduling logic, implementing a custom High-Noise Power Law distribution. This forces the model to prioritize target identity over the guide video's original context.
• Compute: Trained for 60+ hours on NVIDIA RTX PRO 6000 Blackwell GPUs, iterating through over 300GB of experimental files.

📊 Dataset Specifications

• Volume: 300 high-quality head swap video pairs.
• Base Resolution: Trained on 512x512 buckets.
• Aspect Ratio: Primarily Landscape data. Horizontal videos will yield significantly better results than portrait ones.
• Framing: Optimized for Close-ups (face near the camera). Wide shots may result in lower identity fidelity.

💡 Inference & Tuning Tips

CRITICAL: The Frame 0 Requirement

This model was specifically trained to use the first frame to condition the entire generation.

• Initial Swap: For 100% accuracy and similarity, it is essential to perform a high-quality head swap on Frame 0 before running the inference.
• Cutouts: While the model might work if you simply paste a photo over the face in the first frame, it will not be 100% effective in all cases.
• Pro Tip: Use my previous BFS Image Models to prepare a high-quality Frame 0 swap for the best results.

Optimization

• LoRA Strength:
  • Strength 1.0: Best for motion fidelity. Keeps the movement as close as possible to the guide video.
  • Strength > 1.0: Improves hair capture and identity from Frame 0 but causes the model to deviate from the guide video's original motion.
• Multi-pass Workflows: You can try to improve results by creating multi-pass workflows, running the LoRA at different strengths to refine the output.
• Prompting: Detailed prompts currently have no effect. The LoRA was trained with a single unique trigger and does not respond to scene or subject descriptions yet.

⚠️ Known Issues (Alpha Release)

• Identity Leakage: The model may occasionally default back to the original hair from the guide video.
• Hard Cuts: Jump cuts break the temporal attention chain. If a character disappears and reappears, the model might "reset" to the original guide video features.

🚀 Future & Support

I am already working on the next version. My goal is to allow the model to work with only a reference photo, eliminating the need for the initial first-frame swap. I have already tested this approach, but the results are not yet up to my standards for a full release.

I still have many other trained versions to test. If I find an iteration that performs better than this one, I will release it here.

Maintaining R&D and renting Blackwell GPUs is expensive. If you find this LoRA useful, please consider a donation to support the development of v2 (with SAM 3 support).

👉 Support here: Buy Me a Coffee