LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes

Chuyan Chen, Haoxing Chen, Kun Chen, Zhenglin Cheng, Long Cui, Ruishan Fang, Zhangxuan Gu, Zhicheng Huang, Zhenzhong Lan, Yuanting Lei, Haoquan Li, Jianguo Li, Rongchuan Li, Sidu Li, Tao Lin, Deyuan Liu, Jiacheng Liu, Lin Liu, Yuxuan Lou, Zhisheng Lu, Yuxin Ma, Shuheng Shen, Peng Sun, Chaoyang Wang, Hongjun Wang, Xiaomei Wang, Yongxin Wang, Chengzhang Wu, Hongru Wu, Jun Xie

LLaDA-Image is a 6B Diffusion Transformer that decouples visual-prior learning from language alignment to enable efficient, open-source image generation and editing.

How can we build a high-performance image generator by leveraging a frozen, pre-trained vision-language understanding backbone?

Current image generators often couple visual-prior learning with language alignment from the start, forcing reliance on expensive, lossy paired image-text data during the most compute-intensive training stages. LLaDA-Image addresses this by prioritizing image-only pre-training, using a frozen vision-language model to derive conditions from visual regions rather than external captions, and employing a single-stream Diffusion Transformer for unified generation and editing. On Qwen-Image-Bench, the model sets a new state-of-the-art for open-source systems, achieving overall scores of 53.53 and 53.38 on English and Chinese tracks respectively.

Paper Primer

The system architecture hinges on a modular design: a frozen diffusion language model (dLLM) handles multimodal understanding, while a Residual Query Adapter and Transformer Connector project these semantic features into the conditioning space of a single-stream Diffusion Transformer (DiT). This setup allows the model to perform text-to-image generation and reference-preserving editing within one framework, using a dedicated reference-image pathway that bypasses the VLM entirely.

LLaDA-Image establishes a new state-of-the-art for open-source image generators on the Qwen-Image-Bench.

The model outperformed all compared open-source baselines in Quality, Aesthetics, and Alignment dimensions across both English and Chinese tracks. It exceeded the next-best open-source baseline, Z-Image Turbo, by 1.87 points on the English track and 0.67 points on the Chinese track.

To enable efficient deployment, the authors use TwinFlow distillation to produce LLaDA-Image Turbo, which compresses the multi-step generation process into 2–4 sampling steps. The training recipe is real-data-dominant, with 98% of the 220M training samples being real images, and over 90% of the pipeline dedicated to image-only pre-training.

Why does the model use image-only pre-training instead of standard paired image-text data?

The authors observe that paired captions are expensive and lossy, often leading to detail loss at low resolutions. By using image-only data, the model learns a strong visual generative prior at scale without the artifacts or limitations associated with synthetic-heavy caption mixtures.

How does the model handle image editing without requiring a separate task-specific backbone?

The framework reuses the shared DiT backbone but introduces a parallel reference-image stream. This stream injects SigLIP-VQ semantic features and VAE latents directly into the DiT, allowing the model to preserve identity and structure in unedited regions while following new instructions.

Introduction and Motivation

We frame the need for a generation model that leverages a frozen VLM without costly paired data.

Current high‑fidelity diffusion models depend on massive paired image‑text datasets from the start of training. Collecting and curating such captions is expensive, and the captions often lose fine‑grained visual detail after down‑sampling. Moreover, tying visual‑prior learning to language alignment forces the model to allocate compute to learning a language interface before it has a solid image generation foundation, creating a bottleneck that slows convergence and inflates data requirements.

The paper seeks a way to train a strong image generator without relying on large paired caption corpora, by separating visual‑prior learning from language alignment and reusing a frozen vision‑language model for semantic conditioning.

The shift toward leveraging frozen VLM backbones enables efficient, high‑quality image generation without the heavy cost of paired caption data.

Architecture and Components

LLaDA-Image separates semantic understanding from pixel synthesis, linking them with a lightweight connector.

LLaDA-Image splits image generation into two cleanly decoupled stages—semantic understanding via a frozen VLM and pixel‑level synthesis via a Diffusion Transformer—bridged by a lightweight connector that lets both text‑to‑image and editing share the same backbone.

It treats high‑level meaning and low‑level pixel generation as separate modules linked by a thin adapter, so each can be optimized for its own task while still talking to each other.

How does the Residual Query Adapter differ from fine‑tuning the frozen VLM?

Fine‑tuning would update all VLM weights, requiring large compute and risking loss of its pretrained knowledge. The RQA instead adds a tiny set of learnable query tokens that attend to the frozen VLM’s hidden states, extracting the needed generation cues while leaving the VLM parameters untouched.

Append the residual queries to the input sequence: $c' = [\text{"cat"}, 0.3, 0.7]$.

Run the frozen VLM forward pass on $c'$, yielding hidden states $h_{\text{vlm}} = [0.5, 0.2, 0.1]$ (illustrative values).

Project $h_{\text{vlm}}$ through the connector to obtain conditioning tokens $h_{\text{cond}} = [0.45, 0.18]$.

These conditioning tokens are then fed to the DiT together with a noised image latent, completing the generation pipeline.

The RQA injects only a few scalar values that steer the frozen VLM toward generation‑relevant features, avoiding any weight updates while still providing a strong signal to the downstream DiT.

Single‑stream DiT forward pass with optional reference‑image branch.

**Figure 3.** Overview of the LLaDA-Image architecture. Learnable query tokens aggregate information from the input tokens through cross-attention in the RQA. The resulting query representations are concatenated with the original inputs and encoded by LLaDA 2.0 Mini, after which the connector maps the hidden states into the DiT conditioning space. The FLUX.2 VAE supplies image latents, while editing additionally conditions the generator on SigLIP-VQ features and clean VAE latents from the reference image. All modal tokens are concatenated and processed by the single-stream DiT to predict the flow-matching velocity.

**Disclosure.** All results displayed on these two pages are generated outputs, spanning text-to-image synthesis, bilingual text rendering. This informal reader challenge is a qualitative demonstration, not a controlled perceptual study.

Training Pipeline

We detail the training pipeline that equips the frozen backbone with chain‑of‑thought fine‑tuning.

Training a generator that relies on a frozen vision‑language backbone is painful because the backbone never sees the generation distribution. The authors resolve this by first applying Chain‑of‑Thought supervised fine‑tuning to the backbone, then training the generator on top of the adapted model.

CoT SFT teaches the frozen backbone to reason step‑by‑step on multimodal prompts, turning a static recognizer into a dynamic problem‑solver.

Tokenization yields $5{,}000 + 5{,}000 + 5{,}000 = 15{,}000$ text tokens; the image contributes $1{,}384$ VQ tokens, totaling $16{,}384$.

A block‑diagonal mask creates four independent attention blocks, preventing any token from attending to tokens of the other three sentences.

Sampling $r = 0.6$ gives $\rho = \cos(0.6\pi/2) \approx 0.31$, so $31\%$ of answer tokens are replaced by the mask token.

Cross‑entropy is evaluated on the masked $0.31 \times$ answer tokens, driving the model to reconstruct them from the surrounding context.

CoT SFT forces the backbone to practice conditional generation on realistic multimodal inputs, bridging the gap between static understanding and dynamic synthesis.

Apply Chain‑of‑Thought SFT to the frozen VLM backbone using the packed token format.

Pre‑train (PT) the generator on image‑only data at $256^2$ resolution.

Mid‑train (MT) the generator on image‑only data at $512^2$ resolution with AR buckets.

Fine‑tune (SFT) the generator for text‑to‑image at $512^2 arrow 1024^2$ resolution.

Jointly train generation and editing (T2I + I2I) to enable multi‑step manipulation.

Run the multi‑step LLaDA‑Image pipeline to produce high‑fidelity samples.

Distill the final model into a Turbo checkpoint (2–4 inference steps) via TwinFlow.

**Figure 2.** Training and release path. CoT SFT prepares the frozen understanding backbone (dashed); the solid path trains the generator and yields the Base and TwinFlow-distilled Turbo checkpoints. PT and MT denote pre-training and mid-training, respectively.

Generation Performance

LLaDA-Image sets open-source SOTA on Qwen-Image-Bench, surpassing the next best model by 1.87 points.

LLaDA-Image establishes open‑source state‑of‑the‑art on Qwen‑Image‑Bench, beating the next best open‑source model by 1.87 points on the English track.

Table 3 shows LLaDA‑Image achieving an overall score of 53.53 versus Z‑Image Turbo’s 51.32.

A creator‑oriented benchmark that evaluates image generation across five dimensions using 1,000 stratified prompts in English and Chinese.

**Figure 1.** **Qwen-Image-Bench Results.** LLaDA-Image achieves open-source SOTA on both English and Chinese tracks; models are ordered by their two-track average.

LLaDA‑Image outperforms all open‑source competitors on Qwen‑Image‑Bench, delivering higher quality, aesthetics, and alignment without extra prompting tricks.

Text Rendering Capabilities

LLaDA‑Image excels on long‑text and multi‑region benchmarks, achieving top open‑source scores.

LLaDA‑Image scores 0.923 on the English LongText‑Bench and 0.913 on the Chinese subset, the highest among open‑source models.

Table 5 reports these values for LLaDA‑Image.

LLaDA‑Image not only leads the open‑source tier on LongText‑Bench, but also maintains stable accuracy as the number of text regions grows—from 0.892 on two‑region prompts down to 0.857 on five‑region prompts—demonstrating consistent performance under increasing visual‑text complexity.

A benchmark that tests whether a model can faithfully render long strings of text in images, covering both English and Chinese, and exposing errors that become more frequent as the text length increases.

How does LongText‑Bench differ from short‑text or single‑region evaluations?

Short‑text tests involve only a few characters, so omission or duplication errors have little impact. LongText‑Bench uses long strings and multiple regions, making each missing or extra character significantly lower the score, thus revealing weaknesses in sustained textual rendering.

**Table 5.** Comparison of Long-Text Rendering Ability on LongText-Bench.

**Table.** Legacy diagnostic benchmarks.

**Table 4.** Comparison of Text-to-Image Generation Performance on Qwen-Image-Bench (CN).

Extended Benchmarking

Additional experiments report LLaDA‑Image’s top scores on legacy benchmarks.

LLaDA‑Image leads the GenEval benchmark with an overall score of 0.85, surpassing the nearest competitor by 0.02.

Overall 0.85 is the highest among 22 models; the next best model scores 0.83.

LLaDA‑Image Turbo attains 87.48 on the DPG‑Bench, beating the runner‑up by 0.13.

Overall 87.48 is the top score; the second‑place model scores 87.35.

Image Editing Performance

We test how well LLaDA‑Image edits images compared to prior models.

The paper’s core claim is that a frozen vision‑language backbone lets a diffusion model edit images without extra training. Here we evaluate that claim on the bilingual GEdit‑Bench benchmark.

LLaDA‑Image’s semantic consistency exceeds its perceptual quality on the English track.

Table 9 shows GSC = 8.043 versus GPQ = 7.182 for LLaDA‑Image EN.

The remaining perceptual‑quality gap indicates that LLaDA‑Image still lags behind specialized editing models, pointing to a clear target for refinement.

CoT SFT and RL Analysis

Assess how Chain-of-Thought supervised fine‑tuning changes multimodal benchmark performance.

CoT supervised fine‑tuning is the sole stage that modifies the frozen vision‑language backbone, so we isolate its impact before any generation components are added.

**Table 11.** Multimodal understanding results before and after CoT supervised fine-tuning. The LLaDA 2.0 Uni column reproduces the scores reported in its technical report (Inclusion AI et al., 2026); the CoT SFT column is evaluated by us under the protocol described in App. B.

Aspect-Ratio Bucket Configurations

Defines aspect‑ratio buckets and resizing rules for consistent training across resolutions.

During image‑only mid‑training we introduce a set of aspect‑ratio buckets and keep them fixed through supervised fine‑tuning. The two target budgets are called “512²‑pixel” and “1024²‑pixel”, referring to the approximate total pixel count rather than a square resolution.

Because the selected bucket shares almost the same aspect ratio as the source image, the excess after scaling is only a few pixels, so cropping removes negligible content. In practice this step is a down‑sampling operation, since training images are typically larger than the bucket dimensions.

**Table 10.** Aspect-ratio bucket configurations at the $512^2$-pixel and $1024^2$-pixel budgets. Each tuple reports $(W, H)$ in pixels.

Equal‑area buckets prevent any single rank from becoming a memory or compute outlier, which would otherwise cause training‑time stragglers. Compared with forcing every sample into a square, this scheme avoids geometric distortion and aggressive cropping while teaching the model to generate a wide spectrum of output aspect ratios.

Qualitative Examples

Appendix C showcases diverse visual examples spanning artifacts to landscapes.

The Han‑dynasty ceramic maid stands alone on a museum display, hands folded before her, robe hanging straight down. Its surface shows faint red, black, and painted pigments, warm‑gray clay, fine cracks, and localized loss, all highlighted by soft left‑side lighting against a warm‑gray wall.

A Ming‑era cloisonné double‑ear vase (1368‑1644) occupies the center of a light‑gray backdrop. The vessel’s four‑petal “sea‑peony” silhouette is gilded at the rim, with symmetrical gilt dragon ears, a raised lotus‑petal band, and a body covered in intricate blue‑green enamel dragons outlined in gold.

A male traveler in his thirties poses on a wooden observation platform overlooking a glacier. He wears a gray sport‑sunglass, orange‑brown jacket, black tee, gray trousers, and a small backpack, while the glacier below shows pale blue‑white ice, fissures, and gray‑rock debris.

A realistic early‑morning photograph captures a Mediterranean hillside town built of yellow, cream, and warm‑orange stone houses climbing the slope. Red‑brown tiled roofs, varied windows, and occasional green‑plant pots line narrow, winding stone streets that lead up a series of steps.

The scene depicts a Jiangnan water‑town centered on a winding canal flanked by white‑walled, gray‑tiled dwellings. A low stone arch bridge spans the water, a narrow boat glides beneath, and nearby shops display simple signs, one reading “Tea House.”

A wide‑angle aerial view shows a rugged green grass slope eroded by sea wind, leading up to a towering sea‑eroded cliff that drops vertically into deep blue water. The cliff’s gray‑brown face is punctuated by visible strata and wave‑carved grooves, while distant headlands fade into a blue‑gray sky.

Captioning and Filtering

Describes the captioning pipeline, prompts, and filtering criteria for generating high‑quality image‑text pairs.

The captioner produces a comprehensive description of each image, enumerating main subjects, attributes, actions, spatial relations, scene context, and visual style. After generation, structural checks discard any caption that is incomplete, a refusal response, or a degenerate repetition.

Qwen3.6-35B-A3B then evaluates both image quality and caption faithfulness using the image, its caption, and the transcription of any visible text. Samples containing watermark signals or private information are filtered out, and any caption that hallucinates visual content or misdescribes visible text is rejected.

The Chinese captioning prompt instructs annotators to generate high‑density, faithful Chinese descriptions, covering subjects, scene, composition, style, lighting, and any visible text, while avoiding aesthetic commentary. The short prompt variant must be a concise 10‑50 word user‑style instruction without punctuation.

The English captioning prompt follows the same structure: produce an objective, information‑dense English description that lists subjects, scene, composition, style, lighting, and visible text verbatim. The short prompt is a 5‑30 word user‑style command without a trailing period.

The unified filtering prompt defines four boolean or enumerated fields for quality control: detection of private information, detection of watermarks, assessment of render‑text usability, and evaluation of caption faithfulness (faithful, minor issue, or hallucinated).

Prompt Examples Part 1

Exact prompts used for the report’s text‑to‑image examples.

Prompt #1 (city‑park portrait): a candid street‑style portrait of a young East‑Asian woman walking along a rain‑wet park path, half‑frame height, turning back at a large tree, detailed clothing, lighting, and background as described in the original Chinese prompt.

Prompt #2 (indoor portrait with “LLaDA‑Image is Released!”): a realistic phone‑camera style indoor portrait of a young white woman beside a huge picture frame that displays the bold text “LLaDA‑Image is Released!”, with specific clothing, lighting, and a glass vase of dried roses.

Prompt #3 (Xiaohongshu‑style cafe portrait): a lightly vintage street portrait of a young East‑Asian woman leaning against an old café wall, holding an instant camera and a paper bag, with a blackboard reading “今日特调” and “冰美式”.

Prompt #4 (early‑morning market documentary): a vertical street‑documentary shot of a middle‑aged male vendor arranging vegetables at a market, with fresh produce, a straw‑hat customer, and handwritten price signs.

Prompt #5 (evening seaside boardwalk group shot): three young East‑Asian people posing on a boardwalk railing at dusk, wind lifting hair and clothing, with a gray‑blue sea and lighthouse in the background.

Prompt #6 (restaurant birthday celebration): a warm indoor medium shot of a young East‑Asian couple holding a pink birthday cake, surrounded by pasta, wine, and soft lighting.

Prompt #7 (library children’s portrait): a seven‑year‑old East‑Asian girl reading a picture book at a low table by a window, wearing a pink cardigan and gray skirt, with crayons and a book titled “森林朋友”.

Prompt #8 (home‑scene child building blocks): a four‑year‑old girl sitting cross‑legged on a carpet, assembling red blocks, surrounded by toys and natural afternoon light.

Prompt #9 (winter‑evening amusement‑park old‑photo style): a five‑year‑old boy beside a carousel entrance, making a V‑sign, dressed in a down jacket and hat, with a dated warm‑yellow film look.

Prompt #10 (travertine‑lake landscape): a realistic natural‑landscape view of snow‑fed travertine pools with colorful mineral‑rich water, surrounding pine forest, and distant mountains.

The image displays two side-by-side comparisons of a statue, labeled "Boogu-Image 0.1 Turbo" and "Gemini 3.1 Flash Image".

The image displays two separate photographs of ancient Chinese-style ceramic figurines, labeled "Z-Image Turbo" (top) and "GPT-Image 2" (bottom).

A Han Dynasty pottery female attendant figurine stands independently on a museum display platform. The figure has her hands folded in front of her body, with a long robe hanging straight down. The styling is simple and restrained, and the facial features are sculpted in an archaic style. The surface retains small amounts of faded vermilion, black, and painted pigments, and the pottery is a warm grayish-brown color. Fine cracks, soil deposits, and local damage are visible. Soft, natural light falls from the left side, and the background is a warm gray-beige wall. The image highlights the figure's silhouette and historical traces, presented as a realistic archaeological photograph with a vertical composition.

The figure displays two images of ancient Chinese-style ceramic figurines, labeled "LLaDA-Image Turbo" (top) and "Qwen-Image 2512" (bottom).

**Figure.** GPT-Image 2 and Z-Image Turbo.

**Figure.** Gemini 3.1 Flash Image and Boogu-Image 0.1 Turbo

Figure: A terracotta figurine of a standing woman, likely from the Tang Dynasty, displayed on a pedestal.

The image displays two side-by-side or stacked visual outputs labeled "Boogu-Image 0.1 Turbo" and "Gemini 3.1 Flash Image," both depicting ornate, turquoise-colored cloisonné vases with gold dragon-shaped handles.

The image displays two side-by-side or stacked visual outputs from different models, labeled "LLaDA-Image Turbo" and "Qwen-Image 2512", both depicting an ornate cloisonné vase with dragon handles.

**Figure.** Comparison of vase images generated by Z-Image Turbo and GPT-Image 2.

An ornate, bulbous vase featuring intricate cloisonné enamel work in shades of turquoise, blue, and reddish-brown, accented with gold-colored metal filigree. The vase has a flared, petal-shaped rim and two prominent, sculpted golden dragon handles on either side of the neck. The body of the vase is decorated with swirling, cloud-like patterns and is set against a plain, neutral gray background.

**LLaDA-Image Turbo**

**Z-Image Turbo**

The image displays two panels, each showing a person standing on a wooden viewing platform with a glacier and mountains in the background. The top panel is labeled "Boogu-Image 0.1 Turbo" and shows a close-up portrait of the person with arms raised. The bottom panel is labeled "Gemini 3.1 Flash Image" and shows a wider shot of the same person in the same pose on the platform, including more of the surrounding environment and other people.

一张冰川观景区的男性旅行游客照。拍摄于晴朗上午。主体是一位三十岁左右的东亚男性，戴着深灰色运动墨镜，站在木质观景平台中央，双手带着手套高高张开，头微微抬起，脸上露出自然放松的笑容。人物脸型偏方长，额头较宽，黑色短发被风吹得自然蓬松。鼻梁较直，鼻尖略宽，嘴唇厚度适中。皮肤呈自然偏暖的肤色，面颊有轻微日晒痕迹，整体保持偏哑光质感。他穿橙棕色户外夹克、黑色T恤、深灰色户外长裤和登山鞋。肩上背着一个小型登山包。背景是一条从巨大雪山之间缓慢向山谷延伸的冰川，冰川表面呈浅蓝白色，可以看到自然裂隙和灰色碎石带，两侧山体为黑灰色岩石与积雪交错。天空蓝得清透，几朵白云位于主峰附近。平台周围还有护栏和少量游客，整体规模感明确，同时保持普通游客旅行纪念照的自然感。

**GPT-Image 2.**

**Figure 2512.** Qwen-Image 2512

**Figure.** A person standing on a wooden viewing platform with arms outstretched, wearing a brown jacket, black shirt, grey pants, and sunglasses. In the background, there is a large glacier surrounded by snow-capped mountains under a clear blue sky. Other people are visible in the distance on the platform.

A realistic photograph of a European hillside town at dawn, featuring Mediterranean-style stone houses in shades of pale yellow, cream, and warm orange with reddish-brown tiled roofs. The town is built into the slope, with narrow stone alleys and staircases connecting different levels. In the foreground, there are terraces with potted plants, while the background shows blue-gray mountains and a calm body of water. The morning light illuminates the building facades, creating soft shadows and a clear, natural atmosphere.

The image displays two side-by-side or stacked visual outputs labeled "LLaDA-Image Turbo" and "Qwen-Image 2512".

This image shows a hillside village overlooking a body of water.

**Figure 2.** GPT-Image 2

This image depicts a scenic view of Positano, Italy, characterized by its iconic colorful buildings cascading down a steep hillside toward the Mediterranean Sea. In the foreground, a narrow stone staircase winds between traditional Mediterranean-style buildings with balconies adorned with flowers. The background features a dramatic mountain slope covered in lush vegetation, overlooking the deep blue sea. The perspective is from an elevated position, looking down the path toward the town center and the coastline.

The image displays two panels, each showing a canal scene in a traditional Chinese water town. The top panel is labeled "Boogu-Image 0.1 Turbo" and the bottom panel is labeled "Gemini 3.1 Flash Image".

**Qwen-Image 2512**

**SenseNova U1.5 Preview**

The image contains a descriptive text block in Chinese, titled "SenseNova U1.5 Preview". The text provides a detailed prompt or description of a realistic coastal cliff landscape, specifying elements such as the high-angle perspective, deep green grassy foreground with exposed rocks, narrow paths, massive sea cliffs, crashing waves, and a vast sky with clouds.

**Qwen-Image 2512.**

**Figure.** SenseNova U1.5 Preview

This image shows a coastal landscape with steep, grassy cliffs dropping into the ocean.

Prompt Examples Part 2

Additional image‑generation prompts and their translations.

A golden snub‑nosed monkey perches on a thick branch in the Qinling Mountains, its golden‑orange fur and pale blue facial skin rendered in high‑resolution, telephoto‑lens detail under soft morning light.

An adult giant panda sits quietly on a damp Sichuan bamboo‑forest floor, clutching fresh bamboo and displaying lifelike black‑and‑white fur texture, illuminated by even overcast natural light.

Read the original paper

Open the simplified reader on Paperglide

Browse all simplified papers