Virtual Try-on for Flux

BETA. Flux + Backend V0 only feature.

Garment / Clothing - with a model - Recommended

Generated images

Garment / Clothing - with a model - Recommended

Generated images

Sports virtual try-on

Generated images

Garment / Clothing - Flat lay - must indicate "flat lay" in title - not recommended

Generated images

More examples in the gallery

Overview

Virtual try-on is the term for the process of generating images of a person wearing a specific garment or clothing item. This can be useful for e-commerce websites, fashion designers, or even for personal use.

See pricing for the cost of this feature.

Usage

To use Astria's virtual try-on feature, you need to create:

1. A faceid fine-tune of the garment you want to try on. The faceid class name should be clothing, shirt or pants, coat or swimming suit. Class name is used to identify the garment in the image and remove the original model in case a reference image with a fashion model is used.
2. Optional: A model of the person you want to dress up ideally through fine-tuning

Garment reference image

When creating a faceid of the garment, two types of images are possible for use:

1. Reference image with a human model. This is recommended as it allows the virtual try-on AI model to infer relative position to the body and get a better understanding of the garment.
2. Flat lay image - a plain image of the garment. In this case, image title must include the wording flat lay in the faceid tune description.

Tips for inference

1. For the virtual try-on to work, generated image or input image must show elbows or lower part body. Use waist up or long shot in your prompt.
2. To reference the garment use the <faceid:xxxxxx:1> syntax, where xxxxxx is the faceid of the garment.
3. If a face of a specific model is desired, use a fine-tuned LoRA fine-tuned.
4. Use a short prompt to describe the background, pose and composition. For example woman wearing a black shirt and jeans, fashion editorial plain white background.
5. In case a reference image to the garment includes a fashion model, a better alternative to this, could be to use the input image as a reference pose as well as img2img to bring some of the background features into the image. See above example for dress and respective prompt below.
6. Use 9:16 aspect-ratio i.e: 768x1280 width x height for generating the image.
7. Use controlnet pose, preferably preprocessed stick figure. Set controlnet conditioning scale to a low 0.3-0.5 and
8. Turn on super resolution, super resolution details and inpaint faces for better results.
9. Avoid occlusions covering the garment in the generated image such as hands or a bag - The garment should be clearly visible.

Examples

Dress example prompt:

<lora:1533312:1.0>  <faceid:1350339:1> front view of ohwx woman wearing dress, fashion editorial plain, white background --control_guidance_end 0.5
num_images=1
negative_prompt=
seed=
steps=
cfg_scale=
controlnet=pose
input_image_url=https://sdbooth2-production.s3.amazonaws.com/lttvv8rz1olcf28cah48igiydwqi
mask_image_url=
denoising_strength=
controlnet_conditioning_scale=0.5
controlnet_txt2img=false
super_resolution=true
inpaint_faces=true
face_correct=false
film_grain=false
face_swap=false
hires_fix=true
backend_version=0
ar=1:1
scheduler=dpm++2m_karras
color_grading=
use_lpw=true
w=768
h=1280

Guns and roses flat lay t-shirt example prompt:

<lora:1533312:1.0>  <faceid:1823239:1.0> closeup view of ohwx woman wearing Guns And Roses shirt, fashion editorial plain, white background --control_guidance_end 0.5
num_images=1
negative_prompt=
seed=
steps=
cfg_scale=
controlnet=
input_image_url=
mask_image_url=
denoising_strength=
controlnet_conditioning_scale=0.5
controlnet_txt2img=true
super_resolution=true
inpaint_faces=true
face_correct=false
film_grain=false
face_swap=false
hires_fix=true
backend_version=0
ar=1:1
scheduler=dpm++2m_karras
color_grading=
use_lpw=true
w=768
h=1280