Não testei ainda, mas vi muitos exemplos, atualmente meu foco é LTX então no fim os modelos que lançam para Wan eu não dou tanta atenção, o único do ecossistema Wan que eu treinei alguma coisa recentemente foi o Bernini, modelo de edição, fora isso meu foco total é em treinamento de modelos, é difícil eu usar os modelos para inferência, só treinamento mesmo.
Hey everyone! 👋 New FaceID IC-LoRA for LTX-2 🎭 Not perfect, but it works nicely!
You give it a close-up face photo + a prompt, and it keeps that person's identity in the video.
Grab it on HF 👇
https://t.co/ELXCHMeeVp
To run it you'll need my custom BFS Nodes in ComfyUI — the node is what actually applies the custom TASS-RoPE + overlap that injects the reference. Just feed it a clean, close-up face.
⚠️ Prompt format (in this order for best results):
ref_t2v: [facial description] → [action/scene]
e.g. "ref_t2v: A light-skinned man with short dark hair, blue eyes and light stubble is folding clothes in a laundry room, medium-wide shot."
Days of A LOT of testing went into this — several approaches and papers before landing on v1. I also tried an ArcFace projector to translate the reference straight into the guts of LTX 😅 In the end it's only subtle (optional in the node); the identity mostly comes from the overlap + a good descriptive prompt.
@ltx_model I have created and continue to training various experimental models, focused on editing (inpainting with or without reference), motion transfer, and head swapping.
https://t.co/Il4t6vQyFj
https://t.co/mi8ijH1ut2
https://t.co/8MesLN6PTU
@sucrasm@ComfyUI There's one from https://t.co/MPPctfsdBn; you can combine it with Obscura Remove + Mask in LTX and you can remove watermarks in an interesting way.
@darklig45208568@drawthingsapp The base model will definitely do simple head swaps, now put some object on the face and tell me if it works, or try changing the angle of the face a lot. Everyone tests the simple things but they don't test the complex scenarios, which are the most important.
@mix_buchi_ Did you read the instructions on how to use it? And about possible problems using the model with Lora de Distill? Remember that since the editing depends on a prompt, a CFG greater than 1 can be crucial.
@brubbleR @toyxyz3 You're either not using it correctly, or you're trying to use it for a purpose different from what it's designed for. Read the model descriptions and instructions for use.
@oliviscusAI Para quem tiver interesse existem meus modelos de head swap para imagem e alguns teste que eu fiz com head swap em video. Só procurar por BFS no huggingface ou civitai.
@Gradio In words of @Tremeschine (As X is blocked in Brazil): "Thanks Gradio for sharing my projects!! Gradio helped me a lot streamline implementing nice user interfaces, and I'm excited to see how it evolves in the future, DepthFlow thanks you!