Ant's LLaDA-Image Goes Open Source: 90% Text-Free Pretraining Tops Open-Source T2I Charts
机器之心 · wechat · 2026-09-07
InclusionAI has released and fully open-sourced LLaDA-Image, a unified text-to-image and instruction-editing model built on a from-scratch 6B single-stream DiT paired with the LLaDA2.0-mini diffusion LM for understanding.
Key method: "learn to draw before learning to listen"
- Over 90% of 220M cumulative generation training samples use image-only supervision: pretraining skips captions entirely, with masked-image completion providing both condition and target to build visual priors; image-text pairs are reserved for later SFT language alignment.
- Six-stage pipeline from CoT SFT of the understanding backbone through pure-image pretraining, text alignment, unified T2I/I2I training, and TwinFlow distillation.
- Engineering details: affine-free RMSNorm across the DiT fixes long-run training scale drift; Muon optimizer; TwinFlow shares one DiT backbone for generator and fake-score roles, cutting inference from 50 steps to 2-4.
Results: 53.53/53.38 on Qwen-Image-Bench (EN/ZH), first among open-source models and between GPT-Image 1 and Imagen 4.0 Ultra; strong long-text rendering (0.923/0.913) and editing scores (7.336/7.294 on GEdit-Bench). Counting (GenEval 0.53) remains a weakness.
Weights (incl. FP8 and Turbo), training/inference code, and the full data recipe are open on GitHub, Hugging Face, and ModelScope.
More from Multimodal
- Magnific teases October 28 release: 'We're making it move' — cuenca · 2026-10-27
- Midjourney --sref demo: retro Americana Batman art series — gizakdag · 2026-10-27
- AI video is a new medium, not cinema or animation — and it's perfect for depicting dreams — Kyrannio · 2026-10-27
- Dad turns his toddler's 31-year-old quote into a song via SUNO kids' choir — NathanpmYoung · 2026-10-27
- Band's local-AI music video hits 200k views: 900 generations kept 64, 3-second clips beat face drift — Hwang0500 · 2026-10-27
- Adobe Firefly video editor walkthrough: one-sentence Quick Cut turns 32s into 20s — LudovicCreator · 2026-10-26