JD.com Deploys AI‑Powered Virtual Try‑On Ahead of 618 Festival

What JD.com claims

JD.com announced that its AI virtual try‑on feature is now live across the main JD app. According to the company, users can upload a single portrait, the model extracts a skeletal pose, estimates body dimensions, and renders a realistic image of the selected garment on the user. The preview appears in roughly ten seconds, supports both men’s and women’s apparel, and includes one‑tap color changes and AI‑driven outfit pairing.

The rollout is timed for the upcoming 618 shopping festival, a three‑day sales period that typically generates over ¥300 billion in GMV. JD’s press release suggests the tool will lower return rates and boost conversion, positioning the company against rivals like Taobao and Douyin that are also experimenting with AI‑driven commerce.

What’s actually new

Model architecture

The feature builds on a pose‑guided generative network similar to recent work such as VITON (2022) and TryOnGAN (2024). JD’s engineering blog (see the official announcement) indicates they combine a HRNet‑based pose estimator with a conditional diffusion model that takes the user’s silhouette and the target garment’s mask as inputs. The diffusion step runs for only 5 inference iterations, which explains the sub‑10‑second latency on JD’s proprietary inference chips.

Data pipeline

JD reportedly leveraged its own catalog of over 10 million product images and a curated dataset of 1.2 million user‑generated photos (opt‑in only) to train the model. The data includes depth cues from the phone’s LiDAR sensor where available, improving the accuracy of body‑measurement extraction. This is a step beyond earlier virtual‑try‑on services that relied solely on 2‑D keypoints.

Integration with the shopping flow

The try‑on button is embedded directly on the product detail page. After the preview, the UI offers a “Swap Color” button that triggers a lightweight style‑transfer network, and a “Complete the Look” suggestion that draws from a recommendation engine trained on purchase histories. The entire pipeline runs on JD’s edge servers, reducing round‑trip time compared to cloud‑only solutions.

Limitations and practical concerns

1. Measurement accuracy – The system infers measurements from a single photo. While the diffusion model can produce visually plausible results, the underlying dimensions can be off by 5‑10 cm, especially for users wearing loose clothing in the uploaded image. This may still lead to fit‑related returns.
2. Privacy – JD stores the uploaded photo for up to 24 hours to enable the preview. The privacy policy (linked here) does not detail whether the images are used for further model training, which could raise regulatory questions under China’s Personal Information Protection Law.
3. Category coverage – At launch the feature supports tops, dresses, and sportswear. Pants, shoes, and accessories are absent, limiting its impact on high‑value categories where returns are most costly.
4. Device dependency – The 10‑second turnaround assumes a modern smartphone with a decent GPU. Users on older devices experience latency up to 30 seconds, which may deter usage.
5. Bias in generated images – Early user reports indicate the model struggles with body shapes that deviate significantly from the training distribution (e.g., very tall or plus‑size users). The rendered garments sometimes appear stretched or misaligned, which could affect user trust.

How this fits into JD’s broader AI commerce push

Beyond virtual try‑on, JD has been rolling out AI‑generated scripts for live‑stream hosts, chat‑based shopping assistants, and automated inventory forecasting. The try‑on tool is the most visible consumer‑facing component, but it shares the same underlying infrastructure: a suite of large‑scale diffusion and transformer models hosted on JD’s in‑house AI accelerator.

Analysts at iResearch (see their recent report PDF) estimate that AI‑enabled fitting tools could shave return rates by 1‑2 percentage points, translating to roughly ¥1‑2 billion in saved logistics costs for JD during the 618 period. However, the actual impact will depend on user adoption, which remains untested at scale.

Bottom line

JD.com’s AI virtual try‑on is a technically solid implementation of pose‑guided diffusion, integrated tightly with the shopping experience. It offers a faster preview than many competitors and adds modest stylistic features. The main hurdles are measurement fidelity, privacy handling, and limited category support. If JD can expand the model to cover more product types and improve accuracy for diverse body shapes, the tool could become a useful lever for reducing returns during high‑volume sales events like 618.

---

For developers interested in the underlying tech, JD has open‑sourced a trimmed version of the pose estimator on their GitHub org. The diffusion component remains proprietary.