Tencent open-sources HY-WU for virtual try-on

HY-WU is an open-source image editing framework from Tencent, specifically built for high-fidelity virtual try-on, outfit transfer, and style synthesis.

Give it a base photo of a person and a reference shot of an outfit, and it cleanly transfers the clothes onto your subject while leaving the face, pose, and background completely untouched.

On-the-fly LoRA generation

Traditional virtual try-on usually forces a painful tradeoff. You either spend minutes fine-tuning an adapter for every new piece of clothing, or you pass reference images directly into a generic image-to-image pipeline and watch fine textures turn to mush.

HY-WU bypasses that entirely. It treats parameters as instant working memory. By encoding reference images and prompt instructions together, it synthesizes an instance-conditioned micro-LoRA on the fly and injects it straight into the frozen backbone during the forward pass.

Basically, you get the precision of custom fine-tuning with zero test-time training latency. Beyond swapping outfits between real people, it can even take clothes off a toy figurine and dress an adult human in them, alongside rock-solid facial identity transfer.

Frontier scores with massive hardware demands

In human preference evaluations, HY-WU consistently beats open-source rivals like Qwen ImageEdit, Long Cat ImageEdit, and Flux 2, while outperforming closed-source tools such as Cream 4.5 and GPT Image 1.5.

It only trails closed-source giants Nano Banana 2 and Nano Banana Pro, which edge it out slightly at 52.4% and 53.8% win rates respectively.

The catch is the hardware bill. HY-WU pairs an 8B parameter generator with an 80B base model (13B active tokens), demanding at least four 80GB or eight 40GB GPUs for multi-GPU inference.

Tencent has made the inference scripts, model weights, and Gradio interface fully available on GitHub, with a distilled, lightweight checkpoint promised down the road.

Why it matters

If you build e-commerce pipelines, digital fashion workflows, or creative character concepts, this is worth cloning right now. Proving that real-time adapter generation can rival proprietary models without per-task training is a major technical leap for open-source graphics.

That said, the steep compute barrier means solo creators and hobbyists will have to sit on their hands. Until Tencent releases the smaller distilled checkpoint, HY-WU remains a powerhouse reserved mostly for well-funded studios and enterprise clusters.