Trygve Wiig trygvebw
trygvebw
/ blur_latent_noise.py
Last active
April 29, 2024 07:50
A way to use high classifier-free guidance (CFG) scales with Stable Diffusion by applying an unsharp mask to the model output, while avoiding the artifacts and excessive contrast/saturation this usually produces
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| # This is an abbreviated demonstration of how to perform this technique. The code | |
| # is a simplified version of that in my own custom codebase, and can't be plugged | |
| # into other ways of using Stable Diffusion (e.g. Diffusers or A1111) without changes. | |
| # In essence, the observation that the CFG formula: | |
| # | |
| # output_noise = uncond_noise + (cond_noise - uncond_noise) * scale | |
| # | |
| # looks a lot like the formula for the unsharp mask, a common way to sharpen or add local contrast to images: | |
| # |
trygvebw
/ find_noise.py
Last active
June 21, 2024 15:31
A "reverse" version of the k_euler sampler for Stable Diffusion, which finds the noise that will reconstruct the supplied image
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| import torch | |
| import numpy as np | |
| import k_diffusion as K | |
| from PIL import Image | |
| from torch import autocast | |
| from einops import rearrange, repeat | |
| def pil_img_to_torch(pil_img, half=False): | |
| image = np.array(pil_img).astype(np.float32) / 255.0 |
trygvebw
/ generate.py
Last active
July 29, 2023 20:01
My Stable Diffusion image generation function. Abbreviated – will not run because of a few missing utility functions and classes.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| def normalize_latent(x, max_val, quantile_val): | |
| x = x.detach().clone() | |
| for i in range(x.shape[0]): | |
| if x[[i], :].std() > 1.0: | |
| x[[i], :] = x[[i], :] / x[[i], :].std() | |
| s = torch.quantile(torch.abs(x[[i], :]), quantile_val) | |
| s = torch.maximum(s, torch.ones_like(s) * max_val) | |
| x[[i], :] = x[[i], :] / (s / max_val) | |
| return x |