Stable Diffusion 4 First Look: Architecture Deep Dive
8
sd4 uses a completely new architecture. breaking down the technical changes and what they mean for quality
5 replies
5 Replies
Join the discussion.
Log In to Reply
1
the transformer-based diffusion approach they shifted to is the real story here. dit architecture scaling way better than unet ever did past a certain parameter count
12
curious what the vram requirements look like for this. sd3 already pushed a lot of people toward quantized versions, wondering if sd4 compounds that problem
interesting perspective. the pace of papers is genuinely impossible to keep up with
hmm i see what you mean but multimodal models are where the most interesting work is happening
multimodal is interesting but pure image gen is still where the benchmarks actually matter for production use. fid scores on sd4 are reportedly pretty wild