SenseNova-U1.5-8B-MoT-TailSteer

SenseNova-U1.5-8B-MoT-TailSteer

An image generation and editing model that stays steadier across many successive edits, built on SenseNova-U1.5-8B-MoT with the same loading steps. For designers and researchers who edit one image repeatedly.

Made by sensenova on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a version of SenseNova-U1.5-8B-MoT with an extra final training stage called TailSteer. It aims to cut the drift that builds up when each edited image becomes the next input. The complete weights keep the same architecture, layout and loading steps as the base model.

Who it suits: People who edit one image through many rounds, and researchers who measure how much an edit pipeline changes the parts it should leave alone.

Industry | Who uses it | How they use it | Example
Creative production | Teams refining images step by step | Edit one image repeatedly | Add objects to a scene one at a time
Visual design | Designers working from references | Combine several reference images in one scene | Introduce objects into an atrium image
Research | Image editing researchers | Measure how much the unchanged area drifts | Run the Add and Preserve benchmark

You can skip it if you only make single edits, since the card reports scores close to the base model there.

Where to use it:

• The official SenseNova inference package, with only the model ID changed
• Python with the transformers library, on a machine with a large GPU
• Windows, macOS or Linux, to download the files

Use case | Who | What they get back
Multi-turn editing | Designers | A scene whose structure stays legible after many edits
Multi-reference editing | Creative teams | Objects added without rebuilding the scene
Text-to-image and single edits | Anyone using the base model | Results the card says stay close to the base

How to set it up: Reuse the environment you already have for the base model. The card says no adapters, merging or extra runtime are needed, and only the model ID or local path changes.

• Python and pip
• The official SenseNova inference package and its environment
• A GPU with room for about 35.1 GB of BF16 weights

1. Set up the official SenseNova inference environment.
2. Download the TailSteer files.
3. Replace the original model ID with this one.

pip install -U huggingface_hub
hf download sensenova/SenseNova-U1.5-8B-MoT-TailSteer

How to use it: Run the package's editing script with the TailSteer model path, an input image and an instruction. For the next edit, feed the output back in. Use TailSteer for the whole run, with no model switching or mask.

python examples/editing/inference.py \
  --model_path sensenova/SenseNova-U1.5-8B-MoT-TailSteer \
  --image input.png --prompt "Change the jacket to blue." --output output.png

Tool | What it does
examples/editing/inference.py | Runs one edit on an input image from a text instruction
--model_path | Points the pipeline at the TailSteer model
--image, --prompt, --output | Set the input picture, the instruction and the result file

Example: You pass a photo as input.png with the instruction "Add a small red boat on the lake." The script writes output.png with the boat added. The card says the rest of the scene is preserved better over repeated edits than with the base model.

For newcomers and experts: If you are new, start with one edit using the command above and look at output.png. Then feed that image back in for a second edit and compare the two.

For experts, TailSteer is a training stage added at the end, and it keeps the base model's learned abilities for the most part. The card measures drift with Add and Preserve, which asks for a small object in a red box each turn across 100 images and ten turns. The cost is a slightly lower RISE accuracy, 29.44 against 31.39 for the base.

Input image and instruction
TailSteer model, same pipeline
Edited image
Feed it back for the next edit

Good to know:

Title: PSNR at turn 10 outside the edit, higher is better
Base model: 11.23 dB
After TailSteer: 18.22 dB

• Single-turn scores: close to the base model on most benchmarks, lower on RISE accuracy
• Weights: complete weights, based on SenseNova-U1.5-8B-MoT
• License: Apache 2.0
• Cost to run: no API key or paid service named, but it needs a capable GPU
• Data it reaches: downloads files from Hugging Face, nothing else is named
• Not for: use where RISE accuracy matters most, since it dropped slightly

On Hugging Face, 13 November 2026

• Task: any to any
• Library: transformers
• Parameters: 17.5B
• Weights: about 35.1 GB in BF16; about 8.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 198
• Likes: 5
• License: apache-2.0
• Based on: sensenova/SenseNova-U1.5-8B-MoT
• Tags: transformers, safetensors, neo_chat, feature-extraction, tailsteer, image-editing, text-to-image, any-to-any

Links

• Model: https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT-TailSteer
• Files: https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT-TailSteer/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sensenova/SenseNova-U1.5-8B-MoT-TailSteer

Then follow the usage in its model card. Nothing in this kit runs by itself.