How Alibaba’s Taobao Starry Model Delivers Precise, Consistent E‑commerce Image Edits

1. Taobao Starry · Image Editing: Precise Consistent E‑commerce Editing

Current image editing techniques often struggle with consistency in e‑commerce scenarios, where merchants need only subtle adjustments without changing the product’s main subject or layout.

To meet this demand, Alibaba introduced the Taobao Starry · Image Editing model, focusing on high consistency for product‑level images, enabling precise, high‑fidelity edits while keeping the subject and layout unchanged.

2. The "Last Mile" of E‑commerce Image Editing

Although mainstream editing tools produce impressive visual effects, they often fall short in e‑commerce because they lack consistency, precise positioning, and domain‑specific capabilities such as text editing or lighting adjustments.

Inconsistent results : Minor changes can unintentionally alter the product’s appearance or background.

Imprecise positioning : Models may guess wrong locations, reducing controllability.

Missing e‑commerce features : Functions like text editing or lighting control are either poorly supported or produce unsatisfactory outcomes.

Most tools are designed for generic scenarios and do not optimize for product‑level stability, high‑fidelity details, or exact layout preservation, creating a “last‑mile” gap.

3. Consistency‑Centric Solution

1. Data: Building a Consistency Dataset for E‑commerce

The core is not to rebuild images from scratch but to make precise, lossless modifications while respecting the original design. Alibaba constructed the first large‑scale, high‑consistency e‑commerce editing dataset, covering over 20 editing tasks and reaching a million‑scale of high‑quality samples.

2. Model: Plug‑in, High‑Consistency Editing Framework

The design follows a "base model + mode‑conversion module + consistency‑enhancement module" hierarchy, offering flexibility and extensibility to adapt various foundation models.

Multi‑stage multilingual training strategy improves e‑commerce understanding:

Stage 1 (Mode‑Conversion) : Trains the conversion module with both open‑source and high‑quality e‑commerce data, enabling the model to handle a wide range of edit commands.

Stage 2 (Consistency Enhancement) : Fine‑tunes the consistency module solely on high‑quality e‑commerce data to boost performance on precise, consistent edits.

Mixed Chinese‑English instruction training : Allows accurate handling of bilingual prompts for product names, attributes, and complex operations.

4. Taobao Starry · Image Editing in Action

The model excels at background replacement, product swapping, color adjustment, object addition/removal, and text modification, with special upgrades for e‑commerce pain points such as intelligent text removal, portrait lighting enhancement, hairstyle and expression changes, clothing replacement, and robust Chinese instruction support.

Examples

Below are representative prompts and results (images omitted for brevity).

5. Summary and Outlook

Current image editing can be seen but not fully controlled. In e‑commerce, "accurate edits without changing appearance" is the true standard.

The proposed Taobao Starry solution, centered on consistency, combines high‑quality data and a plug‑in architecture to overcome stability, precision, and Chinese language limitations of mainstream models, showing clear advantages for e‑commerce‑specific tasks.

Future work will deepen fine‑grained control, multimodal collaborative editing, and real‑time interaction, moving the model from merely usable to genuinely useful for merchants.

Technical report: https://arxiv.org/abs/2510.04483