SenseTime releases an 8B model for visual understanding and image editing

A smaller unified visual model

SenseTime's SenseNova U1.5-Lite-Preview brings image generation and editing into an 8B-MoT model that also handles visual understanding. The company's own announcement is dated August 3, preceding the August 21 news brief. It describes an iteration of the NEO-Unify approach introduced with U1, rather than a model first unveiled on the later reporting date.

The central feature is a shared visual system that can interpret what is in an image and then change it. SenseTime emphasizes native 4K generation, more accurate Chinese and English lettering, and tighter control over complex layouts.

Editing a region while keeping the rest

The announcement describes an encoder-free visual design and a revised generation head intended to reduce grid-like artifacts. It also says the editing data and training tasks were reorganized to better preserve a subject's identity, scene structure and content outside the requested edit.

Supported demonstrations include reference-image combinations and edits directed by boxes, coordinates or markers. Those controls make the intended region more explicit. For a poster, the user might want to alter a heading while keeping the photograph and layout; for a product image, the requested change might concern a background rather than the object itself.

This is a more demanding objective than generating a fresh image with roughly the same theme. A successful edit must satisfy the new instruction and retain the earlier decisions that were not meant to change. SenseTime's examples illustrate that goal, while remaining company-selected demonstrations rather than a measurement of every possible editing task.

Weights and hosted access are separate

SenseTime links the preview checkpoint on Hugging Face and ModelScope. It subsequently announced U1.5 Lite through the SenseNova Token Plan, giving users another route that does not require running the weights locally. The hosted model identifier is sensenova-u1.5-lite.

The hosted guide describes separate image-generation and image-editing endpoints compatible with the OpenAI image API specification. It supports reference images, optional prompt expansion and output dimensions up to 4096 by 4096. Returned image URLs expire after 24 hours, with an encoded-data response offered as another format.

Request settings that affect the finished asset

The hosted guide requires output dimensions to be multiples of 32, with each side from 512 to 4096 pixels and aspect ratios no wider or taller than three to one. Reference images can use publicly reachable URLs or Base64 data URLs; bare encoded strings without the data-URL prefix are not supported. These are practical interface constraints for an application that prepares source assets automatically.

Prompt expansion is enabled by default through the prompt_extend setting. That means a controlled comparison needs to record whether the service rewrote the request before generation. The guide also exposes a watermark switch and an encoded-data response option. The latter changes delivery format, while the former changes a visible output property; neither changes the underlying distinction between hosted inference and running the preview weights on one's own hardware.

What the API does inside an agent

SenseTime explicitly distinguishes this image service from a conversational chat model. An agent integrates it as an image tool, while another component manages the conversation and decides when to request a generation or edit. This explains why adding its name to an ordinary chat-model selector is not sufficient.

The Token Plan announcement advertised a free public-beta allowance. That is a time-bounded launch offer, not evidence of permanently free production use. The lasting product distinction is the availability of both a preview checkpoint for local experimentation and a documented hosted route for visual work.