wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-02-15 07:52:35 +08:00
Code Issues Actions 17 Packages Projects Releases Wiki Activity

Use @ashen-uncensored formula, which works better!!!

Browse Source
This commit is contained in:
Raphael Walker 2023-12-06 17:23:02 +01:00
parent 757bfb7d9d
commit 20d8a06bbe
No known key found for this signature in database
GPG Key ID: E6F58BE3395D3AA8
2 changed files with 10 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
comfy/samplers.py
View File

@ -261,26 +261,22 @@ def sampling_function(model, x, timestep, uncond, cond, cond_scale, model_option
sag_sigma = model_options["sag_sigma"] sag_sigma = model_options["sag_sigma"]
sag_threshold = model_options.get("sag_threshold", 1.0) sag_threshold = model_options.get("sag_threshold", 1.0)
# or is it x - uncond?
# or do I have to use the sigma ?
x0_est = uncond_pred
# this method is added by the sag patcher # this method is added by the sag patcher
uncond_attn = model.get_attn_scores() uncond_attn = model.get_attn_scores()
degraded = create_blur_map(x0_est, uncond_attn, x - uncond_pred, sag_sigma, sag_threshold) degraded = create_blur_map(uncond_pred, uncond_attn, sag_sigma, sag_threshold)
# todo, optimize this: doing it this way creates an extra call that we don't even use degraded_noised = degraded + x - uncond_pred
(_, sag) = calc_cond_uncond_batch(model, cond, uncond, degraded, timestep, model_options) # TODO optimize this: doing it this way creates an extra call that we don't even use
(_, sag) = calc_cond_uncond_batch(model, cond, uncond, degraded_noised, timestep, model_options)
return uncond_pred + (cond_pred - uncond_pred) * cond_scale + (uncond_pred - sag) * sag_scale # Unless I've misunderstood the paper, this is supposed to be (uncond_pred - sag) * sag_scale.
# but this is what the automatic1111 implementation does, and it works better??
return uncond_pred + (cond_pred - uncond_pred) * cond_scale + (degraded - sag) * sag_scale
else: else:
return uncond_pred + (cond_pred - uncond_pred) * cond_scale return uncond_pred + (cond_pred - uncond_pred) * cond_scale
def create_blur_map(x0, attn, noise, sigma=3.0, threshold=1.0): def create_blur_map(x0, attn, sigma=3.0, threshold=1.0):
# reshape and GAP the attention map # reshape and GAP the attention map
_, hw1, hw2 = attn.shape _, hw1, hw2 = attn.shape
b, lc, lh, lw = x0.shape b, lc, lh, lw = x0.shape
# I think this depends on the model:
# sdxl has 20 heads and the middle of the unet is 4 times smaller
# sd 1.5 has 8 heads and the middle of the unet is 8 times smaller
attn = attn.reshape(b, -1, hw1, hw2) attn = attn.reshape(b, -1, hw1, hw2)
# Global Average Pool # Global Average Pool
mask = attn.mean(1, keepdim=False).sum(1, keepdim=False) > threshold mask = attn.mean(1, keepdim=False).sum(1, keepdim=False) > threshold
@ -302,8 +298,6 @@ def create_blur_map(x0, attn, noise, sigma=3.0, threshold=1.0):
blurred = gaussian_blur_2d(x0, kernel_size=9, sigma=sigma) blurred = gaussian_blur_2d(x0, kernel_size=9, sigma=sigma)
blurred = blurred * mask + x0 * (1 - mask) blurred = blurred * mask + x0 * (1 - mask)
blurred = blurred + noise
return blurred return blurred
def gaussian_blur_2d(img, kernel_size, sigma): def gaussian_blur_2d(img, kernel_size, sigma):

4
comfy_extras/nodes_sag.py
View File

@ -53,8 +53,8 @@ class SagNode:
@classmethod @classmethod
def INPUT_TYPES(s): def INPUT_TYPES(s):
return {"required": { "model": ("MODEL",), return {"required": { "model": ("MODEL",),
"scale": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 30.0, "step": 0.1}), "scale": ("FLOAT", {"default": 0.5, "min": -2.0, "max": 5.0, "step": 0.1}),
"blur_sigma": ("FLOAT", {"default": 3.0, "min": 0.0, "max": 10.0, "step": 0.1}), "blur_sigma": ("FLOAT", {"default": 2.0, "min": 0.0, "max": 10.0, "step": 0.1}),
}} }}
RETURN_TYPES = ("MODEL",) RETURN_TYPES = ("MODEL",)
FUNCTION = "patch" FUNCTION = "patch"