Qwen-Image-Flash: NVIDIA's 4-Step DMD2 Distillation of Qwen-Image

What is Qwen-Image-Flash?

Qwen-Image-Flash is a four-step, DMD2-distilled build of Qwen/Qwen-Image from NVIDIA. Improved Distribution Matching Distillation was run through NVIDIA FastGen, NVIDIA Model Optimizer and NVIDIA AutoModel; only the transformer weights change, so the rest of the Qwen-Image pipeline (Qwen2.5-VL text encoder, Qwen tokenizer, Qwen-Image VAE, 60-layer transformer shape) is inherited unchanged.

• 20.43B parameters in the denoising transformer, 28.85B across the full learned pipeline
• 4 function evaluations instead of the base model's full schedule
• Packaged FlowMatch Euler scheduler with a static shift-3 trajectory, producing sigmas [1.0, 0.9, 0.75, 0.5, 0.0]
• Distilled with English captions; inherited Chinese-language capability is not evaluated in the model card

Running the distilled checkpoint

The official release is a Diffusers checkpoint, and the distillation internalizes the teacher's guidance: sampling needs num_inference_steps=4 together with true_cfg_scale=1.0. Dropping the weights into a Qwen-Image workflow that still applies CFG 4.0 would double the guidance.

from diffusers import QwenImagePipeline
import torch

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",
    width=1024, height=1024,
    num_inference_steps=4,
    true_cfg_scale=1.0,
).images[0]

The model card lists Diffusers 0.38.0 with Transformers 5.12.1, SGLang Diffusion, vLLM-Omni and TensorRT-LLM VisualGen as the supported runtimes, with H100 and B200 as the validated hardware. ComfyUI is not part of the official engine list: treat the single-file transformer weights as architecture-compatible with the Qwen-Image family rather than as an officially supported runtime.