DEV Community

Cover image for I Built a Photorealistic AI Virtual Try-On App Entirely on a $150 Phone (And How We Pivoted When It Failed)

I Built a Photorealistic AI Virtual Try-On App Entirely on a $150 Phone (And How We Pivoted When It Failed)

Harun - solo dev on October 17, 2026

UPDATE 🚨 : guys we built another securities platform if you want to read it go here : https://dev.to/koda2026/iam-12-i-built-an-enterprise-ai-secur...
Collapse
 
adamthedeveloper profile image
Adam - The Developer ✨ •

awesome work as usual. i do have some questions though:

  1. Your pitch is β€œyour photo never leaves your phone,” but in v1.0 the photo is sent to a public Hugging Face Space. Who can access uploaded images there, how long are they retained, and have you told users this before they upload? If it’s the latter, how does the privacy-first claim hold up?

  2. Isn’t β€œsplit inference” normally a model partitioned across device and server? Here the phone only handles the UI and the full model runs remotely, which is just a standard client-server call to a hosted API. What does the split mean in practice, and what part of the pipeline is yours rather than IDM-VTON’s?

  3. The token being in the client bundle is fine as a stopgap, but β€œread-only” doesn’t protect your GPU quota. Anyone can pull it from the bundle and burn through it. Have you rotated it since publishing, and is there rate limiting in place until the Worker proxy ships?

Collapse
 
effessdev profile image
EffessDev •

Oh πŸ˜…

Collapse
 
akhourianmolkumar profile image
Akhouri Anmol Kumar •

Failed.

Collapse
 
akhourianmolkumar profile image
Akhouri Anmol Kumar •

I consider it as Specious product.

Collapse
 
effessdev profile image
EffessDev •

Why is everyone roasting him this much πŸ˜…

Thread Thread
 
adamthedeveloper profile image
Adam - The Developer ✨ •

was i too harsh as well? haha

Thread Thread
 
akhourianmolkumar profile image
Akhouri Anmol Kumar •

you did absolutely great, you should be more harsher next time
you did what is necessary.

Collapse
 
effessdev profile image
EffessDev •

How are you guys building stuff this fast 🀯 I remember you writing a post here about the plan, and instantly it's done?!

Collapse
 
koda2026 profile image
Harun - solo dev •

haha, pure stubbornness and a lot of late nights in termux! πŸ˜‚

but honestly, the speed came from a massive, necessary pivot. we realized running a full diffusion model locally was hitting a hard hardware wall on budget phones. so we implemented a split inference architecture.

it’s actually a concept currently in active research for privacy-focused vton: the phone handles the ui, strict data control, and auth, while the cloud handles the heavy diffusion lifting via secure websockets. it let us ship the privacy-first promise without melting the device or compromising the user's biometric data.

glad you checked it out! the journey from "all paths are impossible" to "βœ” AI Generation Complete" was wild. πŸ―πŸš€

Collapse
 
akhourianmolkumar profile image
Akhouri Anmol Kumar •

product is just specious

Collapse
 
technogamerz profile image
π“π‘πž π‹πšπ³π² 𝐆𝐒𝐫π₯ •

Superhero πŸ˜