Nvidia Qwen-Image-Flash: 4-Step DMD2 Distillation of Qwen-Image

NVIDIA released Qwen-Image-Flash, a four-step, DMD2-distilled version of Qwen/Qwen-Image that cuts generation cost dramatically while keeping the base model's architecture.

What Is Qwen-Image-Flash?

Qwen-Image-Flash is a distilled few-step checkpoint of Qwen-Image built with Improved Distribution Matching Distillation (DMD2) using NVIDIA FastGen, NVIDIA Model Optimizer, and NVIDIA AutoModel. The distilled student weights replace the base transformer while the rest of the pipeline stays the same, so the model remains a drop-in Qwen-Image family member.

Key specs:

• Transformer: 20.43B parameters, same MMDiT architecture as Qwen-Image
• Steps: 4 function evaluations (NFE), down from the base model's default schedule
• Scheduler: packaged FlowMatch Euler with a static shift-3 trajectory, producing sigmas [1.0, 0.9, 0.75, 0.5, 0.0]
• Pipeline: Qwen2.5-VL text encoder, Qwen tokenizer, Qwen-Image VAE

The accompanying paper, Qwen-Image-Flash: Beyond Objective Design, describes a unified few-step recipe for both text-to-image generation and instruction-guided image editing in just 4 NFEs.

Availability

The official release is a Diffusers checkpoint. Run it with QwenImagePipeline from diffusers, using 4 inference steps, true_cfg_scale=1.0, and the packaged scheduler:

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