search

mcmonkeyprojects / sd-dynamic-thresholding

Dynamic Thresholding (CFG Scale Fix) for Stable Diffusion (eSwarmUI, ComfyUI, and Auto WebUI)
MIT License
1.1k stars 103 forks source link

please support sdxl #56

Closed nishidiaomao closed 1 year ago

nishidiaomao commented 1 year ago

dynamic_thresholding.py->def combine_denoised(self, x_out, conds_list, uncond, cond_scale):

add this code

if not isinstance(uncond, torch.Tensor): uncond = uncond['crossattn']

mcmonkey4eva commented 1 year ago

Fixed, thank you!

The behavior of outputs from dynthresh seems to be fundamentally different on SDXL, so test with care (don't just copy over working settings from SDv1 to SDXL usage).