Curbing drift from repeated edits
China's SenseTime released on Hugging Face by Nov. 6 the weights for "SenseNova-U1.5-8B-MoT-TailSteer," a further-trained version of "SenseNova-U1.5-8B-MoT," a model that can generate and edit images. When an edited image keeps being used as the input for the next edit, small changes in parts that were not requested accumulate. According to the model description, a final training stage was added to curb this "drift," while the original model's capabilities were largely preserved. @aitrackerbot, an X account that tracks new model releases, reported the release on Nov. 6.
SenseTime's sensenova-u1.5-8b-mot-tailsteer weights are now public on Hugging Face.
According to Hugging Face records, the repository was created on Nov. 4 (Nov. 5 Japan time), and the time it was made public is not shown. The paper on the method was submitted to arXiv on Nov. 4. The authors are affiliated with MMLab at the Chinese University of Hong Kong, SenseTime Research and Shanghai AI Laboratory. As far as ORIENS could find, there is no notice of this release on SenseTime's official website or in its GitHub updates.
A weakness cited in U1.5's description
SenseTime released a trial version of U1.5 on July 31 and the full version on Aug. 20. The description of the full version lists "Complex editing drift" among known limitations, saying content can drift during wide-area edits or multiple edits. The new weights amount to an update addressing this weakness. On Oct. 30, the company also released the weights for SenseNova Loop, a model in a different series.
Better preservation outside the box after 10 edits
The main results in the model description are figures from "Add & Preserve," an evaluation prepared by the developer. It applies 10 successive edits that add an object inside a red box on 100 images and measures how much the area outside the box changed from the original image. PSNR at the 10th edit (a value indicating deviation from the original image; higher is better) rose by 6.99 dB from 11.23 to 18.22. LPIPS (a value indicating visual difference; lower is better) fell by 31.3% from 0.607 to 0.417. COA, which shows the share of images in which the requested object appeared, was nearly unchanged at 94.93% versus 94.78%.
| Metric | Original model | TailSteer |
|---|---|---|
| COA (%, higher is better) | 94.93 | 94.78 |
| PSNR@10 (higher is better) | 11.23 | 18.22 |
| SSIM@10 (higher is better) | 0.166 | 0.489 |
| LPIPS@10 (lower is better) | 0.607 | 0.417 |
Preservation figures keep falling as edits accumulate
Drift remains with TailSteer as well. PSNR keeps falling from 27.32 at the 1st edit to 18.22 at the 10th. LPIPS at the 1st edit was 0.161 for TailSteer versus 0.154 for the original model.
The paper also includes results from external evaluations. The success rate on MSE-Bench, which measures 5 edits, rose from 50.00% to 55.00%. On EdiVal-Bench, however, instruction completion rose from 40.21% to 42.31% with the released settings, while it fell with 2 variants using different training settings.
Short of GPT-Image-2.5 Sunburst's figures
The paper also measured other companies' models on the same evaluation. SenseNova's TailSteer version fell short of OpenAI's "GPT-Image-2.5 Sunburst" (98.29%, 18.55, 0.540, 0.378) on all 4 metrics: COA, PSNR, SSIM and LPIPS. The paper writes that open models using TailSteer match or exceed powerful closed models, but the example it shows in figures is the FLUX.2 [klein] 9B version.
On LPIPS at the 10th edit, the original model of Alibaba's "Qwen Image-2.1" scored 0.270, lower than the 0.417 of SenseNova's TailSteer version. Meanwhile, FLUX.1 Kontext [dev] has a high PSNR of 18.64 but a COA of only 34.77%. Models that often fail to add the requested object cannot be compared on preservation figures alone.
Single edits decline on 6 of 7 items
The TailSteer weights can also be used for text-to-image generation and single edits. Scores fell on 6 of the 7 items shown in the model description, with only GEdit ZH rising, from 8.30 to 8.31. The largest declines were on RISE, from 31.39% to 29.44%, and Qwen-IB EN, from 53.43 to 52.24. The paper says text-to-image generation and GEdit and ImgEdit scores are close to those of the original model.
| Evaluation | Original model | TailSteer |
|---|---|---|
| Qwen-IB EN | 53.43 | 52.24 |
| Qwen-IB ZH | 53.25 | 52.58 |
| BizGen ES (%) | 45.85 | 45.73 |
| GEdit EN | 8.33 | 8.27 |
| GEdit ZH | 8.30 | 8.31 |
| RISE (%) | 31.39 | 29.44 |
| ImgEdit | 4.43 | 4.41 |
Unclear points remain about evaluation conditions
Add & Preserve is an evaluation created by the developer: images were chosen from Open Images and MONET, box positions were set by hand, and GPT-6.1-sol was used to annotate the instructions. The data is public, so outside parties can run the same evaluation.
The paper does not describe any human evaluation, nor does it state which model was used to judge COA or how box boundaries were handled. The abstract says drift decreased over 15 to 20 edits, but the range evaluated with figures is mainly up to 10 edits, and the results for 20 edits are mostly image comparisons. ORIENS's research found no third-party reproductions or comparisons. As of Nov. 6, Hugging Face showed 2 downloads and 0 likes.
Drop-in replacement with the same architecture, licensed under Apache 2.0
The weights can be downloaded without an application. The license is Apache 2.0, and the original model's license and attribution notices also apply. The bundled documents state explicitly that the weights were modified.
The model description says "Same architecture. Same official model layout. Same loading API.," and says it can be used simply by replacing the model specified in the official inference software. These are complete weights that need no additional components, and in successive editing the edited result is fed back into the same process. No region specification or model switching is needed at inference time, it said. ORIENS has not tested the model.
Hugging Face displays the size as about 18 billion parameters (BF16). The model description gives no information on required GPU memory, an API or pricing. It also lists no URL for a TailSteer-specific code repository, and the project page cited in the paper could not be opened as of Nov. 6.
A wave of research on successive editing
Research on curbing drift was also published in quick succession in October. ACEdit submitted to arXiv on Oct. 15 a method that trains a model using its own outputs as inputs. MTEdit presented data on about 500,000 successive edits and an evaluation on Oct. 16. TailSteer's model description cites as a feature that it does not require training on correct edit pairs or successive-editing data. The comparison tables in the TailSteer paper do not include these 2.
No link to revenue shown
SenseTime's results for the first half of 2026, announced on Aug. 26, showed revenue of 2.91 billion yuan, of which generative AI accounted for 2.33 billion yuan. Net profit under International Financial Reporting Standards was 620 million yuan, and the adjusted net loss was 390 million yuan. No explanation from the company on the relationship between this weight release and revenue has been found.
