Question about training code and character lora

[karmag466]

First of all thank you so much for your willing to open the training code. Is there any chance to know how the progress going with that? I don't mean to pressure, but I'd love to know if it's about months or weeks?

Also can you provide informations about training character lora - do we need to train on videos or we can put just images in dataset? Images works very well on ltx and used to work on wan. Not sure how it will be with scail-2
Maybe it's good to put this in documentation since many people who decide to train lora wants to do a character lora.

[fengjia-guo]

Hi, thanks for your interest in training code and character LoRAs. We're still organizing the SAT training code for easier usage, and it will hopefully be available this week.

We haven't trained character LoRAs on SCAIL-2 ourselves yet, so we can’t confirm whether image‑only datasets will work. We strongly encourage the community to experiment and share their findings with us. Once we see reliable success cases, we’ll document those best practices.

I'm personally working on a community version that is more easy-to-use and VRAM-friendly, based on more popular frameworks such as DiffSynth-Studio or musubi-tuner, so that training SCAIL-2 LoRAs or finetunes might be more convenient.

[karmag466]

Hi, thanks for your interest in training code and character LoRAs. We're still organizing the SAT training code for easier usage, and it will hopefully be available this week.

We haven't trained character LoRAs on SCAIL-2 ourselves yet, so we can’t confirm whether image‑only datasets will work. We strongly encourage the community to experiment and share their findings with us. Once we see reliable success cases, we’ll document those best practices.

I'm personally working on a community version that is more easy-to-use and VRAM-friendly, based on more popular frameworks such as DiffSynth-Studio or musubi-tuner, so that training SCAIL-2 LoRAs or finetunes might be more convenient.

Thank you. I think announcing releasing training code would be a great thing for example on StableDiffusion subreddit - people really loves your model there. Some publicity about this will be very positive, especially if the authors announce it.

[fengjia-guo]

Thank you. I think announcing releasing training code would be a great thing for example on StableDiffusion subreddit - people really loves your model there. Some publicity about this will be very positive, especially if the authors announce it.

Thank you for your advice. We appreciate the community's passion for SCAIL-2, and we'll announce the release of training code properly once it is available.

[fengjia-guo]

Hi, thank you for your interest and patience! We’ve now open-sourced the training code on sat-scail2 branch. A VRAM-friendly version based on musubi-tuner is available here.