LLaDA-Image: 6B diffusion model with fully open training recipes for photorealistic images
iScienceLuvr · x · 2026-09-04
LLaDA-Image pairs a 6B Diffusion Transformer trained from scratch with a frozen vision-language module built on the LLaDA2.0-Mini diffusion LM backbone. Instead of leaning on paired image-text data, it builds a strong visual prior via image-only pre-training and mid-training across 220M samples (98% real images), using parameter-free RMSNorm and the Muon optimizer throughout. The model produces photorealistic images and follows fine-grained editing instructions, and is distilled into LLaDA-Image-Turbo for 2-4 step fast inference. Training recipes are fully open.
More from Multimodal
- HeyGen's Professional Voice Clone needs just 30 min of audio to sound exactly like you — rohanpaul_ai · 2026-10-21
- Flux 3 Max ranks #1 on physics with 1.5M+ human judgments — isidentical · 2026-10-21
- Claude Code directed a full music video: 6.5 hours, 765M input tokens — colinmcnamara · 2026-10-21
- Kling 4.0 debuts native 30-second AI video with 10 keyframes, extendable to 2 minutes — lmoroney · 2026-10-21
- Mistral's Voxtral transcribes an 1891 Eiffel phonograph recording at 6.5% WER — IsForAt · 2026-10-21
- Redditor remakes SD clown dance video at 25 steps with motion boost and fixed prompts — roychodraws · 2026-10-21