Consolidate to single app.py entry point

- Rename app_diffusers.py to app.py (uses OvisImagePipeline)
- Backup original modular app.py as app_old.py for reference
- Update README.md app_file and short_description
- Update CLAUDE.md documentation

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

CLAUDE.md

@@ -12,12 +12,9 @@ Ovis-Image is a 7-billion parameter text-to-image generation model optimized for
 # Install dependencies
 pip install -r requirements.txt
 
-# Run web UI (custom implementation)
+# Run web UI (uses diffusers pipeline)
 python app.py
 
-# Run web UI (diffusers-based, simpler)
-python app_diffusers.py
-
 # Standalone inference test
 python ovis_image/test.py \
   --model_path <path/to/ovis_image.safetensors> \
@@ -40,9 +37,9 @@ Models auto-download from HuggingFace Hub. Set `HF_TOKEN` environment variable i
 - `ovis_image/model/layers.py` - Transformer blocks, attention, embeddings
 - `ovis_image/model/ops.py` - Attention backends (Flash3, SDPA, eager)
 
-**Two Entry Points:**
-- `app.py` - Loads individual components (modular, educational)
-- `app_diffusers.py` - Uses `OvisImagePipeline.from_pretrained()` (simpler)
+**Entry Point:**
+- `app.py` - Uses `OvisImagePipeline.from_pretrained()` for simple, clean loading
+- `app_old.py` - Backup of modular implementation that loads individual components (for reference)
 
 ## Configuration
 

README.md

@@ -5,10 +5,10 @@ colorFrom: purple
 colorTo: red
 sdk: gradio
 sdk_version: 6.0.2
-app_file: app_diffusers.py
+app_file: app.py
 pinned: false
 license: apache-2.0
-short_description: Official demo for Ovis-Image
+short_description: Ovis-Image UI with Apple-style CSS
 thumbnail: >-
   https://cdn-uploads.huggingface.co/production/uploads/636f4c6b5d2050767e4a1491/yMWqt9g4PC2SGCrgk6kRj.png
 ---

app.py

@@ -4,74 +4,27 @@ import gradio as gr
 import spaces
 import random
 import numpy as np
-from safetensors.torch import load_file
-from huggingface_hub import hf_hub_download
 
 from diffusers.utils import logging
 from PIL import Image
 
-from ovis_image.model.tokenizer import build_ovis_tokenizer
-from ovis_image.model.autoencoder import load_ae
-from ovis_image.model.hf_embedder import OvisEmbedder
-from ovis_image.model.model import OvisImageModel
-from ovis_image.sampling import generate_image
-from ovis_image import ovis_image_configs
+from diffusers import OvisImagePipeline
+
 
 logging.set_verbosity_error()
 
-# DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 MAX_SEED = np.iinfo(np.int32).max
 
 device = "cuda"
 _dtype = torch.bfloat16
 hf_token = os.getenv("HF_TOKEN")
 
-print("init ovis_image")
-model_config = ovis_image_configs["ovis-image-7b"]
-ovis_image = OvisImageModel(model_config)
-ovis_image_path = hf_hub_download(
-    repo_id="AIDC-AI/Ovis-Image-7B", 
-    filename="ovis_image.safetensors", 
-    token=hf_token,
-)
-model_state_dict = load_file(ovis_image_path)
-missing_keys, unexpected_keys = ovis_image.load_state_dict(model_state_dict)
-print(f"Load Missing Keys {missing_keys}")
-print(f"Load Unexpected Keys {unexpected_keys}")
-ovis_image = ovis_image.to(device=device, dtype=_dtype)
-ovis_image.eval()
-
-print("init vae")
-vae_path = hf_hub_download(
-    repo_id="AIDC-AI/Ovis-Image-7B", 
-    filename="ae.safetensors", 
+pipe = OvisImagePipeline.from_pretrained(
+    "AIDC-AI/Ovis-Image-7B",
     token=hf_token,
+    torch_dtype=torch.bfloat16
 )
-autoencoder = load_ae(
-    vae_path,
-    model_config.autoencoder_params,
-    device=device,
-    dtype=_dtype,
-    random_init=False,
-)
-autoencoder.eval()
-
-print("init ovis")
-# ovis_path = hf_hub_download(
-#     repo_id="AIDC-AI/Ovis-Image-7B", 
-#     subfolder="Ovis2.5-2B", 
-#     token=hf_token,
-# )
-ovis_tokenizer = build_ovis_tokenizer(
-    "AIDC-AI/Ovis2.5-2B",
-)
-ovis_encoder = OvisEmbedder(
-    model_path="AIDC-AI/Ovis2.5-2B",
-    random_init=False,
-    low_cpu_mem_usage=True,
-    torch_dtype=torch.bfloat16,
-).to(device=device, dtype=_dtype)
-
+pipe.to("cuda")
 
 examples = [
     "Five shimmering goldfish weave through crevices between stones; four are red-and-white, while one is silver-white. By the pond's edge, a golden shaded British Shorthair cat watches them intently, counting on blind luck. Watercolor style.",
@@ -582,60 +535,51 @@ def infer(
 
     print(f'inference with prompt: {prompt}, size: {height}x{width}, seed: {seed}, steps: {num_inference_steps}, cfg: {guidance_scale}')
 
-    image = generate_image(
-        device=next(ovis_image.parameters()).device,
-        dtype=_dtype,
-        model=ovis_image,
-        prompt=prompt,
-        autoencoder=autoencoder,
-        ovis_tokenizer=ovis_tokenizer,
-        ovis_encoder=ovis_encoder,
-        img_height=height,
-        img_width=width,
-        denoising_steps=num_inference_steps,
-        cfg_scale=guidance_scale,
-        seed=seed,
-    )
-    # bring into PIL format and save
-    image = image.clamp(-1, 1)
-    image = image.cpu().permute(0, 2, 3, 1).float().numpy()
-    image = (image * 255).round().astype("uint8")
-
-    return image[0], seed
+    generator = torch.Generator().manual_seed(seed)
+    image = pipe(
+        prompt,
+        negative_prompt="",
+        height=height,
+        width=width,
+        num_inference_steps=num_inference_steps,
+        true_cfg_scale=guidance_scale,
+        generator=generator,
+    ).images[0]
+
+    return image, seed
+
+custom_theme = gr.themes.Soft(
+    primary_hue=gr.themes.colors.blue,
+    secondary_hue=gr.themes.colors.slate,
+    neutral_hue=gr.themes.colors.gray,
+    spacing_size=gr.themes.sizes.spacing_lg,
+    radius_size=gr.themes.sizes.radius_lg,
+    text_size=gr.themes.sizes.text_md,
+    font=[gr.themes.GoogleFont("Inter"), "SF Pro Display", "-apple-system", "BlinkMacSystemFont", "system-ui", "sans-serif"],
+    font_mono=[gr.themes.GoogleFont("JetBrains Mono"), "SF Mono", "ui-monospace", "monospace"],
+).set(
+    body_background_fill='#f5f5f7',
+    body_background_fill_dark='#000000',
+    button_primary_background_fill='#0071e3',
+    button_primary_background_fill_hover='#0077ed',
+    button_primary_text_color='#ffffff',
+    block_background_fill='#ffffff',
+    block_background_fill_dark='#1d1d1f',
+    block_border_width='0px',
+    block_shadow='0 2px 12px rgba(0, 0, 0, 0.08)',
+    block_shadow_dark='0 2px 12px rgba(0, 0, 0, 0.4)',
+    input_background_fill='#ffffff',
+    input_background_fill_dark='#1d1d1f',
+    input_border_width='1px',
+    input_border_color='#d2d2d7',
+    input_border_color_dark='#424245',
+    input_shadow='none',
+    input_shadow_focus='0 0 0 4px rgba(0, 113, 227, 0.15)',
+)
 
 with gr.Blocks(
     title="Ovis-Image",
     fill_height=False,
-    theme=gr.themes.Soft(
-        primary_hue=gr.themes.colors.blue,
-        secondary_hue=gr.themes.colors.slate,
-        neutral_hue=gr.themes.colors.gray,
-        spacing_size=gr.themes.sizes.spacing_lg,
-        radius_size=gr.themes.sizes.radius_lg,
-        text_size=gr.themes.sizes.text_md,
-        font=[gr.themes.GoogleFont("Inter"), "SF Pro Display", "-apple-system", "BlinkMacSystemFont", "system-ui", "sans-serif"],
-        font_mono=[gr.themes.GoogleFont("JetBrains Mono"), "SF Mono", "ui-monospace", "monospace"],
-    ).set(
-        body_background_fill='#f5f5f7',
-        body_background_fill_dark='#000000',
-        button_primary_background_fill='#0071e3',
-        button_primary_background_fill_hover='#0077ed',
-        button_primary_text_color='#ffffff',
-        block_background_fill='#ffffff',
-        block_background_fill_dark='#1d1d1f',
-        block_border_width='0px',
-        block_shadow='0 2px 12px rgba(0, 0, 0, 0.08)',
-        block_shadow_dark='0 2px 12px rgba(0, 0, 0, 0.4)',
-        input_background_fill='#ffffff',
-        input_background_fill_dark='#1d1d1f',
-        input_border_width='1px',
-        input_border_color='#d2d2d7',
-        input_border_color_dark='#424245',
-        input_shadow='none',
-        input_shadow_focus='0 0 0 4px rgba(0, 113, 227, 0.15)',
-    ),
-    css=apple_css,
-    js=js_code,
 ) as demo:
     # Two-column layout - variant='panel' prevents responsive stacking
     with gr.Row(equal_height=False, variant="panel", elem_id="main-row"):
@@ -712,7 +656,7 @@ with gr.Blocks(
             result = gr.Image(
                 label="Result",
                 show_label=False,
-                type="numpy",
+                type="pil",
                 format="png",
             )
 
@@ -731,5 +675,11 @@ with gr.Blocks(
         outputs=[result, seed],
     )
 
+    # Load JS after DOM is ready (like Qwen-Image)
+    demo.load(None, None, None, js=js_code)
+
 if __name__ == '__main__':
-    demo.launch()
+    demo.launch(
+        theme=custom_theme,
+        css=apple_css,
+    )

app_diffusers.py → app_old.py

@@ -4,27 +4,74 @@ import gradio as gr
 import spaces
 import random
 import numpy as np
+from safetensors.torch import load_file
+from huggingface_hub import hf_hub_download
 
 from diffusers.utils import logging
 from PIL import Image
 
-from diffusers import OvisImagePipeline
-
+from ovis_image.model.tokenizer import build_ovis_tokenizer
+from ovis_image.model.autoencoder import load_ae
+from ovis_image.model.hf_embedder import OvisEmbedder
+from ovis_image.model.model import OvisImageModel
+from ovis_image.sampling import generate_image
+from ovis_image import ovis_image_configs
 
 logging.set_verbosity_error()
 
+# DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 MAX_SEED = np.iinfo(np.int32).max
 
 device = "cuda"
 _dtype = torch.bfloat16
 hf_token = os.getenv("HF_TOKEN")
 
-pipe = OvisImagePipeline.from_pretrained(
-    "AIDC-AI/Ovis-Image-7B",
+print("init ovis_image")
+model_config = ovis_image_configs["ovis-image-7b"]
+ovis_image = OvisImageModel(model_config)
+ovis_image_path = hf_hub_download(
+    repo_id="AIDC-AI/Ovis-Image-7B", 
+    filename="ovis_image.safetensors", 
+    token=hf_token,
+)
+model_state_dict = load_file(ovis_image_path)
+missing_keys, unexpected_keys = ovis_image.load_state_dict(model_state_dict)
+print(f"Load Missing Keys {missing_keys}")
+print(f"Load Unexpected Keys {unexpected_keys}")
+ovis_image = ovis_image.to(device=device, dtype=_dtype)
+ovis_image.eval()
+
+print("init vae")
+vae_path = hf_hub_download(
+    repo_id="AIDC-AI/Ovis-Image-7B", 
+    filename="ae.safetensors", 
     token=hf_token,
-    torch_dtype=torch.bfloat16
 )
-pipe.to("cuda")
+autoencoder = load_ae(
+    vae_path,
+    model_config.autoencoder_params,
+    device=device,
+    dtype=_dtype,
+    random_init=False,
+)
+autoencoder.eval()
+
+print("init ovis")
+# ovis_path = hf_hub_download(
+#     repo_id="AIDC-AI/Ovis-Image-7B", 
+#     subfolder="Ovis2.5-2B", 
+#     token=hf_token,
+# )
+ovis_tokenizer = build_ovis_tokenizer(
+    "AIDC-AI/Ovis2.5-2B",
+)
+ovis_encoder = OvisEmbedder(
+    model_path="AIDC-AI/Ovis2.5-2B",
+    random_init=False,
+    low_cpu_mem_usage=True,
+    torch_dtype=torch.bfloat16,
+).to(device=device, dtype=_dtype)
+
 
 examples = [
     "Five shimmering goldfish weave through crevices between stones; four are red-and-white, while one is silver-white. By the pond's edge, a golden shaded British Shorthair cat watches them intently, counting on blind luck. Watercolor style.",
@@ -535,51 +582,60 @@ def infer(
 
     print(f'inference with prompt: {prompt}, size: {height}x{width}, seed: {seed}, steps: {num_inference_steps}, cfg: {guidance_scale}')
 
-    generator = torch.Generator().manual_seed(seed)
-    image = pipe(
-        prompt,
-        negative_prompt="",
-        height=height,
-        width=width,
-        num_inference_steps=num_inference_steps,
-        true_cfg_scale=guidance_scale,
-        generator=generator,
-    ).images[0]
-
-    return image, seed
-
-custom_theme = gr.themes.Soft(
-    primary_hue=gr.themes.colors.blue,
-    secondary_hue=gr.themes.colors.slate,
-    neutral_hue=gr.themes.colors.gray,
-    spacing_size=gr.themes.sizes.spacing_lg,
-    radius_size=gr.themes.sizes.radius_lg,
-    text_size=gr.themes.sizes.text_md,
-    font=[gr.themes.GoogleFont("Inter"), "SF Pro Display", "-apple-system", "BlinkMacSystemFont", "system-ui", "sans-serif"],
-    font_mono=[gr.themes.GoogleFont("JetBrains Mono"), "SF Mono", "ui-monospace", "monospace"],
-).set(
-    body_background_fill='#f5f5f7',
-    body_background_fill_dark='#000000',
-    button_primary_background_fill='#0071e3',
-    button_primary_background_fill_hover='#0077ed',
-    button_primary_text_color='#ffffff',
-    block_background_fill='#ffffff',
-    block_background_fill_dark='#1d1d1f',
-    block_border_width='0px',
-    block_shadow='0 2px 12px rgba(0, 0, 0, 0.08)',
-    block_shadow_dark='0 2px 12px rgba(0, 0, 0, 0.4)',
-    input_background_fill='#ffffff',
-    input_background_fill_dark='#1d1d1f',
-    input_border_width='1px',
-    input_border_color='#d2d2d7',
-    input_border_color_dark='#424245',
-    input_shadow='none',
-    input_shadow_focus='0 0 0 4px rgba(0, 113, 227, 0.15)',
-)
+    image = generate_image(
+        device=next(ovis_image.parameters()).device,
+        dtype=_dtype,
+        model=ovis_image,
+        prompt=prompt,
+        autoencoder=autoencoder,
+        ovis_tokenizer=ovis_tokenizer,
+        ovis_encoder=ovis_encoder,
+        img_height=height,
+        img_width=width,
+        denoising_steps=num_inference_steps,
+        cfg_scale=guidance_scale,
+        seed=seed,
+    )
+    # bring into PIL format and save
+    image = image.clamp(-1, 1)
+    image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+    image = (image * 255).round().astype("uint8")
+
+    return image[0], seed
 
 with gr.Blocks(
     title="Ovis-Image",
     fill_height=False,
+    theme=gr.themes.Soft(
+        primary_hue=gr.themes.colors.blue,
+        secondary_hue=gr.themes.colors.slate,
+        neutral_hue=gr.themes.colors.gray,
+        spacing_size=gr.themes.sizes.spacing_lg,
+        radius_size=gr.themes.sizes.radius_lg,
+        text_size=gr.themes.sizes.text_md,
+        font=[gr.themes.GoogleFont("Inter"), "SF Pro Display", "-apple-system", "BlinkMacSystemFont", "system-ui", "sans-serif"],
+        font_mono=[gr.themes.GoogleFont("JetBrains Mono"), "SF Mono", "ui-monospace", "monospace"],
+    ).set(
+        body_background_fill='#f5f5f7',
+        body_background_fill_dark='#000000',
+        button_primary_background_fill='#0071e3',
+        button_primary_background_fill_hover='#0077ed',
+        button_primary_text_color='#ffffff',
+        block_background_fill='#ffffff',
+        block_background_fill_dark='#1d1d1f',
+        block_border_width='0px',
+        block_shadow='0 2px 12px rgba(0, 0, 0, 0.08)',
+        block_shadow_dark='0 2px 12px rgba(0, 0, 0, 0.4)',
+        input_background_fill='#ffffff',
+        input_background_fill_dark='#1d1d1f',
+        input_border_width='1px',
+        input_border_color='#d2d2d7',
+        input_border_color_dark='#424245',
+        input_shadow='none',
+        input_shadow_focus='0 0 0 4px rgba(0, 113, 227, 0.15)',
+    ),
+    css=apple_css,
+    js=js_code,
 ) as demo:
     # Two-column layout - variant='panel' prevents responsive stacking
     with gr.Row(equal_height=False, variant="panel", elem_id="main-row"):
@@ -656,7 +712,7 @@ with gr.Blocks(
             result = gr.Image(
                 label="Result",
                 show_label=False,
-                type="pil",
+                type="numpy",
                 format="png",
             )
 
@@ -675,11 +731,5 @@ with gr.Blocks(
         outputs=[result, seed],
     )
 
-    # Load JS after DOM is ready (like Qwen-Image)
-    demo.load(None, None, None, js=js_code)
-
 if __name__ == '__main__':
-    demo.launch(
-        theme=custom_theme,
-        css=apple_css,
-    )
+    demo.launch()