You can now fine-tune Wan 2.1 14b using Replicate and Ostris’s AI-toolkit 🎉 - autocaption training images - publish to Hugging Face - save intermediate checkpoints to Weights & Biases All of the outputs below are text-to-video. Still much to do!

Takes about 20mins to train 14b on an H100 with 2000 steps

Togoda is Google on Steroids with AI summaries . 🚀 The only thematic AI search engine.👀 It's 100% private with third party proxy. 🧨 Try it today & experience the difference! 👉Follow us @togoda_com 👈 🚀Help us grow & share this post!🚀

@ostrisai Awesome thank you! The space looks straightforward, will try it out 🙏🏽 Running the H100 for an hour costs ~7$ Do you know how long it would take to fine tune the model using e.g. 20 images leaving the settings all at their default values?

@ostrisai 20 minutes

@ostrisai Great job here 💪 From the models you trained, are 2000 steps and the lora rank at 32 the ideal params for great results in photo-realistic ?

@ostrisai I haven't trained enough to know best, but these are using @ostrisai's defaults. I've found they work well so far.

@ostrisai Awesome, will there be also option to use wan 2.1 for image generation on replicate?

@ostrisai Can be. Why do you want to?

@ostrisai How many images did you use to train it?

@ostrisai Q. When you write captions for the images do you need to mention the trigger word too?

@ostrisai Yes it's best to use in the right context, if you do not the trigger will get added automatically by AI toolkit downstream