>>82624617>So you only trained the clip_l?No, i just trained all blocks. There's no point in training CLIP. It's a different architecture, clip does almost nothing in flux.
Inside MMDiT, there is a part that deals with text (circled in red). The idea is that training this part should give you an effect similar to text encoder training in SDXL.
Also, my feeling so far is that double blocks are mostly responsible for content and composition and single blocks are responsible for style, so for a character lora you train (a part of ? ) the double blocks and don't train single blocks, but i haven't tried all possible combinations yet
