search

kijai / ComfyUI-Diffusers-X-Adapter

ComfyUI diffusers wrapper node X-Adapter testing
https://showlab.github.io/X-Adapter/
Apache License 2.0
230 stars 8 forks source link

Image quality(through 1.5 controlnet) is poorer than SDXL(direct) #2

Open eblis88 opened 9 months ago

eblis88 commented 9 months ago

How to fix it?

kijai commented 9 months ago

I don't really know yet, different model combos work completely different, some don't match at all and some give really unique results, this is part of the experimentation at this point.

I did manage to allow larger resolutions now which helps a bit, still need to be 2x the 1.5 resolution, and still only squares, but it allows for far better end results.

TomLucidor commented 1 month ago

Probably need to test between all the major flavors of SD1.5 vs all the major flavors of SDXL and/or Pony. Borrowing concept from this article Some small questions to consider:

  1. How would the introduction of Flux dev (some will use this) and AuraFlow (Pony7 will use this) change the situation? For example, they have stronger architecture to stuff half of an LLM inside as an additional embedding.
  2. Why only squares? Can't training also accommodate things like alternate aspect ratios that is default in SD1.5 and corresponding 2x sizes in SDXL?