SenseNova-U1.5 Technical Report: VAE-Free Native 4K Image Generation
Secret_Yak2496 · reddit · 2026-09-16
SenseNova released the technical report for SenseNova-U1.5, an 8B native unified model for image understanding, generation and editing that skips external visual encoders and VAEs — images map directly to visual tokens, each covering a 32×32-pixel region.
Key methods: spatially joint reconstruction via Pixel Shuffle + 3×3 convs (fixing U1's patch-boundary artifacts from independent token decoding), resolution-aware noise conditioning extended to 4096×4096, and a 'specialize, then unify' scheme — four RL experts (aesthetics, bilingual text rendering, infographics, editing) consolidated via multi-expert on-policy distillation. Training added 59M text-image pairs from 78 sources and 38M editing examples, with 88.2% of effective generation volume above 1024². Known issues remain with dense small text, small faces/hands and multi-reference drift. Open-sourced on GitHub and Hugging Face.
Related event: SenseTime Open-Sources SenseNova U1.5, an 8B Native Unified Model(2 posts)→
More from Multimodal
- Midjourney wallpaper recipe: sref + profile mix for eerie film-blur portraits — michaelrabone · 2026-11-02
- Cavallaero Teaser: A Painterly Rider Races Through a Golden Forest — CommunicationTop4244 · 2026-11-02
- Anima's snapref Offers a New Reference-Based Fix for Character Consistency — Intrepid_Swim_8669 · 2026-11-02
- Pencil-textured folding fan: surreal product concept render goes viral — Delahuntagram · 2026-11-02
- Free Midjourney SREF gallery adds 1,200+ handpicked style codes overnight — tisch_eins · 2026-11-02
- Adobe's EditAgent beats prior image-editing agents but needs ~10x the compute — emmanuelvivier · 2026-11-02