I really love these lectures from experienced researchers/engineers. Listening to people who are down-to-earth (real-world problems) is very enjoyable! (Lucas Beyer (bl16) on SigLip)

After my last experiment, curious now if anyone has done a study on comparing applying lora to different layers of DiTs

I got sent this cool blog. The author took our course on diffusion models, figured "there should be a better way" and then thought his way to inventing flow matching without even knowing it! Legend 😁 tommyc.xyz/posts/simple-d… Tanishq Mathew Abraham, Ph.D. lookit

CSGO builds on top of B-LoRA ! Looking forward to the style transfer dataset by InstantX. CSGO: arxiv.org/abs/2408.16766 B-LoRA: b-lora.github.io/B-LoRA/

Was looking at B-LoRA for Flux and the first thing that pops up is you can't just do prompt injection as proposed in the paper because in UNet, the encoder_hidden_states a.k.a prompt emebds don't change per layer.

Flux inversion results are pretty interesting. I’ll post some samples on real world images today