Upload folder using huggingface_hub

README.md

@@ -1,13 +1,33 @@
 ---
-title: Flux Cmo Compositional
-emoji: 🐠
-colorFrom: yellow
-colorTo: indigo
+title: FLUX.1-dev CMO Compositional
+emoji: 🎨
+colorFrom: gray
+colorTo: purple
 sdk: gradio
 sdk_version: 6.20.0
-python_version: '3.12'
 app_file: app.py
-pinned: false
+short_description: FLUX.1-dev fine-tuned with CMO for compositional generation
+python_version: "3.12"
+startup_duration_timeout: 30m
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# FLUX.1-dev CMO — Compositional Text-to-Image
+
+This Space demonstrates **Correlation-Weighted Multi-Reward Optimization (CMO)** applied to FLUX.1-dev.
+CMO improves compositional text-to-image generation by analyzing the correlation structure among
+concept-wise rewards (object existence, attributes, numeracy, size, spatial relations) and adaptively
+weighting conflicting concepts during training.
+
+## Model
+
+- **Base:** [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev)
+- **LoRA adapter:** [Bruece/FLUX.1-dev-CMO](https://huggingface.co/Bruece/FLUX.1-dev-CMO)
+
+## Usage
+
+Enter a compositional prompt (e.g. "A red apple is on the left of the yellow banana") and click Run.
+The model is loaded in bfloat16 with a tiny VAE for fast preview.
+
+## License
+
+FLUX.1-dev carries a non-commercial license. See [LICENSE](https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md).

app.py

@@ -0,0 +1,186 @@
+import os
+
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+
+import spaces  # MUST come before any CUDA-touching import
+import torch
+import gradio as gr
+import random
+import numpy as np
+from diffusers import DiffusionPipeline, AutoencoderTiny, AutoencoderKL
+from peft import PeftModel
+
+dtype = torch.bfloat16
+device = "cuda"
+
+# Tiny VAE for fast preview, good VAE for final output (same pattern as official FLUX.1-dev space)
+taef1 = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to(device)
+good_vae = AutoencoderKL.from_pretrained(
+    "black-forest-labs/FLUX.1-dev", subfolder="vae", torch_dtype=dtype
+).to(device)
+
+# Load base FLUX.1-dev with tiny VAE, then apply CMO LoRA adapter
+BASE_MODEL = "black-forest-labs/FLUX.1-dev"
+CMO_LORA = "Bruece/FLUX.1-dev-CMO"
+
+pipe = DiffusionPipeline.from_pretrained(
+    BASE_MODEL, torch_dtype=dtype, vae=taef1
+).to(device)
+
+# Load and merge the CMO LoRA adapter
+pipe.transformer = PeftModel.from_pretrained(pipe.transformer, CMO_LORA)
+pipe.transformer = pipe.transformer.merge_and_unload()
+pipe = pipe.to(device)
+
+torch.cuda.empty_cache()
+
+MAX_SEED = np.iinfo(np.int32).max
+MAX_IMAGE_SIZE = 2048
+
+
+def _estimate(prompt, num_inference_steps, *args, **kwargs):
+    return min(180, 30 + int(num_inference_steps * 4))
+
+
+@spaces.GPU(duration=_estimate)
+def generate(
+    prompt: str,
+    seed: int = 42,
+    randomize_seed: bool = True,
+    width: int = 1024,
+    height: int = 1024,
+    guidance_scale: float = 4.5,
+    num_inference_steps: int = 40,
+    progress: gr.Progress = gr.Progress(track_tqdm=True),
+):
+    """Generate an image from a text prompt using FLUX.1-dev fine-tuned with CMO.
+
+    CMO (Correlation-Weighted Multi-Reward Optimization) improves compositional
+    text-to-image generation by adaptively weighting conflicting concept rewards
+    (object existence, attributes, numeracy, size, spatial relations).
+
+    Args:
+        prompt: Text description of the image to generate.
+        seed: RNG seed for reproducibility.
+        randomize_seed: If True, pick a random seed each run.
+        width: Output image width in pixels.
+        height: Output image height in pixels.
+        guidance_scale: Classifier-free guidance scale.
+        num_inference_steps: Number of denoising steps.
+    """
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    seed = int(seed)
+    generator = torch.Generator().manual_seed(seed)
+    image = pipe(
+        prompt=prompt,
+        height=height,
+        width=width,
+        num_inference_steps=num_inference_steps,
+        guidance_scale=guidance_scale,
+        generator=generator,
+    ).images[0]
+    return image, seed
+
+
+CSS = """
+#col-container { max-width: 1100px; margin: 0 auto; }
+.dark .gradio-container { color: var(--body-text-color); }
+"""
+
+EXAMPLES = [
+    ["A red apple is on the left of the yellow banana"],
+    ["Two Tyrannosaurus rexes engaged in a boxing match"],
+    ["a photo of a black kite and a green bear"],
+    ["A brown cow wearing yellow sunglasses in a pastel chalk drawing"],
+    ["a cat holding a sign that says hello world"],
+    ["The green plant was on top of the white nightstand"],
+]
+
+with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.Markdown(
+            """# FLUX.1-dev CMO — Compositional Text-to-Image
+FLUX.1-dev fine-tuned with **Correlation-Weighted Multi-Reward Optimization (CMO)** for improved compositional generation.
+[[Paper](https://huggingface.co/papers/2603.18528)] [[Code](https://github.com/TheDarkKnight-21th/CMO)] [[Model](https://huggingface.co/Bruece/FLUX.1-dev-CMO)]
+"""
+        )
+        with gr.Row():
+            prompt = gr.Textbox(
+                label="Prompt",
+                show_label=False,
+                max_lines=1,
+                placeholder="Enter your prompt (e.g. 'A red apple is on the left of the yellow banana')",
+                container=False,
+                scale=4,
+            )
+            run_button = gr.Button("Run", variant="primary", scale=1)
+
+        result = gr.Image(label="Result", show_label=False)
+
+        with gr.Accordion("Advanced settings", open=False):
+            seed = gr.Slider(
+                label="Seed",
+                minimum=0,
+                maximum=MAX_SEED,
+                step=1,
+                value=42,
+            )
+            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+            with gr.Row():
+                width = gr.Slider(
+                    label="Width",
+                    minimum=256,
+                    maximum=MAX_IMAGE_SIZE,
+                    step=32,
+                    value=1024,
+                )
+                height = gr.Slider(
+                    label="Height",
+                    minimum=256,
+                    maximum=MAX_IMAGE_SIZE,
+                    step=32,
+                    value=1024,
+                )
+            with gr.Row():
+                guidance_scale = gr.Slider(
+                    label="Guidance Scale",
+                    minimum=1,
+                    maximum=15,
+                    step=0.1,
+                    value=4.5,
+                )
+                num_inference_steps = gr.Slider(
+                    label="Number of inference steps",
+                    minimum=1,
+                    maximum=50,
+                    step=1,
+                    value=40,
+                )
+
+        gr.Examples(
+            examples=EXAMPLES,
+            fn=generate,
+            inputs=[prompt],
+            outputs=[result, seed],
+            cache_examples=True,
+            cache_mode="lazy",
+        )
+
+        gr.on(
+            triggers=[run_button.click, prompt.submit],
+            fn=generate,
+            inputs=[
+                prompt,
+                seed,
+                randomize_seed,
+                width,
+                height,
+                guidance_scale,
+                num_inference_steps,
+            ],
+            outputs=[result, seed],
+            api_name="generate",
+        )
+
+demo.launch(mcp_server=True)

requirements.txt

@@ -0,0 +1,6 @@
+diffusers
+peft
+accelerate
+torchvision
+safetensors
+sentencepiece