What SenseNova-U1.5 actually shipped

OpenSenseNova released SenseNova-U1.5-8B-MoT (Preview) on July 31, 2026. On August 20 the same repo posted the non-preview checkpoint, sensenova/SenseNova-U1.5-8B-MoT, plus an 8-step LoRA. Both sit under Apache 2.0. Weights, inference code, and the U1.5 pre-training launcher are public.

The job of the model is one native stack for text-to-image, instruction editing, multi-image reference, and region control. It is not a text-to-image model with a separate inpainter bolted on.

Hugging Face lists the checkpoint at 18B parameters. The name 8B-MoT is the Mixture-of-Transformers layout from the U1 paper: about 8.2B understanding parameters and 8.2B generation parameters, not an 8-billion-parameter monolith.

What the Preview claimed

The Preview notes list native 4K generation, stronger texture, material, lighting, and realism, better Chinese and English text, denser layouts, and editing that keeps subject identity and unedited regions. Region control is masks, bounding boxes, and visual markers.

The architectural change they name is ConvDecoder patch-joint reconstruction. U1 decoded each RGB patch with an MLP head. At high resolution that can show as a grid, a seam, or a texture break. U1.5 reshapes visual tokens into a 2D grid, upsamples with Pixel Shuffle, and lets neighboring patches mix through 3×3 convolutions.

Those sentences are OpenSenseNova’s. They are not an independent lab.

The published generation table

Qwen-Image Bench Overall, from the Preview notes. Higher is better. The dagger row uses Prompt Enhance, a Cosmos3-style upsampler, not the raw checkpoint.

The asterisk on Qwen Image 2 is theirs: they ran that column. Raw U1.5 Preview beats U1. It does not beat HunyuanImage 3.0 or Qwen Image 2 on this overall row.

The published editing table

WeEdit Average, same document:

U1.5 Preview is a step up from U1 and a hair above GPT-Image-1.5 on that average. Nano-Banana-Pro is in another band. ImgEdit-Bench Overall on the same page is 4.37 for U1.5 Preview against 3.90 for U1.

How you run it

The upstream environment is Python 3.11, PyTorch 2.8, and CUDA 12.8. Clone the GitHub repo, uv sync, then point --model_path at the current checkpoint:

python examples/t2i/inference.py \
  --model_path sensenova/SenseNova-U1.5-8B-MoT \
  --prompt "A cinematic mountain lake at sunrise, realistic photography." \
  --width 2048 --height 2048 \
  --device_map auto \
  --output output.png

Editing uses examples/editing/inference.py. The 8-step LoRA is for the August 20 base checkpoint, not the July Preview. A community Q8 GGUF of the Preview is 19.9 GB. There is also a free browser demo at SenseNova-Studio.

What still breaks

The August 20 card says the official release improves on Preview. It still lists oversaturated detail (try lowering cfg_scale), errors in dense or mixed Chinese–English type, imperfect counts and alignment, unstable small faces and hands, and drift on broad or multi-turn edits.

That is the honest product: an open unified image model you can host, not a Photoshop replacement, and not the top row of its own generation table without Prompt Enhance.

Source cutoff: OpenSenseNova GitHub README and U1.5 Preview notes, Hugging Face Preview and U1.5-8B-MoT cards, and the U1 paper, as retrieved August 22, 2026.