Qwen-Image-2.1
An open-source image generation model released by Alibaba's Qwen team. Model weights are available for download on the Hugging Face platform, with support for local deployment, fine-tuning, and secondary development.
Core Facts
Qwen-Image-Edit is an image editing model launched by Alibaba's Tongyi Qianwen team, with capabilities for image generation and local editing.
Qwen Image is a text-to-image model launched by Alibaba.
With the image editing workflow of Qwen Image 2.1, the face-swapping process can be simplified to two images plus a prompt, with no masking operations required.
Timeline (last 90 days)
On an RTX PRO 6000, using BF16 precision, 40 steps and a batch size of 1, Qwen-Image-2.1 took an average of 9.378 seconds and a median of 9.375 seconds to generate 14 images of about 1 MP.
In the evaluation, the time for a single 2K image was about 5.55 times that of 1 MP and the peak VRAM about 1.54 times, lower than a pure quadratic prediction, because the official pipeline most likely performs the main denoising in latent space and the high-resolution stage may employ tiled or windowed attention.
In the food editing sample generated by Qwen-Image-2.1, the prompt asked for a top-down shot, but the output is closer to a 45-degree oblique high angle.
In the infographic "HOW RAIN BECOMES A RIVER" generated by Qwen-Image-2.1, the title and the four labels CLOUD, RAIN, SOIL and RIVER are all correct, but it is already raining in the cloud panel, and the first two stages are not as distinct as the labels suggest.
In this fixed-seed test round, Qwen-Image-2.1 succeeded on the first try for all four prompts that explicitly required rendered text.
The blue origami whale sticker PNG generated by Qwen-Image-2.1 has a valid alpha channel with values ranging from 0 to 255; of 1,048,576 pixels, 989,473 are not fully opaque, proving it is genuine transparency rather than a white or checkerboard background painted into the image.
Qwen-Image-2.1's warm-cache pipeline load time on an RTX PRO 6000 was 4.427 seconds.
In a test of 15 images, Qwen-Image-2.1 completed all of them, with no VRAM overflow or CUDA errors.
When generating about 1 MP images on an RTX PRO 6000, Qwen-Image-2.1's average peak allocated VRAM was 36.801 GiB, a bit over thirty percent of the 95.0 GiB of available VRAM reported by PyTorch.
When using stageload's release(module) to move parameters and buffers to the meta device and running Qwen-Image-2.1 in a stage-by-stage 'load at start, release at end' manner (the 16.3 GB text encoder and the 13.3 GB transformer loaded only one at a time), the peak on the same Mac dropped to 19.0 GB, and swap did not grow in either run.
40 more timeline events
All Facts (20)
Qwen-Image-Edit is an image editing model launched by Alibaba's Tongyi Qianwen team, with capabilities for image generation and local editing.
90%VerifiedQwen Image is a text-to-image model launched by Alibaba.
90%VerifiedWith the image editing workflow of Qwen Image 2.1, the face-swapping process can be simplified to two images plus a prompt, with no masking operations required.
85%VerifiedQwen's image generation capability is based on Diffusion Model technology.
80%VerifiedTesters mentioned that Qwen Image 2.1's image editing and reference image capabilities look promising.
70%VerifiedQwen-Image-2.1 is available through multiple channels including GitHub, ModelScope, and Hugging Face.
65%VerifiedQwen-Image-2.1 supports up to 10 reference images and provides precise local control.
65%VerifiedQwen-Image-2.1 natively supports the generation and editing of RGBA image layers, and can directly output images with an alpha channel.
65%VerifiedAlibaba's Qwen-Image-2.1 has open weights, uses a 7B architecture, integrates generation and editing in a single model, and natively supports transparent layers.
65%UnverifiedImage models can maintain INT8 or FP8 precision when offloaded, and Z-Image generation time on a B200 with a 10 GiB VRAM limit dropped from 38.4 seconds to 2.3 seconds
70%UnverifiedUsers who have already installed ComfyUI typically only need to update to a version that supports the model and download the corresponding model weight files to load and use Qwen-Image-2.1 via nodes.
60%UnverifiedQwen-Image-2.1 adopts a design of using a single open-weight checkpoint to support both image generation and editing (One open-weight checkpoint for both generation and editing).
60%UnverifiedQwen-Image-2.1 adopts a lightweight architecture with only 7B parameters.
60%UnverifiedThe official documentation lists application scenarios supported by Qwen-Image-2.1 including panoramic images, infographics, and virtual try-on.
60%UnverifiedIn a test of 15 images, Qwen-Image-2.1 completed all of them, with no VRAM overflow or CUDA errors.
50%UnverifiedOn an RTX PRO 6000, using BF16 precision, 40 steps and a batch size of 1, Qwen-Image-2.1 took an average of 9.378 seconds and a median of 9.375 seconds to generate 14 images of about 1 MP.
50%UnverifiedWhen generating about 1 MP images on an RTX PRO 6000, Qwen-Image-2.1's average peak allocated VRAM was 36.801 GiB, a bit over thirty percent of the 95.0 GiB of available VRAM reported by PyTorch.
50%UnverifiedQwen-Image-2.1's warm-cache pipeline load time on an RTX PRO 6000 was 4.427 seconds.
50%UnverifiedThe blue origami whale sticker PNG generated by Qwen-Image-2.1 has a valid alpha channel with values ranging from 0 to 255; of 1,048,576 pixels, 989,473 are not fully opaque, proving it is genuine transparency rather than a white or checkerboard background painted into the image.
50%UnverifiedIn this fixed-seed test round, Qwen-Image-2.1 succeeded on the first try for all four prompts that explicitly required rendered text.
50%