Skip to content

[Feature] SenseNova-U1.5-8B-MoT #1841

Description

@GreenShadows

Feature Summary

SenseNova-U1.5 is a natively unified multimodal model built on NEO-unify, featuring new patch encoding and decoding layers.

Detailed Description

Image Image Image

"SenseNova-U1's original MLP head reconstructs each RGB patch independently. At high resolutions, this can expose token boundaries as grid patterns, seams, or texture discontinuities. SenseNova-U1.5 instead adopts progressive spatial reconstruction via ConvDecoder: visual tokens are reshaped into a two-dimensional feature grid and upsampled through multiple Pixel Shuffle stages, with intermediate 3×3 convolutions enabling neighboring patches to interact and merge into a more continuous image.

Qualitative results show that this design substantially suppresses grid-like artifacts throughout pre-training. It also improves robustness during downstream adaptation, reducing the likelihood that similar artifacts re-emerge when fine-tuning on new domains."

HF: https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT-Preview

GitHub: https://github.com/OpenSenseNova/SenseNova-U1

Alternatives you considered

No response

Additional context

No response

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions