Inspiration
Inspiration
Creators kept asking for a fast, no-learning-curve way to make ecommerce and social images pop.
Background removal tools were either too slow, too pricey, or too “one-trick.”
We wanted AI that feels instant and controllable—from cleanup to full character stylization (LEGO/Barbie/3D).
What it does
One-click background remover (web-based), smart object erase, and background replace.
AI Action Figure Generator: turn selfies into stylized 3D/LEGO/Barbie-like characters.
Batch processing for sellers: consistent cutouts, size presets, color correction.
Before/after preview with WebGL/WebGPU, export in PNG/JPG/WebP, up to high-res.
Templates for ads/social, plus brand color/size presets.
How I built it
Frontend: React + Next.js (TS), Tailwind; WebAssembly + WebGL/WebGPU for realtime preview.
Models: PyTorch + Diffusers (SDXL/ControlNet), SAM for segmentation; ONNX Runtime/TensorRT for speed.
Backend: FastAPI (Python) for inference endpoints; Node.js BFF; Redis queues; gRPC/REST.
Infra: Docker + Kubernetes; GPU nodes on AWS/GCP; S3 + CloudFront for assets.
Data/ops: PostgreSQL (users/billing), Sentry/Prometheus/Grafana, PostHog for product analytics.
Auth & billing: Auth0/Clerk (magic links, OAuth) + Stripe subscriptions/credits.
Challenges I ran into
Achieving clean hair/fur edges without halos at small image sizes.
Balancing speed vs. quality on commodity GPUs and occasional CPU fallbacks.
Preventing style drift in character generation while keeping outputs diverse.
Building a safe queue that handles burst traffic without timing out web users.
Making batch tools intuitive for non-technical sellers.
Accomplishments that I'm proud of
Sub-second previews for most 2–4MP photos with crisp mattes on hair and semi-transparent objects.
A prompt-free Ghost/figure pipeline that produces consistent poses and lighting.
Templateable workflows: users can save brand presets and reproduce a look reliably.
Clear, minimal UI that gets a usable result in under 30 seconds end-to-end.
What I learned
Good UX beats raw model scores: controls + instant feedback matter more than SOTA numbers.
Post-processing (alpha refinement, color harmonization) often delivers bigger perceived gains than another model hop.
Observability early (traces + product analytics) pays off when diagnosing edge-case images.
What's next for PixPretty AI Photo Editor
Mobile web & desktop app with offline (on-device) inference for quick edits.
Style LoRA marketplace and brand kits (shared palettes, typography, shadows).
Smart scenes: scene-aware relighting and shadow synthesis for composites.
Team collaboration: comments, versioning, and approval flows.
Public API/SDK so marketplaces and CMSs can batch-generate listings automatically.

