Virtual Try-On Still to Generic Image-to-Video Animate (fal VTON)

A dedicated virtual try-on model composites a garment photo onto a model photo to produce a photoreal try-on still, then a separate general-purpose image-to-video model animates that still into a short motion clip, with no source performance video required to drive the motion.

Properties

family: image-to-video
modality: garment image + model image -> try-on still -> video
whyMultiModel: Garment compositing (matching fabric drape, print and fit to a person's pose and geometry) and video motion synthesis are different learned tasks; the try-on model has no temporal capability at all, and the i2v model has no garment-fitting capability, so a still-image try-on specialist is chained into a separate general-purpose animator. This differs from the tracked diffusion-transformer-video-try-on (MagicTryOn), which requires an actual person video and preserves that video's real motion inside one video-native DiT; here the input is two stills and the motion is generated from scratch by a generic animator, not preserved from a real performance. It also differs from the tracked idm-vton-virtual-try-on, which stops at a still image and never reaches video.
steps: [object Object], [object Object]
controls: garment image and model image inputs; try-on output resolution (864x1296); i2v motion/scene prompt; clip duration.
exampleStack: fal-ai/fashn/tryon/v1.6 (garment composite) -> fal-ai/bytedance/seedance/v1/pro/image-to-video (animate).
useCases: [E-commerce garment video previews generated from static product and model photos, no film shoot][Fashion social content showing fabric motion without a physical try-on session][Rapid per-SKU try-on video variants at catalogue scale]
pitfalls: The i2v stage has no access to real garment-motion physics, so cloth drape and fabric behavior during motion are hallucinated rather than derived from an actual performance, unlike video-native try-on models that condition on a real source video; front/back consistency across the generated motion is not guaranteed since the still only shows one angle. The template also lists a fal-ai/any-llm/vision endpoint whose role in the graph is not documented on the public page, so it is not counted toward the two-model bar.
source: fal VTON workflow template (fal.ai)

Provenance

Source: https://fal.ai/workflows/templates/template/fal-vton

Properties

family

image-to-video

modality

garment image + model image -> try-on still -> video

whyMultiModel

Garment compositing (matching fabric drape, print and fit to a person's pose and geometry) and video motion synthesis are different learned tasks; the try-on model has no temporal capability at all, and the i2v model has no garment-fitting capability, so a still-image try-on specialist is chained into a separate general-purpose animator. This differs from the tracked diffusion-transformer-video-try-on (MagicTryOn), which requires an actual person video and preserves that video's real motion inside one video-native DiT; here the input is two stills and the motion is generated from scratch by a generic animator, not preserved from a real performance. It also differs from the tracked idm-vton-virtual-try-on, which stops at a still image and never reaches video.

steps

[object Object], [object Object]

controls

garment image and model image inputs; try-on output resolution (864x1296); i2v motion/scene prompt; clip duration.

exampleStack

fal-ai/fashn/tryon/v1.6 (garment composite) -> fal-ai/bytedance/seedance/v1/pro/image-to-video (animate).

useCases

[E-commerce garment video previews generated from static product and model photos, no film shoot][Fashion social content showing fabric motion without a physical try-on session][Rapid per-SKU try-on video variants at catalogue scale]

pitfalls

The i2v stage has no access to real garment-motion physics, so cloth drape and fabric behavior during motion are hallucinated rather than derived from an actual performance, unlike video-native try-on models that condition on a real source video; front/back consistency across the generated motion is not guaranteed since the still only shows one angle. The template also lists a fal-ai/any-llm/vision endpoint whose role in the graph is not documented on the public page, so it is not counted toward the two-model bar.

source

fal VTON workflow template (fal.ai)