core
Qwen-Image Technical Report
Chenfei Wu, Jiahao Li, Jingren Zhou · Alibaba (Qwen) · 2025-08 · arXiv:2508.02324 · code
Why it matters
A 20B MM-DiT whose distinguishing contribution is a data pipeline built for complex text rendering, an area every prior open system was visibly weak at. Uses a vision-language model as the text encoder rather than CLIP or T5, and became the base most 2025-26 open editing and RL work builds on.
What this paper does
read: full textBefore this
Open text-to-image systems could follow prompts reasonably well but broke down on rendered text. The paper states that even strong commercial systems such as GPT Image 1 and Seedream 3.0 struggle with multi-line text, non-alphabetic scripts such as Chinese, localized text insertions, and seamless blending of text with surrounding imagery. Editing systems separately struggled to satisfy visual consistency, keeping untouched regions pixel-faithful, and semantic coherence, preserving global meaning through structural edits, at the same time.
The problem
Build a single foundation model that both renders complex, especially logographic, text reliably and edits images while holding those two conflicting editing objectives together.
The idea
Text rendering quality is treated as primarily a data problem rather than an architecture problem, solved with a progressive curriculum that grows text complexity and resolution together, while editing consistency is solved by feeding the model both a semantic and a pixel-faithful view of the input image.
How it works
Generation happens in a VAE latent space compressed 8x8 spatially with 16 channels, not pixels. Qwen-Image is a Multimodal Diffusion Transformer trained with flow matching, whose loss is the mean squared error between a predicted and ground-truth velocity field mapping a Gaussian noise sample toward the clean image latent, so the network predicts velocity rather than noise or clean data directly. Text conditioning comes from the last hidden layer of Qwen2.5-VL, a vision-language model rather than a CLIP or T5 text encoder, fed into the diffusion transformer through joint attention using Multimodal Scalable RoPE, which assigns text tokens 2D positions so they interleave with image tokens rather than being cross-attended separately. Training follows a five-stage curriculum in which resolution grows from 256x256 through 640x640 to 1328x1328, non-text data is progressively replaced by text-bearing data, data quality is refined, the data distribution is rebalanced, and synthetic data is added last, generated as pure text renders on clean backgrounds, text composited onto physical media like paper or wood, and complex layouts such as slide decks and UI mockups. For editing, the VAE-encoded latent of the input image is concatenated with the noised target latent along the sequence dimension, and MSRoPE is extended with a frame dimension so the model can tell multiple images apart; this reconstructive VAE signal is combined with the semantic features Qwen2.5-VL already extracts, giving the editing module both a pixel-faithful and a meaning-faithful view of the source image.
Evidence
GenEval reaches 0.87 for the base model and 0.91 after RL post-training, which the paper presents as the first system to cross 0.9 on this benchmark. DPG-Bench reaches 88.32. On the text-rendering benchmarks built for this report, ChineseWord scores 58.30 overall, with 97.29% accuracy on the 3,500 most common (Level-1) characters falling to 40.53% on Level-2 and 6.48% on Level-3 rare characters; LongText-Bench reaches 0.943 on English and 0.946 on Chinese, the paper's highest reported score on the Chinese track. For editing, GEdit-Bench overall score is 7.56 on the English track and 7.52 on Chinese, and ImgEdit overall score is 4.27. On novel view synthesis over the GSO dataset the model reaches SSIM 0.884 against 0.859 for FLUX.1 Kontext. The fetched text does not state the CFG or guidance scale used for these numbers.
Limitations
The paper does not carry a dedicated limitations section, but its own numbers concede that rare and complex Chinese characters remain hard, with ChineseWord accuracy falling from 97.29% on common characters to 6.48% on the rarest tier. On the AI Arena human-evaluation leaderboard the model trails the leading Imagen 4 Ultra Preview by roughly 30 Elo points. The authors also exclude Chinese-language prompts from their closed-source arena comparison because most competing commercial APIs do not reliably support Chinese text generation, which limits how directly the Chinese-text claims can be benchmarked against closed competitors.
Why it matters
It shows that the text-rendering gap plaguing open text-to-image models was substantially a data-curriculum problem rather than an irreducible architectural limitation, and it demonstrates a concrete recipe, dual semantic-plus-reconstructive conditioning, for reconciling the visual-consistency and semantic-coherence demands of unified generation-and-editing models. Its open release made it the base most subsequent 2025-26 open editing and RL work builds on.
Abstract, in the authors' own words
We present Qwen-Image, an image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. To address the challenges of complex text rendering, we design a comprehensive data pipeline that includes large-scale data collection, filtering, annotation, synthesis, and balancing. Moreover, we adopt a progressive training strategy that starts with non-text-to-text rendering, evolves from simple to complex textual inputs, and gradually scales up to paragraph-level descriptions. This curriculum learning approach substantially enhances the model's native text rendering capabilities. As a result, Qwen-Image not only performs exceptionally well in alphabetic languages such as English, but also achieves remarkable progress on more challenging logographic languages like Chinese. To enhance image editing consistency, we introduce an improved multi-task training paradigm that incorporates not only traditional text-to-image (T2I) and text-image-to-image (TI2I) tasks but also image-to-image (I2I) reconstruction, effectively aligning the latent representations between Qwen2.5-VL and MMDiT. Furthermore, we separately feed the original image into Qwen2.5-VL and the VAE encoder to obtain semantic and reconstructive representations, respectively. This dual-encoding mechanism enables the editing module to strike a balance between preserving semantic consistency and maintaining visual fidelity. Qwen-Image achieves state-of-the-art performance, demonstrating its strong capabilities in both image generation and editing across multiple benchmarks.
Research line
Reported results
| Benchmark | Value | Guidance | Budget | Source |
|---|---|---|---|---|
| DPG-Bench | 88.32 | — | — | Table 3 |
| GenEval | 0.91 | — | — | Table 4 |
| GenEval | 0.87 | — | — | Table 3 |
Post-RL. Illustrates that RL alignment moves GenEval by ~0.04, comparable to the gap between several distinct systems in this table.
Design-axis choices
Architecture
Method note — the shared flow-matching interpolation
Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.