Chroma support #627

Activity

1. stduhpf commented on Mar 10, 2025

   

   stduhpf

   on Mar 10, 2025

   Contributor

   More actions

   I don't think it will work out of the box, because they modified Flux's DiT block architecture, and the prompt processing is also a bit different. It may not be too difficult to implement though.
2. Green-Sky commented on Aug 1, 2025

   

   Green-Sky

   on Aug 1, 2025

   Contributor

   More actions

   Technically it should work now (see #696 ), but I just tried a new version and it performs very poorly and I am not sure why. It might be the schedule (everyone seems to be using beta), but not sure.

   

   This is the best output I could generate using:

   bin/sd
   --diffusion-model models/chroma-unlocked-v48-detail-calibrated-Q4_0-marduk191.gguf
   --cfg-scale 12
   -t 8
   --t5xxl models/flux-extra/t5xxl_q8_0.gguf
   --vae models/flux-extra/ae-f16.gguf
   --steps 40
   --sampling-method euler # ancestral samplers perform worse
   --diffusion-fa
   --chroma-disable-dit-mask # required for flash attention on cuda still
   -W 1024 -H 768
   --vae-tiling
   -p "photo of an old dog, sitting on the veranda. It is a sunny summer afternoon with a few clouds. There are humans in the background."
   -n "ugly, low quality, blury"
   

   As you can see I had to go real high with the cfg scale and steps to get it somewhat sharp.

   Here is cfg lowered to 6:

   
3. stduhpf commented on Aug 2, 2025

   

   stduhpf

   on Aug 2, 2025

   Contributor

   More actions

   I'm having similar results (with the same kind of flaws) when comparing ComfyUI-zluda (with the recommended workflow) and sdcpp (rocm) with the same prompts and settings. I guess it probably has something to do with the model not being quite done training.

   Maybe --chroma-disable-dit-mask hurts the model more than we initially thought too.
4. Repository owner locked and limited conversation to collaborators on Oct 28, 2025
5. leejet

   converted this issue into a discussion #924 on Oct 28, 2025