JUICE: one flow-matching Transformer unifies image and text generation, hitting 0.88 GenEval
pratyusha_PS · x · 2026-10-23
Most unified models pair an autoregressive LLM for text with a diffusion head for images. JUICE instead trains a single flow-matching Transformer from scratch in the continuous spaces of frozen pretrained encoders (SigLIP2/DINO + Flan-T5), with no LLM initialization, handling generation, understanding, and joint generation with one set of weights.
Key results (2.1B models):
- SigLIP2 variant: 0.87 GenEval, 85.27 DPG-Bench, 0.45 WISE, 73.1 VQAv2; DINO variant: 0.88 GenEval, tying TUNA, the strongest 1–3B baseline
- Parallel text decoder is 3.9–6.4× faster than AR on VQA and 54–163× on captions
- Each modality has its own noise level/schedule, making text-to-image, captioning, VQA, and joint generation paths through one shared noise square
Paper, code, and weights are released.
More from Multimodal
- Designer uses Seedance 2.5 on Runway to design his SF apartment from an iPhone video — c_valenzuelab · 2026-10-23
- MOSS Image Turbo cuts image generation from 50 steps to 4 — _akhaliq · 2026-10-23
- Midjourney horned-woman portrait prompt with full parameter set shared — OVolosin82152 · 2026-10-23
- Dev rebuilds Higgsfield-style one-click video studio on ComfyUI with open weights — Key_Cap_8111 · 2026-10-23
- Creator uses Claude-made previz as reference to shoot subway phone trick with Seedance 2.5 — TinfoilTricorn · 2026-10-23
- Canva launches Craft, its first in-house image generation model — The Decoder · 2026-10-23