SenseNova-U1.5: Towards Native Unified Visual Intelligence

简体中文

L1 masked patch targets and expert OPD fusion, not 8B training.

Paper information

Field	Value
Paper	SenseNova-U1.5: Towards Native Unified Visual Intelligence
Affiliations	SenseTime Research, first author's affiliation
First public release	2026-09-10
Authors' code	SenseNova-U1.
Adapter / method	sensenova-u1-5
Local code	src/auto_research/foundation_latest_20260914.py

Original paper

Background and contribution

Use spatial patch reconstruction instead of tokenizer/VAE, and multi-expert OPD to align understanding and generation post-training.

flowchart LR
  I[Image patches] --> R[Masked reconstruction targets]
  R --> U[Unified model]
  E[Expert teachers] --> O[OPD fusion] --> U

Key figure from the paper

Original Figure 3: SenseNova-U1.5. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism

The deterministic reference computes masked patch targets and merges expert OPD signals, with audited state and boundary checks.

Paper results

The authors report combined understanding and generation results on the unified 8B-MoT. No local paper-scale or inferred A/B result.

Local reproduction

Three-seed metrics use matched fixtures, budget, and seeds and verify state transitions only.

Reproduction limits

Masked target and expert OPD only, no multistage 8B training.