If it seems like the model can't "see" or completely ignores your reference images...
#11
by pwnyourace - opened
What attention backend are you using? If you're using the comfy kitchen CK backend it won't be able to see your images. Not sure the reason but tested by comparing with CK backend and the 2.0+ pytorch cross attention. It's possible sage attention or the flash attention flag could have an effect too. So if you got gibberish from this model or it ignored your input images, try with pytorch cross attention
AMD gfx1150 on ROCm 10.x but happened on 7.xx also. Don't know if it's AMD only. Not trying to report a bug just posting for anyone having poor results.