LLaDA-Image is a 6B Diffusion Transformer (DiT) trained mostly on unlabeled real images, using a frozen Diffusion Language Model (dLLM)-based VLM to self-supply the conditioning signal instead of captions, so 98% of its 220M training samples are real photos and over 90% carry no paired text.
If you’ve tried to reproduce a modern text-to-image model, you know the pain: the compute-heaviest stages want hundreds of millions of clean image–caption pairs, and either you pay to caption them or you paper over the gap with synthetic images that leak stylistic artifacts into your final model. The dominant recipe from Stable Diffusion 3 and its descendants (Flow matching over paired data from step one) makes captions a hard prerequisite for learning the visual prior. LLaDA-Image argues that most of that captioning was unnecessary. It builds the visual prior from raw images, then bolts language on afterward, and reports the top open-source score on Qwen-Image-Bench on both English and Chinese tracks.
The trick that makes image-only pre-training work: a frozen vision-language model looks at the same image the generator is trying to reconstruct, extracts semantic features, and hands those features back as the “caption.” There’s no external text; the image describes itself. To prevent the generator from just copying, they randomly mask most of the image tokens before the VLM sees them, turning the task into sparse-to-dense prediction.
The architecture has three pieces: (1) a frozen LLaDA 2.0 Mini VLM for understanding, (2) a small connector plus a Residual Query Adapter that pulls generation-relevant signals out of the frozen VLM, and (3) a single-stream DiT that does the actual denoising via flow matching. For editing, the reference image bypasses the VLM entirely and enters the DiT through two parallel channels: SigLIP-VQ features for semantics, plus the raw FLUX.2-VAE latent concatenated with the noised target for pixel-level fidelity. Training goes through five stages of increasing resolution and supervision:
# Simplified training loop for one image-only step
img = load_real_image() # no caption needed
crop = random_crop(img, target=256) # mild downsample
tokens = siglip_vq(crop) # visual features
masked = random_mask(tokens, keep=0.25)
cond = connector(frozen_vlm(concat(aux_prompt, masked, rqa(masked))))
x, z, t = clean_latent(crop), gaussian_noise(), uniform(0,1)
x_t = (1-t)*x + t*z
loss = mse(dit(x_t, t, cond), z - x) # flow matching
After the base model converges, they distill it to 2–4 sampling steps using TwinFlow, which reuses one shared DiT backbone (with two output heads) to play both the generator and the fake-score estimator that DMD2 (Distribution Matching Distillation 2) normally requires as a separate network.
The prevailing recipe treats paired image–text data as the foundation of a text-to-image model, poured in from the earliest and most compute-heavy stage. This paper shows the opposite. Learn the visual prior from raw images by having a frozen VLM caption each image to itself through its own hidden states; introduce real captions only late, when the pixel-level model is already competent. The evidence isn’t the leaderboard score. It’s that a real-image-dominant recipe (98% real, 70%+ real even in SFT) reaches open-source SOTA on a creator-oriented benchmark without the synthetic-caption crutch most competitors rely on.
The load-bearing finding is that the recipe works at all: with paired supervision deferred and the pixel budget kept moderate (220M total generation samples, over 90% image-only), LLaDA-Image reaches 53.53 on Qwen-Image-Bench English and 53.38 on Chinese, ahead of the next open-source baseline Z-Image Turbo by 1.87 and 0.67 points. Secondary results:
•
Bilingual text rendering is balanced but not best-in-class: 0.923 English / 0.913 Chinese on LongText-Bench, behind dedicated text-rendering models like Qwen-Image 2512.
•
Editing works from a single unified checkpoint: 7.336 / 7.294 overall on GEdit-Bench English/Chinese, with semantic consistency (8.043) noticeably ahead of perceptual quality (7.182) on the English track. Specialized editing models still win overall.
•
On GenEval the model tops single-object, two-object, and attribute binding, but scores only 0.53 on Counting, dragging the overall to 0.85. The authors flag GenEval as a saturated diagnostic rather than primary evidence.
•
TwinFlow distillation produces a 2–4-step Turbo variant that stays within ~2–3 points of the full model on most benchmarks.
Reach for this if you’re planning to train a domain-specific image generator (fashion, product, medical, satellite) and you have a large stockpile of unlabeled real images but a limited captioning budget. The recipe says: don’t caption the pre-training set at all. Feed raw images through a frozen VLM, mask most of the visual tokens, and let the model learn to reconstruct from the visible patches. Only start paying for captions when you’re ready to align to actual user prompts at the SFT stage.
They release the full stack under inclusionAI/LLaDA-Image: four checkpoints (Base, Turbo, plus FP8 variants of each) on HuggingFace, training and inference code, and stage-by-stage recipes covering data filtering thresholds (ArtiMuse $\geq$ 60, DeQA-Score $\geq$ 4.0), aspect-ratio bucket configurations, and the TwinFlow schedule. The unified checkpoint handles both text-to-image and reference-preserving editing, so you don’t need separate model variants for the two tasks.
A frozen VLM can caption images to itself well enough that you don’t need real captions to learn a visual prior. This decouples the two hardest problems in training a generator (learning pixels vs. learning language) so you can spend real captioning money only where it moves the needle, which is late-stage alignment, not early-stage prior learning.
•
The self-conditioning trick depends on a strong pre-trained VLM being available in your domain. If you’re generating something the frozen VLM can’t parse well (medical imaging, niche art styles), the “caption” it hands back to the DiT will be weak, and the whole pre-training stage degrades.
•
The paper is a systems report, not a controlled ablation. There is no head-to-head comparison of “same 6B DiT, same 220M images, with vs. without image-only pre-training.” The claim that image-only pre-training is what enables the results is plausible but not isolated experimentally.
•
Editing perceptual quality trails dedicated editing models by roughly a full point on GEdit-Bench, and text rendering trails Qwen-Image on LongText-Bench. The unified-checkpoint approach is a real trade-off, not a free lunch.