akhaliq/Qwen-Image-Flash

Qwen-Image-Flash

Custom-frontend Gradio Space that generates images from text prompts using `nvidia/Qwen-Image-Flash` — a four-step, DMD2-distilled version of `Qwen/Qwen-Image`.

Architecture

This Space is built with `gradio.Server` rather than gr.Blocks: it pairs a custom vanilla-HTML/CSS/JS frontend (index.html) with Gradio's backend engine — the queue, concurrency control, ZeroGPU allocation, and gradio_client compatibility — without rendering a Gradio component UI.

• —`app.py` — a gradio.Server (FastAPI subclass) app:
• —@app.get("/") serves the static index.html frontend.
• —@app.api def generate_image(...) / @spaces.GPU is the queued backend endpoint the JS client calls. The 20B pipeline loads lazily inside the GPU-decorated handler (so it never tries pipe.to("cuda") on the CPU-only Space builder).
• —`index.html` — self-contained frontend (no build step). It uses the Gradio JS client (@gradio/client) via const c = await Client.connect(window.location.origin); c.predict("/generate_image", {...}) so requests go through Gradio's queue instead of a raw fetch(). Add a prompt, negative prompt, width/height/steps/seed, click Generate, and download the PNG.

The endpoint is also callable programmatically:

from gradio_client import Client
c = Client("<your-space>")            # or your Space id, e.g. "user/qwen-image-flash"
filedata, seed, info = c.predict(
    prompt="A red fox in a snowy pine forest at golden hour, photorealistic, sharp focus, soft bokeh",
    negative_prompt="",
    width=1024, height=1024,
    num_inference_steps=4,
    seed=42, randomize_seed=False,
    api_name="/generate_image",
)

How the model works

The model is distributed as a full Diffusers QwenImagePipeline. The student retains the base Qwen-Image transformer architecture; its weights are replaced by DMD2-distilled student weights. Because the teacher target used CFG=4.0 during distillation, that guidance is internalized by the student, so inference uses true_cfg_scale=1.0 (and guidance_scale=None / negative_prompt=None by default) to avoid applying guidance a second time. The packaged shift-3 FlowMatch Euler scheduler runs four denoising steps.

The tested output setting is 1024 × 1024; width and height must be divisible by 16 (the app enforces this for you).

Defaults in this Space

A negative prompt box is exposed for experimentation, but note the distillation assumes no negative prompt. Seeds can be fixed or randomized.

Run locally

pip install -r requirements.txt
python app.py
# then open the local URL it prints (e.g. http://127.0.0.1:7860)

Requires a CUDA GPU (Hopper H100 / Blackwell B200 family). The first generate request triggers the ~20B pipeline download + .to("cuda"), so expect a long warm-up. On a Hugging Face Space this runs under ZeroGPU via the @spaces.GPU decorator.

Usage (library)

import torch
from diffusers import QwenImagePipeline

pipe = QwenImagePipeline.from_pretrained(
    "nvidia/Qwen-Image-Flash",
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    prompt="A red fox in a snowy pine forest at golden hour, photorealistic, sharp focus, soft bokeh",
    width=1024,
    height=1024,
    num_inference_steps=4,
    true_cfg_scale=1.0,
    guidance_scale=None,
    negative_prompt=None,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

image.save("qwen-image-flash.png")

License

Use of this model is governed by the NVIDIA Open Model Agreement (additional info: Apache License 2.0). The model does not include a safety checker — deployers should add appropriate safeguards.