Video and live try-on

2.9Video virtual try-on

Video try-on has moved in three waves. In 2024, methods such as ViViD added motion layers to Stable Diffusion 1.5 and ran a second network over the garment [78]. In 2025, video diffusion transformers took over: CatV2TON concatenates garment and person in time, MagicTryOn fine-tunes the 14-billion-parameter Wan2.1 model, and DreamVVT tries the garment on a few keyframes first, then generates the video around them [79, 80, 81]. In 2026, almost every new paper fine-tunes a Wan2.1 or Wan2.2 variant, often with LoRA, and the field is moving to mask-free inputs, synthetic training pairs and keyframe-first pipelines [82, 83, 84, 85].

The papers name the same unsolved problems. Garment detail degrades in close-ups and at higher resolution, and no video try-on paper reports a metric for printed text or logos over time. Fast motion, occlusion and back views still break models. Long videos drift at window boundaries. Per-frame masks and pose maps cost about a second a frame to compute, which is more than a distilled generator needs, so live try-on has to be mask-free [85, 71].

2.10Real-time video generation

Standard image and video generators start from noise and clean it up over 20 to 50 passes, taking seconds for one image. Live try-on needs a new image every 33 to 66 milliseconds. The field got there with four published ideas, and every leading real-time system uses some mix of them:

• Fewer passes, through distillation. A slow, high-quality teacher trains a fast student to do the same job in 1 to 4 passes [87].
• Frame-by-frame generation with memory. Each frame is drawn from the live camera frame and the model's own recent frames, kept in a key-value cache, so the garment does not flicker or change design between frames [88].
• Training on its own mistakes. Small errors normally pile up frame after frame. Self-forcing and related methods train the model on its own imperfect outputs, so it learns to correct drift instead of amplifying it [89].
• Working small and engineering hard. Models draw a compressed image and then upscale it. They run on top GPUs with custom kernels, 8-bit or 4-bit arithmetic, sparse attention and a direct stream with no queue [90, 91].

The public recipe for a small team follows from these results: start from an open video diffusion transformer, train a garment-conditioned video editor, convert it into a causal student with self-forcing and step distillation, and keep a persistent memory of the garment. Before any custom kernel work, published systems of this kind reach about 11 to 24 frames per second at around 480p on one high-end GPU [93, 96, 91]. Many popular real-time repositories cannot be used commercially: CausVid and the LongLive weights are non-commercial, and Krea's code is too, although its weights are Apache-2.0 [87, 92, 93]. The Self-Forcing code and the Wan2.1 models are Apache-2.0 [89, 97].

2.11Live try-on products today

Decart is the clear leader in live try-on. Its Lucy VTON 3.5 model, released in August 2026, takes a live camera stream plus an optional garment photo and text prompt, and returns 720p video over WebRTC, the technology video calls use. Its SDK defines the model at 1280×720 and up to 30 frames per second. Sessions use short-lived client tokens, garments can be switched mid-session, and prompts and garment images are moderated on the server. The list price is USD 0.02 a second, or USD 1.20 a minute, and a faster mode costs double [72, 70, 98]. Decart discloses the principles, not the recipe: causal frame-by-frame generation, training on its own outputs, step distillation, custom GPU kernels and low-precision arithmetic [88, 90].

For FabricVTON, Decart is both the benchmark to beat and a way to prove shopper demand while its own model is built. It will stay ahead on general live video. FabricVTON needs to win only on clothes: prints and logos that stay exact, fabric that looks like the real product, a much lower cost per minute, and a model it controls.

2.12Video base models, data and their licences

The licence rules for images apply to video. Almost every released video try-on checkpoint is non-commercial, but the Wan video models they build on are licensed Apache-2.0. A company can build on Wan, as long as it trains its own weights on its own data. Wan 2.5 and later have no open weights. Several other open video models carry traps: LTX-2 charges above USD 10 million of revenue and bars competing products, and HunyuanVideo's licence does not apply in the EU, UK or South Korea [99, 100].

No public video try-on dataset is clearly cleared for commercial training. The approach the field has settled on is the synthetic-triplet idea applied to video: keep a real video as the target, create the input by re-dressing it synthetically, and pair it with the real garment photo [102, 85]. The model then only ever learns to reproduce real footage.

2.13Gap analysis

2.14Summary of the current understanding