[Diffusion] Support Multi-image Generation and Add Web UI Demo for QwenImage (#97)

Signed-off-by: samithuang <285365963@qq.com>
Signed-off-by: SamitHuang <285365963@qq.com>

docs/.nav.yml

@@ -8,6 +8,7 @@ nav:
       - examples/README.md
       - Offline Inference:
         - Qwen2.5-Omni: user_guide/examples/offline_inference/qwen2_5_omni.md
+        - Qwen2.5-Image: user_guide/examples/offline_inference/qwen_image.md
       - Online Serving: user_guide/examples/online_serving/README.md
     - General:
       - usage/*

examples/offline_inference/diffusion/qwen_image.py

@@ -1,20 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import torch
-
-from vllm_omni.entrypoints.omni import Omni
-
-model_name = "Qwen/Qwen-Image"
-prompt = "a cup of coffee on the table"
-if __name__ == "__main__":
-    m = Omni(
-        model=model_name,
-    )
-    image = m.generate(
-        prompt,
-        height=1024,
-        width=1024,
-        generator=torch.Generator(device="cuda").manual_seed(42),
-    )
-    image[0].save("qwen_image_output.png")

examples/offline_inference/qwen_image/README.md

@@ -0,0 +1,43 @@
+# Qwen-Image Offline Inference
+
+This folder provides two simple entrypoints for experimenting with `Qwen/Qwen-Image` using vLLM-Omni:
+
+- `text_to_image.py`: command-line script for single image generation.
+- `web_demo.py`: lightweight Gradio UI for interactive prompt/seed/CFG exploration.
+
+
+## Local CLI Usage
+
+```bash
+python text_to_image.py \
+  --prompt "a cup of coffee on the table" \
+  --seed 42 \
+  --cfg_scale 4.0 \
+  --num_images_per_prompt 1 \
+  --num_inference_steps 50 \
+  --height 1024 \
+  --width 1024 \
+  --output outputs/coffee.png
+```
+
+Key arguments:
+
+- `--prompt`: text description (string).
+- `--seed`: integer seed for deterministic sampling.
+- `--cfg_scale`: true CFG scale (model-specific guidance strength).
+- `--num_images_per_prompt`: number of images to generate per prompt (saves as `output`, `output_1`, ...).
+- `--num_inference_steps`: diffusion sampling steps (more steps = higher quality, slower).
+- `--height/--width`: output resolution (defaults 1024x1024).
+- `--output`: path to save the generated PNG.
+
+> ℹ️ Qwen-Image currently publishes best-effort presets at `1328x1328`, `1664x928`, `928x1664`, `1472x1140`, `1140x1472`, `1584x1056`, and `1056x1584`. Adjust `--height/--width` accordingly for the most reliable outcomes.
+
+## Web UI Demo
+
+Launch the gradio demo:
+
+```bash
+python gradio_demo.py --port 7862
+```
+
+Then open `http://localhost:7862/` on your local browser to interact with the web UI.

examples/offline_inference/qwen_image/gradio_demo.py

@@ -0,0 +1,214 @@
+import argparse
+from functools import lru_cache
+
+import gradio as gr
+import torch
+
+from vllm_omni.entrypoints.omni import Omni
+
+ASPECT_RATIOS: dict[str, tuple[int, int]] = {
+    "1:1": (1328, 1328),
+    "16:9": (1664, 928),
+    "9:16": (928, 1664),
+    "4:3": (1472, 1140),
+    "3:4": (1140, 1472),
+    "3:2": (1584, 1056),
+    "2:3": (1056, 1584),
+}
+ASPECT_RATIO_CHOICES = [f"{ratio} ({w}x{h})" for ratio, (w, h) in ASPECT_RATIOS.items()]
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Gradio demo for Qwen-Image offline inference.")
+    parser.add_argument("--model", default="Qwen/Qwen-Image", help="Diffusion model name or local path.")
+    parser.add_argument(
+        "--height",
+        type=int,
+        default=1328,
+        help="Default image height (must match one of the supported presets).",
+    )
+    parser.add_argument(
+        "--width",
+        type=int,
+        default=1328,
+        help="Default image width (must match one of the supported presets).",
+    )
+    parser.add_argument("--default-prompt", default="a cup of coffee on the table", help="Initial prompt shown in UI.")
+    parser.add_argument("--default-seed", type=int, default=42, help="Initial seed shown in UI.")
+    parser.add_argument("--default-cfg-scale", type=float, default=4.0, help="Initial CFG scale shown in UI.")
+    parser.add_argument(
+        "--num_inference_steps",
+        type=int,
+        default=50,
+        help="Default number of denoising steps shown in the UI.",
+    )
+    parser.add_argument("--ip", default="127.0.0.1", help="Host/IP for Gradio `launch`.")
+    parser.add_argument("--port", type=int, default=7862, help="Port for Gradio `launch`.")
+    parser.add_argument("--share", action="store_true", help="Share the Gradio demo publicly.")
+    args = parser.parse_args()
+    args.aspect_ratio_label = next(
+        (ratio for ratio, dims in ASPECT_RATIOS.items() if dims == (args.width, args.height)),
+        None,
+    )
+    if args.aspect_ratio_label is None:
+        supported = ", ".join(f"{ratio} ({w}x{h})" for ratio, (w, h) in ASPECT_RATIOS.items())
+        parser.error(f"Unsupported resolution {args.width}x{args.height}. Please pick one of: {supported}.")
+    return args
+
+
+@lru_cache(maxsize=1)
+def get_omni(model_name: str) -> Omni:
+    return Omni(model=model_name)
+
+
+def build_demo(args: argparse.Namespace) -> gr.Blocks:
+    device = "cuda"
+    omni = get_omni(args.model)
+
+    def run_inference(
+        prompt: str,
+        seed_value: float,
+        cfg_scale_value: float,
+        resolution_choice: str,
+        num_steps_value: float,
+        num_images_choice: float,
+    ):
+        if not prompt or not prompt.strip():
+            raise gr.Error("Please enter a non-empty prompt.")
+        ratio_label = resolution_choice.split(" ", 1)[0]
+        if ratio_label not in ASPECT_RATIOS:
+            raise gr.Error(f"Unsupported aspect ratio: {ratio_label}")
+        width, height = ASPECT_RATIOS[ratio_label]
+        try:
+            seed = int(seed_value)
+            num_steps = int(num_steps_value)
+            num_images = int(num_images_choice)
+        except (TypeError, ValueError) as exc:
+            raise gr.Error("Seed, inference steps, and number of images must be valid integers.") from exc
+        if num_steps <= 0:
+            raise gr.Error("Inference steps must be a positive integer.")
+        if num_images not in {1, 2, 3, 4}:
+            raise gr.Error("Number of images must be 1, 2, 3, or 4.")
+        generator = torch.Generator(device=device).manual_seed(seed)
+        images = omni.generate(
+            prompt.strip(),
+            height=height,
+            width=width,
+            generator=generator,
+            true_cfg_scale=float(cfg_scale_value),
+            num_inference_steps=num_steps,
+            num_images_per_prompt=num_images,
+            num_outputs_per_prompt=num_images,
+        )
+        return [img for img in images[:num_images]]
+
+    with gr.Blocks(
+        title="vLLM-Omni Web Serving Demo",
+        css="""
+        /* Left column button width */
+        .left-column button {
+            width: 100%;
+        }
+        /* Right preview area: fixed height, hide unnecessary buttons */
+        .fixed-image {
+            height: 660px;
+            display: flex;
+            flex-direction: column;
+            justify-content: center;
+            align-items: center;
+        }
+        .fixed-image .duplicate-button,
+        .fixed-image .svelte-drgfj2 {
+            display: none !important;
+        }
+        /* Gallery container: fill available space and center content */
+        #image-gallery {
+            width: 100%;
+            height: 100%;
+            display: flex;
+            align-items: center;
+            justify-content: center;
+        }
+        /* Gallery grid: center horizontally and vertically, set gap */
+        #image-gallery .grid {
+            display: flex;
+            flex-wrap: wrap;
+            justify-content: center;
+            align-items: center;
+            align-content: center;
+            gap: 16px;
+            width: 100%;
+            height: 100%;
+        }
+        /* Gallery grid items: center content */
+        #image-gallery .grid > div {
+            display: flex;
+            align-items: center;
+            justify-content: center;
+        }
+        /* Gallery images: limit max height, maintain aspect ratio */
+        .fixed-image img {
+            max-height: 660px !important;
+            width: auto !important;
+            object-fit: contain;
+        }
+        """,
+    ) as demo:
+        gr.Markdown("# vLLM-Omni Web Serving Demo")
+        gr.Markdown(f"**Model:** {args.model}")
+
+        with gr.Row():
+            with gr.Column(scale=1, elem_classes="left-column"):
+                prompt_input = gr.Textbox(
+                    label="Prompt",
+                    value=args.default_prompt,
+                    placeholder="Describe the image you want to generate...",
+                    lines=5,
+                )
+                seed_input = gr.Number(label="Seed", value=args.default_seed, precision=0)
+                cfg_input = gr.Number(label="CFG Scale", value=args.default_cfg_scale)
+                steps_input = gr.Number(
+                    label="Inference Steps",
+                    value=args.num_inference_steps,
+                    precision=0,
+                    minimum=1,
+                )
+                aspect_dropdown = gr.Dropdown(
+                    label="Aspect Ratio (W:H)",
+                    choices=ASPECT_RATIO_CHOICES,
+                    value=f"{args.aspect_ratio_label} ({ASPECT_RATIOS[args.aspect_ratio_label][0]}x{ASPECT_RATIOS[args.aspect_ratio_label][1]})",
+                )
+                num_images = gr.Dropdown(
+                    label="Number of images",
+                    choices=["1", "2", "3", "4"],
+                    value="1",
+                )
+                generate_btn = gr.Button("Generate", variant="primary")
+            with gr.Column(scale=2, elem_classes="fixed-image"):
+                gallery = gr.Gallery(
+                    label="Preview",
+                    columns=2,
+                    rows=2,
+                    height=660,
+                    allow_preview=True,
+                    show_label=True,
+                    elem_id="image-gallery",
+                )
+
+        generate_btn.click(
+            fn=run_inference,
+            inputs=[prompt_input, seed_input, cfg_input, aspect_dropdown, steps_input, num_images],
+            outputs=gallery,
+        )
+
+    return demo
+
+
+def main():
+    args = parse_args()
+    demo = build_demo(args)
+    demo.launch(server_name=args.ip, server_port=args.port, share=args.share)
+
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/qwen_image/text_to_image.py

@@ -0,0 +1,78 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+from pathlib import Path
+
+import torch
+
+from vllm_omni.entrypoints.omni import Omni
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Generate an image with Qwen-Image.")
+    parser.add_argument("--model", default="Qwen/Qwen-Image", help="Diffusion model name or local path.")
+    parser.add_argument("--prompt", default="a cup of coffee on the table", help="Text prompt for image generation.")
+    parser.add_argument("--seed", type=int, default=42, help="Random seed for deterministic results.")
+    parser.add_argument(
+        "--cfg_scale",
+        type=float,
+        default=4.0,
+        help="True classifier-free guidance scale specific to Qwen-Image.",
+    )
+    parser.add_argument("--height", type=int, default=1024, help="Height of generated image.")
+    parser.add_argument("--width", type=int, default=1024, help="Width of generated image.")
+    parser.add_argument(
+        "--output",
+        type=str,
+        default="qwen_image_output.png",
+        help="Path to save the generated image (PNG).",
+    )
+    parser.add_argument(
+        "--num_images_per_prompt",
+        type=int,
+        default=1,
+        help="Number of images to generate for the given prompt.",
+    )
+    parser.add_argument(
+        "--num_inference_steps",
+        type=int,
+        default=50,
+        help="Number of denoising steps for the diffusion sampler.",
+    )
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    generator = torch.Generator(device=device).manual_seed(args.seed)
+
+    omni = Omni(model=args.model)
+    images = omni.generate(
+        args.prompt,
+        height=args.height,
+        width=args.width,
+        generator=generator,
+        true_cfg_scale=args.cfg_scale,
+        num_inference_steps=args.num_inference_steps,
+        num_images_per_prompt=args.num_images_per_prompt,
+        num_outputs_per_prompt=args.num_images_per_prompt,
+    )
+
+    output_path = Path(args.output)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    suffix = output_path.suffix or ".png"
+    stem = output_path.stem or "qwen_image_output"
+    if args.num_images_per_prompt <= 1:
+        images[0].save(output_path)
+        print(f"Saved generated image to {output_path}")
+    else:
+        for idx, img in enumerate(images):
+            save_path = output_path.parent / f"{stem}_{idx}{suffix}"
+            img.save(save_path)
+            print(f"Saved generated image to {save_path}")
+
+
+if __name__ == "__main__":
+    main()

pyproject.toml

@@ -35,6 +35,8 @@ dependencies = [
     "resampy>=0.4.3",
     "diffusers==0.35.2",
     "accelerate==1.12.0",
+    "gradio==5.50",
+    "soundfile>=0.13.1"
     # "vllm==0.11.0",  # TODO: fix the entrypoints overwrite problem
 ]
 
@@ -48,8 +50,6 @@ dev = [
     "isort>=5.12.0",
     "mypy>=1.0.0",
     "pre-commit>=3.0.0",
-    "gradio>=5.49.1",
-    "soundfile>=0.13.1"
 ]
 
 docs = [

vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image.py

@@ -599,6 +599,9 @@ class QwenImagePipeline(
         num_inference_steps = req.num_inference_steps or num_inference_steps
         generator = req.generator or generator
         true_cfg_scale = req.true_cfg_scale or true_cfg_scale
+        req_num_outputs = getattr(req, "num_outputs_per_prompt", None)
+        if req_num_outputs and req_num_outputs > 0:
+            num_images_per_prompt = req_num_outputs
         # 1. check inputs
         # 2. encode prompts
         # 3. prepare latents and timesteps