sensenova / U1-8B-Interleave

Overview

u1-8b-interleave is an interleaved image-text generator from SenseNova. It uses SenseNova U1’s NEO-unify design, so one model handles both words and pixels. It can produce a single response that mixes a written narrative with multiple generated images. This helps you turn long instructions into visuals without switching tools.

What you can build

• Illustrated how-to guides with one image per step
• Travel diaries that alternate short paragraphs and scenes
• Product explainers with a consistent visual style across panels
• Poster-like layouts with dense text, icons, and structured sections
• Reference-guided scenes that follow the composition of an input image

Inputs

• One to five reference images that guide style or layout. Provide JPG, PNG, or WebP files.
• A written instruction describing the story, steps, or layout you want. If you ask multiple questions, separate them with semicolons.
• A target output resolution selection, from HD up to 4K.
• An aspect ratio choice such as 1:1, 16:9, 9:16, 3:2, 2:3, 1:2, or 2:1.
• A quality setting that controls how many denoising steps to run. Valid values are 1 to 60.
• A guidance strength setting that controls how strictly images follow the instruction. Valid values are 0 to 10.
• An optional numeric seed (0 to 9,999,999) to make results repeatable.
• An optional “thinking” mode that lets the model plan before it commits to final images.

Outputs

• One or more PNG images.
• A generated text narrative that matches the image sequence.
• When the run produces multiple images, they are intended to be read in order, as panels or steps.

Recommended settings

• General quality: use 30 to 50 steps for strong detail.
• When prompt adherence matters most: raise guidance to about 4.
• When you want more variation: lower guidance toward 1 to 3.
• Faster drafts with distilled checkpoints: use 8 steps with guidance around 1.

Limitations

• The upstream model reports a maximum context length of 32K tokens. Very long prompts can still hit limits.
• The model is trained around ~2K resolution buckets. Outputs at untrained shapes can lose quality.
• Interleaved generation is an experimental feature. It may lag behind dedicated single-image pipelines.
• Small or complex human figures can show anatomical artifacts.
• Text inside images can contain misspellings, warped glyphs, or inconsistent formatting.
• Low-quality inputs can hurt results. Blurry photos, heavy compression, or scanned images often reduce fidelity.

Safety & compliance

• The upstream SenseNova U1 weights are released under the Apache 2.0 license.
• Only upload images you have rights to use. This includes faces, logos, and copyrighted artwork.
• Review outputs before publishing. Generated images can include incorrect text or misleading visuals.