SenseNova-U1.5-8B-MoT Released: Unified Text-to-Image and Understanding Model Initialized from Qwen3

SenseNova has released SenseNova-U1.5-8B-MoT, a unified model for text-to-image generation, editing, and image understanding. The model is initialized from Qwen3 LLM weights and operates without a separate vision encoder or VAE, processing images as tokens in pixel space.

Architecture and Performance

The model uses a single backbone with shared attention but separate projections, norms, and feed-forward layers for different tasks. In benchmarks against Qwen-Image 20B (without prompt enhancement), U1.5-8B shows mixed results:

• Wins: Outperforms Qwen-Image 20B on Qwen-Image-Bench (English and Chinese) and IGenBench (infographics).
• Losses: Trails Qwen-Image 20B on DPG-Bench, OneIG-ZH, and diversity metrics.
• Understanding: Beats Qwen3-VL-8B on 14 of 19 benchmarks in thinking mode.

Availability and Limitations

The model is released under the Apache-2.0 license and has native support in ComfyUI v0.35.0. A community Q8 GGUF version is available (21.2 GB), though no official GGUF exists yet. Known limitations include instability with small faces, hands, and limbs, as well as errors in dense or mixed Chinese-English text. The technical report does not provide VRAM or speed metrics.