Alibaba Qwen Releases 7B Qwen-Image-2.1 for Unified Image Generation and Editing

The model is smaller than its predecessor. The original Qwen-Image shipped in August 2025 as a 20B model under Apache 2.0, with editing handled by a separate Qwen-Image-Edit checkpoint. Qwen-Image-2.1 folds both functions into one model at about a third of the size. Qwen calls it the most balanced and cost-effective model in the Qwen-Image series. The 7B figure covers the diffusion transformer only; the pipeline also loads an 8B Qwen3-VL encoder, a detail Qwen flags for capacity planning.

The GitHub repository lists four components. The transformer has 32 layers, 7 billion parameters, and a single-stream design with block-causal attention. The text encoder is Qwen3-VL 8B, which encodes text instructions and condition images into one representation. The VAE is a 64-channel RGBA autoencoder with 16x spatial compression and native transparency. The scheduler uses Flow Matching with Euler discrete scheduling and dynamic shifting.

Qwen attributes part of the model's speed to its attention mask. Text tokens use a token-level causal mask, while image tokens use a chunk-level bidirectional mask within each image. Qwen calls this mixed-granularity attention. The condition prefix sits before the noisy latent, so it never attends to it. Its keys and values therefore stay fixed across denoising steps. The model computes text and input images once, at the first step, and reuses that prefix KV cache for every remaining step. Savings grow with the number of reference images, which explains the multi-image speed claim.

Capabilities include native transparency, multi-reference editing, local control, native 2K output, and what Qwen describes as improved aesthetics. For transparency, the model generates RGBA images from text, edits transparent layers, and extracts subjects from photos; Qwen recommends a fixed prompt template for transparent output. Multi-reference editing accepts up to 10 reference images, and README examples include a group photo from six portraits and an outfit from five references. Local edits can target regions using circles, painted annotations, or separate masks, with identity preserved for people and products. Native 2K defaults to 2048 x 2048, with seven supported aspect ratios up to 2752 x 1536. Qwen highlights panoramas, infographics, storyboards, and virtual try-ons.

On Qwen-Image-Bench, Qwen's in-house benchmark, Qwen-Image-2.1 scores 60.28 overall. That places it above Nano Banana 2.0 at 59.82 and every listed open-weight model. FLUX 2 Max, a 32B open model, sits at 55.33. Six closed models score higher, led by GPT Image 2.5 Sunburst at 67.01. The comparison reflects Qwen's own benchmark chart rather than an independent evaluation.

Running the model requires PyTorch 2.4.0 or later, transformers 5.17 or later, Diffusers from source, accelerate, and pillow. A basic example loads QwenImage21Pipeline from Qwen/Qwen-Image-2.1 in bfloat16 and generates an image with 40 inference steps. The same pipeline handles editing when one or more reference images are passed. On smaller GPUs, pipe.enable_model_cpu_offload() reduces memory pressure.

For serving, vLLM-Omni adds FP8 quantization, prefix KV caching, CUDA Graph decode, and tensor parallelism. SGLang adds Cache-DiT, CUDA graphs, multi-GPU parallelism, and component offload. ComfyUI ships native nodes and converted weights. Beyond NVIDIA, the release covers AMD Radeon GPUs via ROCm and eight chip platforms via FlagOS. Qwen also released two prompt-rewriting models, fine-tuned Qwen3.5-VL 9B checkpoints for text-to-image and editing. They expand short prompts into detailed ones and can pick an aspect ratio. The Qwen Research License bars commercial use without a separate agreement.