Switch Space to FLUX.2-klein BSOD image editing

app.py

@@ -1,94 +1,243 @@
-import gradio as gr
-import numpy as np
 import random
+import threading
 
-# import spaces #[uncomment to use ZeroGPU]
-from diffusers import DiffusionPipeline
+import gradio as gr
 import torch
+from diffusers import Flux2KleinPipeline, Flux2Transformer2DModel, GGUFQuantizationConfig
+from PIL import Image, ImageDraw, ImageFont, ImageOps
+
+
+MODEL_ID = "black-forest-labs/FLUX.2-klein-4B"
+GGUF_URL = (
+    "https://huggingface.co/unsloth/FLUX.2-klein-4B-GGUF/resolve/main/"
+    "flux-2-klein-4b-Q4_K_M.gguf"
+)
+MAX_SEED = 2_147_483_647
+MAX_GENERATION_EDGE = 1024
+MIN_GENERATION_EDGE = 256
+SIZE_STEP = 32
+
+PIPELINE = None
+PIPELINE_LOCK = threading.Lock()
+
+BSOD_PROMPT = (
+    "Transform the reference photo into a BSOD-inspired scene. "
+    "Keep the main subject recognizable and preserve the overall composition. "
+    "Use blue-screen-of-death aesthetics, computer hardware, machines, robots, "
+    "cybernetic details, metallic structures, monitor glow, motherboard patterns, "
+    "industrial sci-fi atmosphere, neon blue diagnostics, clean high detail."
+)
+
+CSS = """
+.app-shell {
+    max-width: 1080px;
+    margin: 0 auto;
+}
+.hero {
+    padding: 8px 0 20px;
+}
+.hero h1 {
+    margin-bottom: 8px;
+}
+"""
+
+
+def _device() -> str:
+    return "cuda" if torch.cuda.is_available() else "cpu"
+
+
+def _dtype() -> torch.dtype:
+    return torch.bfloat16 if torch.cuda.is_available() else torch.float32
+
+
+def get_pipeline() -> Flux2KleinPipeline:
+    global PIPELINE
+
+    if PIPELINE is not None:
+        return PIPELINE
+
+    with PIPELINE_LOCK:
+        if PIPELINE is not None:
+            return PIPELINE
+
+        quantization_config = GGUFQuantizationConfig(compute_dtype=_dtype())
+        transformer = Flux2Transformer2DModel.from_single_file(
+            GGUF_URL,
+            config=MODEL_ID,
+            subfolder="transformer",
+            quantization_config=quantization_config,
+            torch_dtype=_dtype(),
+        )
+
+        pipe = Flux2KleinPipeline.from_pretrained(
+            MODEL_ID,
+            transformer=transformer,
+            torch_dtype=_dtype(),
+        )
+        pipe.vae.enable_slicing()
+
+        if torch.cuda.is_available():
+            pipe.enable_model_cpu_offload()
+        else:
+            pipe.to("cpu")
+
+        pipe.set_progress_bar_config(disable=True)
+        PIPELINE = pipe
+        return PIPELINE
+
+
+def _round_to_step(value: int, step: int = SIZE_STEP) -> int:
+    return max(step, int(round(value / step) * step))
+
+
+def _generation_size(image: Image.Image) -> tuple[int, int]:
+    width, height = image.size
+    longest_edge = max(width, height)
+    scale = min(1.0, MAX_GENERATION_EDGE / longest_edge) if longest_edge else 1.0
+
+    resized_width = max(MIN_GENERATION_EDGE, int(width * scale))
+    resized_height = max(MIN_GENERATION_EDGE, int(height * scale))
+
+    gen_width = _round_to_step(resized_width)
+    gen_height = _round_to_step(resized_height)
+
+    gen_width = max(MIN_GENERATION_EDGE, min(MAX_GENERATION_EDGE, gen_width))
+    gen_height = max(MIN_GENERATION_EDGE, min(MAX_GENERATION_EDGE, gen_height))
+    return gen_width, gen_height
+
+
+def _resize_for_model(image: Image.Image, width: int, height: int) -> Image.Image:
+    return image.resize((width, height), Image.Resampling.LANCZOS)
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo"  # Replace to the model you would like to use
 
-if torch.cuda.is_available():
-    torch_dtype = torch.float16
-else:
-    torch_dtype = torch.float32
+def _label_font() -> ImageFont.ImageFont | ImageFont.FreeTypeFont:
+    for font_name in ("DejaVuSans-Bold.ttf", "Arial.ttf"):
+        try:
+            return ImageFont.truetype(font_name, 36)
+        except OSError:
+            continue
+    return ImageFont.load_default()
 
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to(device)
 
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
+def _compose_comparison(original: Image.Image, bsod: Image.Image) -> Image.Image:
+    pad = 28
+    gap = 24
+    header_height = 74
+    bg_color = (10, 16, 30)
+    panel_color = (18, 30, 54)
+    text_color = (223, 236, 255)
+
+    left_w, left_h = original.size
+    right_w, right_h = bsod.size
+    panel_height = max(left_h, right_h)
+
+    total_width = pad * 2 + left_w + right_w + gap
+    total_height = pad * 2 + header_height + panel_height
+    canvas = Image.new("RGB", (total_width, total_height), bg_color)
+    draw = ImageDraw.Draw(canvas)
+    font = _label_font()
+
+    left_panel = (pad, pad + header_height, pad + left_w, pad + header_height + panel_height)
+    right_panel = (
+        pad + left_w + gap,
+        pad + header_height,
+        pad + left_w + gap + right_w,
+        pad + header_height + panel_height,
+    )
+
+    draw.rounded_rectangle(left_panel, radius=20, fill=panel_color)
+    draw.rounded_rectangle(right_panel, radius=20, fill=panel_color)
+
+    left_text_x = pad + 16
+    right_text_x = pad + left_w + gap + 16
+    text_y = pad + 18
+    draw.text((left_text_x, text_y), "original", fill=text_color, font=font)
+    draw.text((right_text_x, text_y), "bsod", fill=text_color, font=font)
+
+    left_y = pad + header_height + (panel_height - left_h) // 2
+    right_y = pad + header_height + (panel_height - right_h) // 2
+
+    canvas.paste(original, (pad, left_y))
+    canvas.paste(bsod, (pad + left_w + gap, right_y))
+    return canvas
 
 
-# @spaces.GPU #[uncomment to use ZeroGPU]
 def infer(
-    prompt,
-    negative_prompt,
-    seed,
-    randomize_seed,
-    width,
-    height,
-    guidance_scale,
-    num_inference_steps,
+    input_image: Image.Image,
+    extra_prompt: str,
+    seed: int,
+    randomize_seed: bool,
+    num_inference_steps: int,
+    guidance_scale: float,
     progress=gr.Progress(track_tqdm=True),
 ):
+    if input_image is None:
+        raise gr.Error("Upload a source image first.")
+
+    original = ImageOps.exif_transpose(input_image).convert("RGB")
+    width, height = _generation_size(original)
+    conditioning = _resize_for_model(original, width, height)
+
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
 
-    generator = torch.Generator().manual_seed(seed)
+    prompt = BSOD_PROMPT
+    if extra_prompt and extra_prompt.strip():
+        prompt = f"{prompt} Extra instructions: {extra_prompt.strip()}"
+
+    pipe = get_pipeline()
+    generator_device = "cuda" if torch.cuda.is_available() else "cpu"
+    generator = torch.Generator(device=generator_device).manual_seed(int(seed))
 
-    image = pipe(
+    result = pipe(
         prompt=prompt,
-        negative_prompt=negative_prompt,
-        guidance_scale=guidance_scale,
-        num_inference_steps=num_inference_steps,
+        image=conditioning,
         width=width,
         height=height,
+        guidance_scale=guidance_scale,
+        num_inference_steps=int(num_inference_steps),
         generator=generator,
     ).images[0]
 
-    return image, seed
+    comparison = _compose_comparison(original, result)
+    return comparison, result, seed
 
 
-examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
-]
-
-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(css=CSS) as demo:
+    with gr.Column(elem_classes=["app-shell"]):
+        with gr.Column(elem_classes=["hero"]):
+            gr.Markdown(
+                """
+                # Make It BSOD
+                Upload a normal photo and get a side-by-side comparison:
+                the left panel stays untouched, the right panel is regenerated
+                in a BSOD, computers, robots, and industrial sci-fi style.
+                """
+            )
 
         with gr.Row():
-            prompt = gr.Text(
-                label="Prompt",
-                show_label=False,
-                max_lines=1,
-                placeholder="Enter your prompt",
-                container=False,
+            input_image = gr.Image(
+                label="Original photo",
+                type="pil",
+                image_mode="RGB",
+            )
+            comparison_image = gr.Image(
+                label="Comparison",
+                type="pil",
             )
 
-            run_button = gr.Button("Run", scale=0, variant="primary")
-
-        result = 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,
+        with gr.Row():
+            extra_prompt = gr.Textbox(
+                label="Extra style instructions",
+                placeholder="Optional: chrome limbs, server room, broken CRTs, robot swarm...",
+                lines=2,
+            )
+            stylized_image = gr.Image(
+                label="BSOD only",
+                type="pil",
             )
 
+        with gr.Accordion("Generation settings", open=False):
             seed = gr.Slider(
                 label="Seed",
                 minimum=0,
@@ -96,59 +245,46 @@ with gr.Blocks(css=css) as demo:
                 step=1,
                 value=0,
             )
-
             randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+            num_inference_steps = gr.Slider(
+                label="Steps",
+                minimum=1,
+                maximum=50,
+                step=1,
+                value=12,
+            )
+            guidance_scale = gr.Slider(
+                label="Guidance scale",
+                minimum=1.0,
+                maximum=10.0,
+                step=0.1,
+                value=4.0,
+            )
+
+        run_button = gr.Button("Make it BSOD", variant="primary")
 
-            with gr.Row():
-                width = gr.Slider(
-                    label="Width",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,  # Replace with defaults that work for your model
-                )
-
-                height = gr.Slider(
-                    label="Height",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,  # Replace with defaults that work for your model
-                )
-
-            with gr.Row():
-                guidance_scale = gr.Slider(
-                    label="Guidance scale",
-                    minimum=0.0,
-                    maximum=10.0,
-                    step=0.1,
-                    value=0.0,  # Replace with defaults that work for your model
-                )
-
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=50,
-                    step=1,
-                    value=2,  # Replace with defaults that work for your model
-                )
-
-        gr.Examples(examples=examples, inputs=[prompt])
-    gr.on(
-        triggers=[run_button.click, prompt.submit],
+        gr.Examples(
+            examples=[
+                ["cold blue datacenter, mechanical arms, diagnostic overlays"],
+                ["retro windows crash screen, motherboard textures, chrome robot face"],
+                ["factory machines, server racks, terminal glow, cybernetic details"],
+            ],
+            inputs=[extra_prompt],
+        )
+
+    run_button.click(
         fn=infer,
         inputs=[
-            prompt,
-            negative_prompt,
+            input_image,
+            extra_prompt,
             seed,
             randomize_seed,
-            width,
-            height,
-            guidance_scale,
             num_inference_steps,
+            guidance_scale,
         ],
-        outputs=[result, seed],
+        outputs=[comparison_image, stylized_image, seed],
     )
 
+
 if __name__ == "__main__":
     demo.launch()

requirements.txt

@@ -1,6 +1,9 @@
-accelerate
-diffusers
-invisible_watermark
-torch
-transformers
-xformers
+accelerate>=1.10.0
+diffusers>=0.37.1
+gguf>=0.17.1
+gradio>=6.5.1
+huggingface_hub>=0.34.0
+Pillow>=11.3.0
+sentencepiece>=0.2.0
+torch>=2.6.0
+transformers>=4.57.0