LLaDA-Image learns to generate before mapping words to images

InclusionAI releases LLaDA-Image, an open visual generation and editing model trained mostly on images alone, accompanied by a four-step Turbo version.

Does an image generator need to begin its training with images paired with descriptions? InclusionAI argues that it does not. Its LLaDA-Image family builds its visual representation almost entirely from images alone before introducing data that connects text with imagery.

This strategy is the project’s main distinguishing feature. Of the roughly 220 million cumulative samples used during generation training, more than 90% rely on visual supervision without corresponding descriptions. Text-image pairs are introduced later, teaching the system to connect instructions with representations it has already learned.

The aim is to separate two forms of learning that are often handled at the same time. The first concerns the appearance of the world, scene composition, shapes, materials, lighting and relationships between objects. The second connects that knowledge to the words used by the person prompting the model.

In a conventional approach based on paired data, the quality of the result partly depends on the quality of the captions. An incomplete description may omit an object, assign the wrong attribute to a subject or reduce a complex scene to a few generic terms. LLaDA-Image first attempts to learn directly from the visual domain, then uses descriptions to make that representation controllable through language.

“Image-only” does not mean the model learns without any form of conditioning. During the first stages of training, an image is analyzed by a visual understanding module. Some of the resulting information is masked, and the generator must reconstruct the missing elements from the remaining context.

The system therefore receives a form of visual description derived from the image itself. This allows it to learn the structure, content and appearance of a scene without depending on a written or automatically generated caption. The authors describe this stage as the construction of a “visual prior,” which is later connected to language.

According to the technical report, 98% of the training corpus consists of real images. The pretraining and intermediate training stages use only this type of content. Synthetic data appears mainly during supervised stages, particularly to improve the rendering of text inside images.

InclusionAI does not disclose the exact list of sources used. The paper refers to material collected from the web and curated datasets, then filtered according to format, visual quality and aesthetic value. The images are also screened for watermarks, private information and various defects.

The descriptions used during language alignment were generated with Qwen3.6-35B-A3B and Qwen3-VL-235B-A22B-Instruct. They were subsequently filtered to remove invented objects, incorrect relationships, transcription errors, repetition, refusals and malformed responses.

The reliance on automatically produced descriptions does not disappear. It simply occurs later and represents a smaller share of the overall training volume. Paired data remains essential to turn a visual representation into a system that can follow instructions in English or Chinese.

Training also progresses through several definition levels. It begins with a nominal 256 × 256 budget, moves to 512 × 512 with several image formats, then introduces text alignment at 512 × 512 and 1,024 × 1,024. Specialized stages subsequently improve portraits, posters, advertisements and scenes containing text.

The main generator is a 6-billion-parameter Diffusion Transformer, or DiT, trained from scratch. It processes conditioning information and the compressed representation of the image to be produced within the same sequence.

The generator is supported by a frozen vision-language understanding module built from LLaDA 2.0 Mini and a SigLIP-VQ encoder. An adapter called the Residual Query Adapter transforms its representations before sending them to the DiT.

The system’s two main components therefore both rely on diffusion models. LLaDA 2.0 Mini processes language using this family of architectures, while the DiT produces images through progressive denoising. They remain separate components connected by an adapter.

The description of LLaDA-Image as a 6-billion-parameter model also requires some context. That figure refers to the DiT responsible for generation. The Hugging Face page categorizes the complete checkpoint as a 7-billion-parameter system and lists a size of approximately 49.3 GB in BF16 because the distributed package includes other components required to run it.

The same generator supports both text-to-image creation and the modification of an existing image. InclusionAI does not provide a separate model dedicated to editing. The text instruction follows the usual language path, while the reference image takes a more direct route.

The reference is converted into both semantic information and a detailed visual representation. The former helps preserve subject identity and overall structure. The latter retains more local characteristics, including textures, backgrounds and the arrangement of the composition.

The model can therefore replace an object, change an item of clothing, modify a setting, add text or transform a style while preserving the parts not targeted by the request. This shared architecture makes it easier to move between generation and editing, but it does not mean that both tasks perform at the same level.

On GEdit-Bench, which contains 606 real-world cases across 11 categories, LLaDA-Image receives an overall score of 7.336 in English and 7.294 in Chinese. These results place it behind several specialized or proprietary systems, including SenseNova, FireRed-Image-Edit, Qwen-Image-Edit 2511 and Seedream 4.5.

Its semantic consistency score of 8.043 on the English track is higher than its perceived quality score of 7.182. The model therefore appears better at preserving the intent and content of a request than at perfecting the finish of every edit.

Standard image generation is its strongest area. On Qwen-Image-Bench, LLaDA-Image