feat: init tkg

.python-version

@@ -0,0 +1 @@
+3.10

app.py

@@ -1,36 +1,44 @@
-import gradio as gr
-import numpy as np
 import random
 
-# import spaces #[uncomment to use ZeroGPU]
-from diffusers import DiffusionPipeline
+import numpy as np
 import torch
 
+from diffusers import StableDiffusionXLPipeline
+import spaces
+
+import gradio as gr
+
+from .tkg import apply_tkg_noise, COLOR_SET_MAP
+
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+
 device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo"  # Replace to the model you would like to use
+model_repo_id = "cagliostrolab/animagine-xl-4.0"  # Replace to the model you would like to use
 
-if torch.cuda.is_available():
-    torch_dtype = torch.float16
-else:
-    torch_dtype = torch.float32
 
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
+pipe = StableDiffusionXLPipeline.from_pretrained(
+    "cagliostrolab/animagine-xl-4.0",
+    torch_dtype=torch.bfloat16,
+    custom_pipeline="lpw_stable_diffusion_xl",
+    add_watermarker=False,
+)
 pipe = pipe.to(device)
 
 MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
+MAX_IMAGE_SIZE = 2048
 
-
-# @spaces.GPU #[uncomment to use ZeroGPU]
+@spaces.GPU
 def infer(
-    prompt,
-    negative_prompt,
-    seed,
-    randomize_seed,
-    width,
-    height,
-    guidance_scale,
-    num_inference_steps,
+    prompt: str,
+    negative_prompt: str,
+    seed: int,
+    randomize_seed: bool,
+    width: int,
+    height: int,
+    guidance_scale: float,
+    num_inference_steps: int,
+    tkg_channels: list[int] = [0, 1, 1, 0],
     progress=gr.Progress(track_tqdm=True),
 ):
     if randomize_seed:
@@ -38,7 +46,27 @@ def infer(
 
     generator = torch.Generator().manual_seed(seed)
 
-    image = pipe(
+    latents = torch.randn(
+        (
+            1, 
+            2, 
+            height // 8, 
+            width // 8,
+        ),
+        generator=generator,
+        device=device,
+    )
+    latents = apply_tkg_noise(
+        latents,
+        shift=0.11,
+        delta_shift=0.1,
+        std_dev=0.5,
+        factor=8,
+        channels=tkg_channels,
+    )
+
+    images = pipe(
+        latents=latents,
         prompt=prompt,
         negative_prompt=negative_prompt,
         guidance_scale=guidance_scale,
@@ -46,27 +74,55 @@ def infer(
         width=width,
         height=height,
         generator=generator,
-    ).images[0]
-
-    return image, seed
-
+    ).images
+
+    w_tkg, wo_tkg = images
+
+    return w_tkg, wo_tkg, seed
+
+def color_name_to_channels(color_name: str) -> list[int]:
+    if color_name in COLOR_SET_MAP:
+        return COLOR_SET_MAP[color_name]
+    else:
+        raise ValueError(f"Unknown color name: {color_name}")
+
+def on_generate(
+    prompt: str,
+    negative_prompt: str,
+    seed: int,
+    randomize_seed: bool,
+    width: int,
+    height: int,
+    guidance_scale: float,
+    num_inference_steps: int,
+    color_name: str,
+    *args,
+    **kwargs
+):
+    tkg_channels = color_name_to_channels(color_name)
+    # TODO: custom channels
+
+    return infer(
+        prompt,
+        negative_prompt,
+        seed,
+        randomize_seed,
+        width,
+        height,
+        guidance_scale,
+        num_inference_steps,
+        tkg_channels=tkg_channels,
+    )
 
 examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
+    # "1girl, arima kana, oshi no ko, hoshimachi suisei, hoshimachi suisei \(1st costume\), cosplay, looking at viewer, smile, outdoors, night, v, masterpiece, high score, great score, absurdres",
+    "1girl, solo, upper body, looking at viewer, straight-on, masterpiece, best quality",
 ]
 
-css = """
-#col-container {
-    margin: 0 auto;
-    max-width: 640px;
-}
-"""
 
-with gr.Blocks(css=css) as demo:
-    with gr.Column(elem_id="col-container"):
-        gr.Markdown(" # Text-to-Image Gradio Template")
+with gr.Blocks() as demo:
+    with gr.Column():
+        gr.Markdown("# TKG Chroma-Key with AnimagineXL 4.0")
 
         with gr.Row():
             prompt = gr.Text(
@@ -79,14 +135,16 @@ with gr.Blocks(css=css) as demo:
 
             run_button = gr.Button("Run", scale=0, variant="primary")
 
-        result = gr.Image(label="Result", show_label=False)
+        with gr.Row():
+            result_w_tkg = gr.Image(label="Result", show_label=False)
+            result_wo_tkg = gr.Image(label="Result", show_label=False)
 
         with gr.Accordion("Advanced Settings", open=False):
             negative_prompt = gr.Text(
                 label="Negative prompt",
                 max_lines=1,
                 placeholder="Enter a negative prompt",
-                visible=False,
+                default="lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurry",
             )
 
             seed = gr.Slider(
@@ -105,7 +163,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=256,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
-                    value=1024,  # Replace with defaults that work for your model
+                    value=832,  
                 )
 
                 height = gr.Slider(
@@ -113,7 +171,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=256,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
-                    value=1024,  # Replace with defaults that work for your model
+                    value=1152,
                 )
 
             with gr.Row():
@@ -122,7 +180,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=0.0,
                     maximum=10.0,
                     step=0.1,
-                    value=0.0,  # Replace with defaults that work for your model
+                    value=5.0,  
                 )
 
                 num_inference_steps = gr.Slider(
@@ -130,10 +188,11 @@ with gr.Blocks(css=css) as demo:
                     minimum=1,
                     maximum=50,
                     step=1,
-                    value=2,  # Replace with defaults that work for your model
+                    value=25,  
                 )
 
         gr.Examples(examples=examples, inputs=[prompt])
+
     gr.on(
         triggers=[run_button.click, prompt.submit],
         fn=infer,
@@ -147,7 +206,7 @@ with gr.Blocks(css=css) as demo:
             guidance_scale,
             num_inference_steps,
         ],
-        outputs=[result, seed],
+        outputs=[result_w_tkg, result_wo_tkg, seed],
     )
 
 if __name__ == "__main__":

pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "app"
+version = "0.1.0"
+description = "Add your description here"
+readme = "README.md"
+requires-python = ">=3.10,<3.13"
+dependencies = [
+    "accelerate>=1.10.0",
+    "diffusers>=0.35.1",
+    "spaces>=0.40.1",
+    "torch>=2.8.0",
+    "transformers>=4.55.4",
+    "xformers>=0.0.32.post2",
+]
+
+[dependency-groups]
+dev = [
+    "gradio>=5.43.1",
+    "ruff>=0.12.10",
+    "ty>=0.0.1a19",
+]

requirements.txt

@@ -1,6 +1,6 @@
 accelerate
 diffusers
-invisible_watermark
 torch
 transformers
-xformers
+xformers
+spaces

tkg.py

@@ -0,0 +1,117 @@
+from typing import NamedTuple
+
+import torch
+import torch.nn.functional as F
+
+
+def get_mean_shifted_latents(
+    latents: torch.Tensor,
+    shift: float = 0.11,
+    delta_shift: float = 0.1,
+    channels: list[int] = [0, 1, 1, 0],  # list of {-1, 0, 1}
+) -> torch.Tensor:
+    shifted_latents = latents.clone()
+
+    for idx, sign in enumerate(channels):
+        if sign == 0:
+            # skip
+            continue
+
+        latent_channel = shifted_latents[:, idx, :, :]
+
+        positive_ratio = (latent_channel > 0).float().mean()
+        target_ratio = positive_ratio + shift * sign
+
+        # gradually shift latent_channel
+        while True:
+            latent_channel += delta_shift * sign
+            new_positive_ratio = (latent_channel > 0).float().mean()
+            if new_positive_ratio >= target_ratio:
+                break
+
+        # replace the channel in the original latents
+        shifted_latents[:, idx, :, :] = latent_channel
+
+    return shifted_latents
+
+
+def get_2d_gaussian(
+    latent_height: int,
+    latent_width: int,
+    std_dev: float,
+    device: torch.device,
+    center_x: float = 0.0,
+    center_y: float = 0.0,
+    factor: int = 8,  # idk why
+):
+    y = torch.linspace(-1, 1, steps=latent_height // factor, device=device)
+    x = torch.linspace(-1, 1, steps=latent_width // factor, device=device)
+
+    y_grid, x_grid = torch.meshgrid(y, x, indexing="ij")
+
+    x_grid = x_grid - center_x
+    y_grid = y_grid - center_y
+
+    gauss = torch.exp(-((x_grid**2 + y_grid**2) / (2 * std_dev**2)))
+    gauss = gauss[None, None, :, :]  # add batch and channel dimensions
+
+    return gauss
+
+
+def apply_tkg_noise(
+    latents: torch.Tensor,
+    shift: float = 0.11,
+    delta_shift: float = 0.1,
+    std_dev: float = 0.5,
+    factor: int = 8,
+    channels: list[int] = [0, 1, 1, 0],
+):
+    batch_size, num_channels, latent_height, latent_width = latents.shape
+
+    shifted_latents = get_mean_shifted_latents(
+        latents,
+        shift=shift,
+        delta_shift=delta_shift,
+        channels=channels,
+    )
+    gauss_mask = get_2d_gaussian(
+        latent_height=latent_height,
+        latent_width=latent_width,
+        std_dev=std_dev,
+        center_x=0.0,
+        center_y=0.0,
+        factor=factor,
+        device=latents.device,
+    )
+    gauss_mask = F.interpolate(
+        gauss_mask,
+        size=(latent_height, latent_width),
+        mode="bilinear",
+        align_corners=False,
+    )
+
+    gauss_mask = gauss_mask.expand(batch_size, num_channels, -1, -1)
+
+    noised_latents = shifted_latents * (1 - gauss_mask) + latents * gauss_mask
+
+    return noised_latents
+
+
+class ColorSet(NamedTuple):
+    name: str
+    channels: list[int]
+
+
+# ref: Figure 28. Additional Result in various color Background with SD
+COLOR_SETS: list[ColorSet] = [
+    ColorSet("green", [0, 1, 1, 0]),
+    ColorSet("cyan", [0, 1, 0, 0]),
+    ColorSet("magenta", [0, -1, -1, -1]),
+    ColorSet("purple", [0, 0, -1, -1]),
+    ColorSet("black", [-1, 0, 0, 1]),
+    ColorSet("orange", [-1, -1, 1, 0]),
+    ColorSet("white", [0, 0, 0, -1]),
+    ColorSet("yellow", [0, -1, 1, -1]),
+]
+
+COLOR_SET_MAP: dict[str, ColorSet] = {c.name: c for c in COLOR_SETS}