# Qwen-Image-2.1

> Source: https://aiwiki.ai/wiki/qwen_image_2_1
> Updated: 2026-09-25
> Fact-checked: 2026-09-25
> Categories: AI Models, Chinese AI, Diffusion Models, Image Generation
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Qwen-Image-2.1." aiwiki.ai, 25 Sept 2026. https://aiwiki.ai/wiki/qwen_image_2_1
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Qwen-Image-2.1** is an open-weight image generation and image editing model released by [Alibaba](https://aiwiki.ai/wiki/alibaba)'s [Qwen](https://aiwiki.ai/wiki/qwen) team on September 20, 2026 [1][3]. A single checkpoint handles [text-to-image](https://aiwiki.ai/wiki/text-to-image_models) generation and instruction-based editing, and it can output either ordinary images or transparent RGBA images depending on the prompt [1][2]. Its visual generation component is a 7-billion-parameter single-stream [diffusion transformer](https://aiwiki.ai/wiki/diffusion_transformer) with 32 layers, conditioned by a [Qwen3-VL](https://aiwiki.ai/wiki/qwen3_vl) 8B vision-language model that serves as the text encoder [1][3]. The weights are published on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) and [ModelScope](https://aiwiki.ai/wiki/modelscope) under the Qwen Research License Agreement, which permits research and evaluation use only; commercial use requires a separate license from Qwen [2][4]. On September 21, 2026, Qwen's official accounts on X stated that outputs "are not part of the licensed Materials" and that users retain the rights to the images they generate; the license text itself was not changed [20][21][22].

The model is a new entry in the [Qwen-Image](https://aiwiki.ai/wiki/qwen_image) series, which began with the 20-billion-parameter, Apache 2.0 Qwen-Image in August 2025. The two most recent flagship releases before it, Qwen-Image-2.0 (February 2026) and [Qwen-Image-3.0](https://aiwiki.ai/wiki/qwen_image_3) (July 2026), were offered as hosted models without downloadable weights [8][17][18]. Qwen-Image-2.0 had already unified generation and editing in one model and, according to its technical report, pairs a Qwen3-VL condition encoder with a diffusion transformer [8][9]. As of September 23, 2026, Qwen-Image-2.1 was the first new Qwen-Image generation model the Qwen organization had published on Hugging Face since Qwen-Image-2512 in December 2025 [18].

## Release

The Qwen team announced Qwen-Image-2.1 on its blog on September 20, 2026, under the title "Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation," and published the weights, code and a Hugging Face demo the same day [1][2][3]. The blog summarizes the release as four improvements: a compact and efficient architecture, native transparency with unified creation and editing, more versatile editing, and more realistic textures and aesthetics [1].

Several inference projects shipped support on release day, according to the official repository: the Hugging Face Diffusers library (a `QwenImage21Pipeline` added in Diffusers pull request #14804), [ComfyUI](https://aiwiki.ai/wiki/comfyui) (with repackaged weights at Comfy-Org/Qwen-Image-2.1), vLLM-Omni, SGLang-Diffusion and LightX2V [3][10]. The repository also points to ModelScope's DiffSynth-Studio for download, online generation and [LoRA](https://aiwiki.ai/wiki/lora) training, to AMD Radeon GPUs through [ROCm](https://aiwiki.ai/wiki/rocm), and to the FlagOS software stack, under which prebuilt images and weights for eight chip platforms were published [3]. For users in mainland China, the repository names the wuli.art site as offering free access to the model's features [3].

## Architecture

According to the official repository, Qwen-Image-2.1 has four components [3]:

| Component | Description |
| --- | --- |
| Transformer | Single-stream DiT, 32 layers, 7B parameters, block-causal attention [1][3] |
| Text encoder | Qwen3-VL 8B, which encodes both text instructions and condition images [3] |
| VAE | 64-channel RGBA [autoencoder](https://aiwiki.ai/wiki/autoencoder) with 16x spatial compression [3] |
| Scheduler | [Flow matching](https://aiwiki.ai/wiki/flow_matching) with Euler discrete sampling and dynamic shifting [3] |

The configuration files published with the weights are consistent with this description. The transformer configuration lists 32 layers, 32 attention heads with a head dimension of 128, and a `causal_condition` flag set to true; its weight index gives a total size of 14,230,249,472 bytes, and the tensor headers of its two weight files list 7,115,124,736 parameters, all stored in BF16. The text encoder is a `Qwen3VLForConditionalGeneration` model whose weight index reports 8,767,123,696 parameters, and the [variational autoencoder](https://aiwiki.ai/wiki/variational_autoencoder) takes four input channels (RGB plus alpha) and uses a 64-channel latent [5]. The "7B" figure in Qwen's materials therefore refers to the image generator alone, not to the full pipeline [1][13].

### Mixed-granularity attention and prefix KV cache

The Qwen team describes the attention design as "mixed-granularity." Text tokens, including the system prefix and any editing instruction, use a token-level causal mask, while image tokens use a chunk-level mask that is bidirectional within an image block. The repository gives the rule as `(q_idx >= kv_idx) or same_image_block` [1][3]. Because the input images and the instruction do not change during denoising, their keys and values are computed at the first denoising step, cached, and reused for every later step. Qwen says this [KV cache](https://aiwiki.ai/wiki/kv_cache) reuse improves inference speed and lowers memory use, with the largest benefit when editing with several reference images [1][3]. The published diagram labels the prefix (system prefix, optional input images and edit instruction) "compute once, reuse K/V" and the target image "compute each step" [1].

## Capabilities

### Native transparency

Transparent image generation is the main new capability. In December 2025 the Qwen team had released Qwen-Image-Layered as a dedicated model supporting transparent image generation; Qwen-Image-2.1 folds transparency into the general model and uses the prompt to decide whether to output a regular image or one with an alpha channel [1]. Qwen's showcase covers single transparent subjects, multi-element transparent compositions, edits to transparent images that keep the transparent background (for example, changing a character's expression or replacing the word "BLOOM" with "Qwen-Image" on a transparent layer), and extraction of a subject from an ordinary RGB photograph as an RGBA layer [1]. The [model card](https://aiwiki.ai/wiki/model_card) and repository recommend a fixed prompt pattern for transparent output: `This is an RGBA image with transparency. <your description>. The image has alpha channel and the background is transparent.` [2][3]

### Editing

Editing improvements fall into four areas in Qwen's account [1]:

- **Multiple references.** Up to 10 input images can be combined into one composition. Qwen's examples include a group photograph assembled from six individual portraits, a virtual try-on from five inputs (a model, clothing, shoes, a bag and a hat) and a furnished room built from 10 furniture images.
- **Local editing.** Edit regions can be marked with colored circles, painted annotations, or a separate mask image that leaves the original image unobscured. Qwen also shows successive local edits assembled into a short animation.
- **Fidelity.** Qwen says the model better preserves facial identity in portrait edits and keeps the text, texture and shape of products consistent.
- **Task coverage.** Qwen demonstrates panorama generation from a selfie, infographic generation from a model photograph, and storyboard generation from a three-view character sheet.

### Output resolution and defaults

The model natively outputs 2K images. The repository's recommended sizes run from 2048 x 2048 for a square image to 2752 x 1536 for 16:9, and the default sampler setting is 40 denoising steps [3].

| Aspect ratio | Width x height |
| --- | --- |
| 1:1 | 2048 x 2048 |
| 4:3 / 3:4 | 2400 x 1792 / 1792 x 2400 |
| 3:2 / 2:3 | 2528 x 1696 / 1696 x 2528 |
| 16:9 / 9:16 | 2752 x 1536 / 1536 x 2752 |

Qwen also released two prompt-rewriting models on September 20, Qwen-Image-2.1-PE-T2I for text-to-image and Qwen-Image-2.1-PE-I2I for editing. The repository describes them as fine-tuned Qwen3.5-VL 9B checkpoints that expand short prompts into the long, detailed prompts the model works best with, and that can also suggest an output aspect ratio [3].

## Benchmark claims

Qwen's launch material includes one benchmark: an overall-score chart on Qwen-Image-Bench, a benchmark the Qwen team itself built and published in May 2026 [1][6][7]. Qwen-Image-Bench uses 1,000 bilingual prompts written with professional artists and scores five pillars (quality, aesthetics, alignment, real-world fidelity and creative generation) with Q-Judger, a judge model based on Qwen3.6-27B that Qwen trained on more than 130,000 expert-annotated prompt-image pairs [6]. The benchmark's own published leaderboard notes that its scores are computed on images generated from Chinese prompts [6].

In the chart published with the Qwen-Image-2.1 announcement, which covers 29 models, Qwen-Image-2.1 scores 60.28 and ranks seventh [1]. Selected entries from that chart:

| Model | Overall score (Qwen-Image-Bench, Qwen's chart) | Parameters shown in chart |
| --- | --- | --- |
| GPT Image 2.5 Sunburst | 67.01 | Undisclosed |
| [GPT Image 2](https://aiwiki.ai/wiki/gpt_image_2) | 64.69 | Undisclosed |
| [Grok Imagine](https://aiwiki.ai/wiki/grok_imagine) 2.0 | 63.47 | Undisclosed |
| Qwen Image 3 Pro | 62.36 | Undisclosed |
| [Muse Image](https://aiwiki.ai/wiki/muse_image) | 62.34 | Undisclosed |
| MAI Image 2.5 Pro | 61.02 | Undisclosed |
| **Qwen-Image-2.1** | **60.28** | **7B** |
| [Nano Banana](https://aiwiki.ai/wiki/nano_banana) 2.0 | 59.82 | Undisclosed |
| GPT Image 1.5 | 59.65 | Undisclosed |
| [Seedream](https://aiwiki.ai/wiki/seedream) 5 Pro | 59.53 | Undisclosed |
| [Nano Banana Pro](https://aiwiki.ai/wiki/nano_banana_pro) | 59.45 | Undisclosed |
| Qwen Image 2.0 Pro | 57.84 | Undisclosed |
| [FLUX.2](https://aiwiki.ai/wiki/flux_2) Max | 55.33 | 32B |
| Qwen Image 2512 | 52.06 | 20B |
| [Hunyuan Image 3.0](https://aiwiki.ai/wiki/hunyuan_image_3) | 50.81 | 80B |
| Qwen Image (August 2025) | 49.23 | 20B |
| GLM Image | 48.19 | 16B |

The chart marks the parameter counts of closed models as undisclosed. Among the entries for which it shows a parameter count, Qwen-Image-2.1 has the highest score, and it scores above several closed systems, including Nano Banana 2.0, GPT Image 1.5 and [Nano Banana Pro](https://aiwiki.ai/wiki/nano_banana_pro), while trailing six others, including Qwen's own hosted Qwen Image 3 Pro [1]. The scores the chart shows for models already on the benchmark's public leaderboard (for example 64.69 for [GPT Image 2](https://aiwiki.ai/wiki/gpt_image_2) and 59.82 for Nano Banana 2.0) match that leaderboard [1][6].

Press coverage stressed that the result is Qwen's own measurement. The Decoder reported that the model "beats most closed models on Qwen's own benchmark, the team claims, though independent benchmarks are still pending" [11], and eWeek wrote that because Alibaba developed the benchmark, its results "should be read as the company's measurement rather than an independent ranking" [13]. 36Kr described the model as ranking first among open-source models, with a score "even exceeding Nano Banana 2.0" [14].

## License

Qwen-Image-2.1 is not released under [Apache 2.0](https://aiwiki.ai/wiki/apache_license), unlike the 2025 open-weight Qwen-Image models. The Hugging Face card lists the license as "qwen-research," and the LICENSE file is the Qwen Research License Agreement, dated September 20, 2026, with Hangzhou Tongyi Laboratory Technology Co., Ltd. as licensor [2][4]. Its main terms [4]:

- Use, reproduction, distribution and modification are granted "FOR NON-COMMERCIAL PURPOSES ONLY," with non-commercial defined as "for research or evaluation purposes only." A second clause says the licensee "shall not use the Materials for any commercial purpose without obtaining a separate commercial license," which is requested by email.
- "Materials" means Qwen's "proprietary Qwen and Documentation," where "Qwen" covers the models, software and algorithms Qwen distributes, including trained weights, parameters, and model, inference, training and fine-tuning code. The agreement has no definition of outputs and no clause on who owns them. Outputs appear in three places: the "Built with Qwen" rule below, a clause that ends the license for anyone who sues claiming the Materials "or any output therefrom" infringe their rights, and the warranty disclaimers.
- Redistributors must pass on a copy of the agreement, mark modified files, and keep a prescribed attribution notice in a "Notice" text file.
- Anyone who uses the Materials "or any outputs or results therefrom to create, train, fine-tune, or improve an AI model that is distributed or made available" must prominently display "Built with Qwen" or "Improved using Qwen" in the related product documentation. "Qwen" may not be the primary name of a derivative product, although descriptive use such as "fine-tuned from Qwen Image" is allowed.
- Qwen keeps ownership of the Materials; subject to compliance with the agreement, licensees own the derivative works and modifications they make.
- The agreement is governed by Chinese law, with the People's Courts in Hangzhou having exclusive jurisdiction.

Early coverage read the agreement as a bar on commercial use. The Decoder wrote on release day that the license "bars commercial use, so business users must apply to Qwen for a separate license" [11], and heise said on September 22 that commercial use, "for example, in one's own product or service," requires a separate agreement with Alibaba [24]. MIXED, in a September 21 article that quotes only the agreement text, summarized the arrangement: "The download is free; using it to make anything you sell is not." [12] Users on the model's Hugging Face page and [GitHub](https://aiwiki.ai/wiki/github) repository asked Qwen to return to Apache 2.0 [26][27].

### Qwen's statement on outputs

Many of the questions concerned generated images. On September 20, an X user asked Qwen to "clarify license around outputs," noting that outputs were "not specifically mentioned" [21]. The official @Alibaba_Qwen account replied on September 21, 2026 (UTC): "Outputs are not part of the licensed Materials. Users retain the rights to images and other content they generate using the model." [21] Just over an hour later, the official Qwen Developers account (@QwenDevs) posted the same two sentences as a standalone answer, saying the team had received "a lot of questions about the license, especially around model outputs" [20].

The statement was made on social media. As of September 25, 2026, the LICENSE file on Hugging Face had not been edited since it was added on September 18, 2026, and the model card still points only to the Qwen Research License Agreement [2][22]. The two sentences address rights in generated content; they do not mention commercial use or the clause barring commercial use of the Materials. GIGAZINE, reporting on September 24, put the two side by side: the license "includes a clause prohibiting commercial use," while Alibaba's official X account "explicitly states" that rights to generated images belong to the user [23].

## Community derivatives

Community developers began repackaging the model within days of release, including GGUF builds of the diffusion transformer, such as unsloth/Qwen-Image-2.1-GGUF [15][16]. One uploader, pottokao, applied [Heretic](https://aiwiki.ai/wiki/heretic_software), an open-source tool for automatic [abliteration](https://aiwiki.ai/wiki/abliteration) (removal of refusal behavior by directional ablation), to the model's text encoder and published the result on September 20, 2026 as Qwen-Image-2.1-Text-Encoder-Heretic, with [GGUF](https://aiwiki.ai/wiki/gguf), FP8 and other quantized variants for ComfyUI and [llama.cpp](https://aiwiki.ai/wiki/llama_cpp) [15][19]. The uploader's card reports that refusals fell from 100 of 100 test prompts for the stock encoder to 5 of 100, with a KL divergence of 0.0220; these are the uploader's own measurements [15]. The card labels the derivative Apache-2.0, stating that the encoder is Qwen3-VL-8B-Instruct used unmodified by Qwen-Image-2.1, and it links refusal-ablated builds of the two prompt-rewriting models and a GGUF build of the diffusion transformer from the same uploader [15]. By contrast, [Unsloth](https://aiwiki.ai/wiki/unsloth)'s GGUF quantizations of the diffusion transformer carry the base model's qwen-research license and must be paired with a separate VAE and Qwen3-VL text encoder [16].

## Early hands-on reports and adoption

GIGAZINE ran the model locally in ComfyUI on a PC with an Intel Arc Pro B70 Creator 32GB graphics card and an AMD Ryzen 5 7600X processor. At 1024 x 1024 pixels and 25 steps, the first image took about 45 seconds and later ones about 30 seconds [23]. The site found that Japanese prompts worked and that asking for a transparent PNG in the prompt produced a fully transparent background. It also reported that the model missed the requested newspaper text on the first try, that image quality was lower than online services such as [ChatGPT](https://aiwiki.ai/wiki/chatgpt) and Gemini, and that a multi-reference editing prompt succeeded only after many attempts [23]. heise wrote that whether the model "is truly that good remains to be seen in tests" [24].

Qwen did not publish hardware requirements. MIXED noted that the README's only provision for smaller graphics cards is a CPU-offloading option [12], while The Decoder wrote that the model runs on "capable consumer GPUs like a 3090" [11].

As of September 25, 2026, Hugging Face reported 37,618 downloads of Qwen/Qwen-Image-2.1 over the previous 30 days and 2,220 likes [25].

## See also

- [Qwen-Image](https://aiwiki.ai/wiki/qwen_image)
- [Qwen-Image-3.0](https://aiwiki.ai/wiki/qwen_image_3)
- [Qwen3-VL](https://aiwiki.ai/wiki/qwen3_vl)
- [Diffusion Transformer (DiT)](https://aiwiki.ai/wiki/diffusion_transformer)
- [AI Image Generation](https://aiwiki.ai/wiki/ai_image_generation)

## References

1. Qwen Team. "Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation." Qwen blog, September 20, 2026. https://qwen.ai/blog?id=qwen-image-2.1
2. "Qwen/Qwen-Image-2.1." Hugging Face model card. https://huggingface.co/Qwen/Qwen-Image-2.1
3. QwenLM. "Qwen-Image-2.1." Official GitHub repository (README), checked September 23, 2026. https://github.com/QwenLM/Qwen-Image-2.1
4. "Qwen Research License Agreement" (release date September 20, 2026). LICENSE file, Qwen/Qwen-Image-2.1 on Hugging Face. https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE
5. Qwen/Qwen-Image-2.1 repository files (model_index.json, transformer/config.json, text_encoder/config.json, vae/config.json and safetensors weight indexes). Hugging Face, checked September 23, 2026. https://huggingface.co/Qwen/Qwen-Image-2.1/tree/main
6. "Qwen/Qwen-Image-Bench." Hugging Face dataset card and leaderboard. https://huggingface.co/datasets/Qwen/Qwen-Image-Bench
7. Qwen Team. "Qwen-Image-Bench: From Generation to Creation in Text-to-Image Evaluation." arXiv:2605.28091, May 2026. https://arxiv.org/abs/2605.28091
8. Qwen Team. "Qwen-Image-2.0: Professional infographics, exquisite photorealism." Qwen blog, February 2026. https://qwen.ai/blog?id=qwen-image-2.0
9. Qwen Team. "Qwen-Image-2.0 Technical Report." arXiv:2605.10730, May 2026. https://arxiv.org/abs/2605.10730
10. Hugging Face Diffusers. "Add Qwen-Image 2.1." Pull request #14804. https://github.com/huggingface/diffusers/pull/14804
11. Bastian, Matthias. "Alibaba's open-weight Qwen-Image-2.1 claims to beat closed models in image generation with just 7 billion parameters." The Decoder, September 20, 2026. https://the-decoder.com/alibabas-open-weight-qwen-image-2-1-claims-to-beat-closed-models-in-image-generation-with-just-7-billion-parameters/
12. Ellison, Shane S. "Qwen-Image-2.1 generates transparent images from open weights, but research use only." MIXED, September 21, 2026. https://mixed-news.com/en/qwen-image-2-1-transparent-rgba-7b-open-weights-research-licence/
13. Jungco, Kezia. "Alibaba's Qwen-Image-2.1 Brings Native Transparency to a 7B Model." eWeek, September 21, 2026. https://www.eweek.com/news/alibaba-qwen-image-2-1-native-transparency/
14. "Qwen-Image-2.1 Leads Open-Source Image Generation: Elevating the Entire Industry to New Heights." 36Kr Europe, September 2026. https://eu.36kr.com/en/p/3991785951198217
15. pottokao. "Qwen-Image-2.1 Text Encoder (Heretic): GGUF, FP8, bf16." Hugging Face model card, created September 20, 2026. https://huggingface.co/pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-GGUF
16. unsloth. "Qwen-Image-2.1-GGUF." Hugging Face repository, created September 21, 2026. https://huggingface.co/unsloth/Qwen-Image-2.1-GGUF
17. Qwen Team. "Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge." Qwen blog, July 21, 2026. https://qwen.ai/blog?id=qwen-image-3.0
18. Hugging Face. Qwen organization model listing, search "Qwen-Image," checked September 23, 2026. https://huggingface.co/models?author=Qwen&search=Qwen-Image
19. p-e-w. "Heretic: Fully automatic censorship removal for language models." GitHub repository. https://github.com/p-e-w/heretic
20. Qwen Developers (@QwenDevs). Post on X about the Qwen-Image-2.1 license and model outputs, September 21, 2026. https://x.com/QwenDevs/status/2101917379785838660
21. Qwen (@Alibaba_Qwen). Reply to @ArchSenex on X, September 21, 2026 (replying to @ArchSenex's question of September 20, 2026). https://x.com/Alibaba_Qwen/status/2101897967703925145
22. Qwen/Qwen-Image-2.1 commit history (LICENSE added September 18, 2026; last commit September 21, 2026). Hugging Face, checked September 25, 2026. https://huggingface.co/Qwen/Qwen-Image-2.1/commits/main
23. "I tried out the image generation and editing AI 'Qwen-Image-2.1,' which has been released as a free open model. It supports Japanese prompts and can also generate transparent PNG images." GIGAZINE, September 24, 2026. https://gigazine.net/gsc_news/en/20260924-qwen-image-2-1/
24. Kirchner, Malte. "Qwen-Image-2.1: Alibaba's new AI image model with transparency feature." heise online, September 22, 2026. https://www.heise.de/en/news/Qwen-Image-2-1-Alibaba-s-new-AI-image-model-with-transparency-feature-11461466.html
25. Hugging Face model API, Qwen/Qwen-Image-2.1 (downloads and likes), checked September 25, 2026. https://huggingface.co/api/models/Qwen/Qwen-Image-2.1
26. "Request: More permissive license for Qwen-Image-2.1." Qwen/Qwen-Image-2.1 community discussion #9, Hugging Face, September 20, 2026. https://huggingface.co/Qwen/Qwen-Image-2.1/discussions/9
27. "[Feedback / Feature Request] Revert Qwen-Image-2.1 to Apache 2.0 to Align with Qwen Family Licensing and Maximize Ecosystem Adoption." QwenLM/Qwen-Image-2.1 GitHub issue #13, September 23, 2026. https://github.com/QwenLM/Qwen-Image-2.1/issues/13