diff --git a/CHANGELOG.md b/CHANGELOG.md index 6f5a0e9..c933a67 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,4 +1,7 @@ -### v2.0.0 - 2024 Mar.01 +### v2.0.0 - 2024 Mar.04 +- Support **SDXL** + +### v2.0.gamma - 2024 Mar.01 - Major **Rewrite** & **Optimization** ### v2.0.beta - 2024 Feb.29 diff --git a/LICENSE b/LICENSE index b0340f4..91b74be 100644 --- a/LICENSE +++ b/LICENSE @@ -1,6 +1,6 @@ MIT License -Copyright (c) 2023 Haoming +Copyright (c) 2024 Haoming Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal diff --git a/README.md b/README.md index 2b535d4..d39d157 100644 --- a/README.md +++ b/README.md @@ -4,7 +4,7 @@ allowing you to adjust the brightness, contrast, and color of the generations. > Also compatible with [Forge](https://github.com/lllyasviel/stable-diffusion-webui-forge)! -**Important:** The color currently only works for **SD 1.5** Checkpoints +> Now supports SDXL! > [Sample Images](#sample-images) @@ -12,7 +12,7 @@ allowing you to adjust the brightness, contrast, and color of the generations. After installing this Extension, you will see a new section in both **txt2img** and **img2img** tabs. Refer to the parameters and sample images below and play around with the values. -**Note:** Since this modifies the underlying latent noise, the composition may change drastically. Using the **Ones** scaling seems to reduce the variations. +**Note:** Since this modifies the underlying latent noise, the composition may change drastically. #### Parameters - **Enable:** Turn on/off this Extension @@ -85,15 +85,16 @@ Refer to the parameters and sample images below and play around with the values. - **Abs:** Calculate using the absolute values of the chosen Tensors instead - `x += abs(F) * y` -

-

+

+ +

#### Scaling Settings By default, this Extension offsets the noise by the same amount each step. But due to the denoising process, this may produce undesired outcomes such as blurriness at high **Brightness** or noises at low **Brightness**. Therefore, I added a scaling option to modify the offset amount throughout the process. -> Essentially, the "magnitude" of the default Tensor gets smaller every step, so offsetting by the same amount will have stronger effects at later steps. This is reversed on the `Alt.` Tensor however. +> Essentially, the "magnitude" of the default Tensor gets smaller every step, so offsetting by the same amount will have stronger effects at the later steps. - **Flat:** Default behavior. Same amount each step. - **Cos:** Cosine scaling. *(High -> Low)* @@ -101,61 +102,34 @@ Therefore, I added a scaling option to modify the offset amount throughout the p - **1 - Cos:** *(Low -> High)* - **1 - Sin:** *(High -> Low)* -> In my experience, **`1 - Sin`** works better for the **default** Tensor while **`1 - Cos`** works better for the **Alt.** Tensor -

-Alt. Disabled
- +

-

-Alt. Enabled
- -

- -

Notice the blurriness and the noises on Flat scaling

- ## Sample Images -- **Checkpoint:** [UHD-23](https://civitai.com/models/22371/uhd-23) -- **Pos. Prompt:** `(masterpiece, best quality), 1girl, solo, night, street, city, neon_lights` -- **Neg. Prompt:** `(low quality, worst quality:1.2)`, [`EasyNegative`](https://huggingface.co/datasets/gsdf/EasyNegative/tree/main), [`EasyNegativeV2`](https://huggingface.co/gsdf/Counterfeit-V3.0/tree/main/embedding) -- `Euler a`; `20 steps`; `7.5 CFG`; `Hires. fix`; `Latent (nearest)`; `16 H.steps`; `0.6 D.Str.`; `Seed:`**`3814649974`** -- `Straight Abs.` +- **Checkpoint:** [Animagine XL V3](https://civitai.com/models/260267) +- **Pos. Prompt:** `[high quality, best quality], 1girl, solo, casual, night, street, city, ` +- **Neg. Prompt:** `lowres, [low quality, worst quality], jpeg` +- `Euler A SGMUniform`; `10 steps`; `2.0 CFG`; **Seed:** `2836968120` +- `Straight Abs.` ; `Flat`

-Base
-Extension Disabled
- +Disabled
+

-Dark
-Brightness: -3; Contrast: 1.5
- +Brightness: 2.5 ; Contrast: -0.5 ; Saturation: 1.5
+R: 2.5; G: 2.5; B: -4
+

-Bright
-Brightness: 2.5; Contrast: 0.5; Alt: Enabled
- +Brightness: -2.5 ; Contrast: 0.5 ; Saturation: 1.5
+R: -1; G: -1; B: 2.5
+

-

-Chill
-Brightness: -2.5; Contrast: 1.25
-R: -1.5; B: 2.5
- -

- -

-Mexican Movie
-Brightness: 3; Saturation: 1.5
-R: 2; G: 1; B: -2
- -

- -

Notice the significant differences even when using the same seed

- ## Roadmap - [X] Extension Released - [X] Add Support for **X/Y/Z Plot** @@ -169,15 +143,18 @@ Therefore, I added a scaling option to modify the offset amount throughout the p - You can enable this in the **Infotext** section of the **Settings** tab - [X] Add Infotext Support *(by. [catboxanon](https://github.com/catboxanon))* - [X] ADD **HDR** Script +- [X] Add SDXL Support - [ ] Add Gradient features -- [ ] Add SDXL Support -

X/Y/Z Plot Support
(Outdated Contrast Value)

-

+

+X/Y/Z Plot Support
+ +

-

X/Y/Z Plot w/ Randomize

-

-

The value is used as the random seed
You can refer to the console to see the randomized values

+

+Randomize
+
+The value is used as the random seed
You can refer to the console to see the randomized values

## API You can use this Extension via [API](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/API) by adding an entry to the `alwayson_scripts` of your payload. @@ -197,9 +174,8 @@ The `args` are sent in the following order in an `array`: - **Scaling Settings:** `str` ## Known Issues -- Does **not** work with `DDIM`, `UniPC` samplers +- Does **not** work with `DDIM`, `UniPC`, `Euler` samplers - Has little effect when used with certain **LoRA**s -- Colors are incorrect when using SDXL checkpoints ## HDR

BETA

diff --git a/samples/00.jpg b/samples/00.jpg index 9ccaa26..8eee988 100644 Binary files a/samples/00.jpg and b/samples/00.jpg differ diff --git a/samples/01.jpg b/samples/01.jpg index ce2e102..48eb847 100644 Binary files a/samples/01.jpg and b/samples/01.jpg differ diff --git a/samples/02.jpg b/samples/02.jpg index 246f203..cadedfa 100644 Binary files a/samples/02.jpg and b/samples/02.jpg differ diff --git a/samples/03.jpg b/samples/03.jpg deleted file mode 100644 index 9bbe8d8..0000000 Binary files a/samples/03.jpg and /dev/null differ diff --git a/samples/04.jpg b/samples/04.jpg deleted file mode 100644 index 0e1d214..0000000 Binary files a/samples/04.jpg and /dev/null differ diff --git a/samples/Bright.jpg b/samples/Bright.jpg deleted file mode 100644 index 997bbc1..0000000 Binary files a/samples/Bright.jpg and /dev/null differ diff --git a/samples/Dark.jpg b/samples/Dark.jpg deleted file mode 100644 index d1f8459..0000000 Binary files a/samples/Dark.jpg and /dev/null differ diff --git a/samples/Method.jpg b/samples/Method.jpg new file mode 100644 index 0000000..b3a8b6a Binary files /dev/null and b/samples/Method.jpg differ diff --git a/samples/Random.jpg b/samples/Random.jpg index cf8c273..0123cfa 100644 Binary files a/samples/Random.jpg and b/samples/Random.jpg differ diff --git a/samples/Scaling.jpg b/samples/Scaling.jpg index 7f1bbfe..4771a38 100644 Binary files a/samples/Scaling.jpg and b/samples/Scaling.jpg differ diff --git a/samples/Scaling_alt.jpg b/samples/Scaling_alt.jpg deleted file mode 100644 index 18c06bf..0000000 Binary files a/samples/Scaling_alt.jpg and /dev/null differ diff --git a/samples/XYZ.jpg b/samples/XYZ.jpg index 36822f5..e68b904 100644 Binary files a/samples/XYZ.jpg and b/samples/XYZ.jpg differ diff --git a/scripts/cc_callback.py b/scripts/cc_callback.py index 16fca98..438d44c 100644 --- a/scripts/cc_callback.py +++ b/scripts/cc_callback.py @@ -62,6 +62,14 @@ class NoiseMethods: return noise / noise.std() +def RGB_2_CbCr(r:float, g:float, b:float) -> float: + """Convert RGB channels into YCbCr for SDXL""" + cb = -0.15 * r - 0.29 * g + 0.44 * b + cr = 0.44 * r - 0.37 * g - 0.07 * b + + return cb, cr + + original_callback = KDiffusionSampler.callback_state @torch.no_grad() @@ -72,6 +80,8 @@ def cc_callback(self, d): if getattr(self.p, "is_hr_pass", False) and not self.vec_cc["doHR"]: return original_callback(self, d) + is_xl: bool = self.p.sd_model.is_sdxl + mode = str(self.vec_cc["mode"]) method = str(self.vec_cc["method"]) source = d[mode] @@ -108,23 +118,42 @@ def cc_callback(self, d): self.vec_cc["b"], ) - for i in range(batchSize): - # Brightness - source[i][0] += target[i][0] * bri - # Contrast - source[i][0] += NoiseMethods.get_delta(source[i][0]) * con + if not is_xl: + for i in range(batchSize): + # Brightness + source[i][0] += target[i][0] * bri + # Contrast + source[i][0] += NoiseMethods.get_delta(source[i][0]) * con - # R - source[i][2] -= target[i][2] * r - # G - source[i][1] += target[i][1] * g - # B - source[i][3] -= target[i][3] * b + # R + source[i][2] -= target[i][2] * r + # G + source[i][1] += target[i][1] * g + # B + source[i][3] -= target[i][3] * b - # Saturation - source[i][2] *= sat - source[i][1] *= sat - source[i][3] *= sat + # Saturation + source[i][2] *= sat + source[i][1] *= sat + source[i][3] *= sat + + else: + # But why... + cb, cr = RGB_2_CbCr(r, b, g) + + for i in range(batchSize): + # Brightness + source[i][0] += target[i][0] * bri + # Contrast + source[i][0] += NoiseMethods.get_delta(source[i][0]) * con + + #CbCr + source[i][1] -= target[i][1] * cr + source[i][2] += target[i][2] * cb + + # Saturation + source[i][1] *= sat + source[i][2] *= sat return original_callback(self, d)