Running 3 AI Background Removal Models in Browser WebGPU/WASM at $0 Server GPU Cost
Cloud background removers all share the same business constraint: GPU inference costs money per image, so free tiers either compress previews to 0.25 MP, slap on watermarks, or require accounts โ and users now face anoth
Cloud background removers all share the same business constraint: GPU inference costs money per image, so free tiers either compress previews to 0.25 MP, slap on watermarks, or require accounts โ and users now face another migration with remove.bg shutting down its standalone web app on December 1, 2026.
When I built RemoveFree, the architecture goal was simple:
- $0 server GPU bill โ inference executes 100% on the user's device.
-
Zero image uploads โ you can verify
0 KBof image traffic in DevTools (or disconnect Wi-Fi after loading). -
Full original
WรHresolution โ no 0.25 MP preview caps, no watermarks, no signup.
Getting that to feel instant on both desktop and mobile required shipping three different ONNX models inside a Web Worker, solving some subtle WebGPU/WASM bugs, and building a <3 ms client-side quality advisor.
Here is the exact architecture and what broke when we benchmarked it across 14 synthetic ground-truth scenes and 50 real CC0 photos.
1. Why One Model Isn't Enough in the Browser
If you pick a single high-accuracy segmentation model (~94โ183 MB), mobile devices choke and first-visit desktop users wait ~38 seconds just for weights to download and compile. Pick only a lightweight portrait model (~6 MB), and e-commerce products or cars collapse (0.0000 IoU).
So RemoveFree splits the problem across three Apache-2.0 / MIT models in a dedicated module Worker (@huggingface/transformers v4.3.2 + onnxruntime-web/wasm):
| Engine in UI | Model Repo | License | Input Size | Runtime | Payload | Median Time (50 real photos) |
|---|---|---|---|---|---|---|
| Fast Universal (Default) | edgetools/u2netp |
Apache-2.0 | 320ร320 |
Self-hosted ONNX + WASM | 4.6 MB | 942.5 ms |
| Portrait Lite | Xenova/modnet |
Apache-2.0 | 512ร512 |
WASM (q8) |
~6.3 MB | 1,155.0 ms |
| Studio HD (Opt-in) | studioludens/birefnet-lite-512 |
MIT | 512ร512 |
WebGPU (fp16) / WASM (fp32) |
~94 MB (WebGPU) / ~183 MB (WASM) | 11,915.5 ms |
Key runtime details
-
Self-hosted Cache API + Hugging Face fallback (
u2netp): Servingu2netp.onnx(~4.57 MB) from same-origin/models/u2netp.onnxcached into Cache Storage (removefree-onnx-v1) eliminates third-party latency on repeat visits and keeps a Hugging Face resolve URL strictly as fallback. -
Multi-threaded WASM via COOP/COEP: Production headers set
Cross-Origin-Opener-Policy: same-originandCross-Origin-Embedder-Policy: credentialless. Inside the worker,ort.env.wasm.numThreadsscales up tomin(4, hardwareConcurrency)wheneverself.crossOriginIsolatedis true. -
Automatic safe fallback: If WebGPU adapter initialization or
birefnet-liteinference throws (driver crash, OOM, shader failure), the client seamlessly falls back tou2netpon WASM (allowFallback = true) instead of dead-ending the user's cutout.
2. Preserving Exact WรH Pixels Without Running 4K Through ONNX
Users care deeply about keeping a 4000ร3000 photo at 4000ร3000 โ not getting back a 512ร384 blurred crop.
The pipeline separates inference geometry from final pixel compositing:
- Only a downscaled copy (
scaleRgbaForModel, max dimension320or512) is transferred (ArrayBuffertransferable) into the Web Worker. - The worker normalizes inputs (ImageNet mean/std
[0.485, 0.456, 0.406]/[0.229, 0.224, 0.225]foru2netp, orAutoProcessorfor Transformers.js models) and outputs either rawlogits(sigmoid-mapped with clamping[-20, 20]) or normalized[0..1]alphavalues into an 8-bitUint8Array. - Back on the compositor side,
upscaleAlphaMask()bilinearly upscales that 8-bit matte to exact original(origW, origH)dimensions and writes it onto untouched original RGB bytes: for solid foreground (alpha >= 250), exact source RGB values are preserved verbatim.
Result across our benchmarks: 14 / 14 (100%) synthetic scenes (including 3840ร2160 4K) and 50 / 50 (100%) real photos preserved exact input dimensions.
3. Three Real Bugs Uncovered by Our 50-Photo CC0 Benchmark
Before finalizing the launch build, we ran all three engines against 14 synthetic ground-truth cases (metrics.json) and 50 real-world CC0 / Public Domain photographs across 7 categories (portraits, hair, jewelry, products, animals, thin_objects, difficult_bg). Testing real images surfaced three bugs synthetic smoke tests missed:
Bug 1: Tensor Input Key Mismatches Silently Triggering Fallbacks
AutoProcessor.from_pretrained emits { pixel_values }, but studioludens/birefnet-lite-512 expects input_image, Xenova/modnet expects input, and edgetools/u2netp expects input.1.
For a while, both alternate models threw inside the worker and silently triggered the fallback back to u2netp โ making our early "model switcher" silently run the same engine underneath. Reading sessions?.model?.inputNames?.[0] dynamically inside the worker fixed it immediately.
Bug 2: MODNet Shader Corruption on Chromium WebGPU
Running Xenova/modnet on Chromium WebGPU introduced patchy holes across otherwise clean human portraits, dropping Sobel edge alignment to 46.2%.
Locking modnet explicitly to quantized WASM (q8, ~6.3 MB) and restricting WebGPU exclusively to birefnet-lite on non-mobile GPUs restored portrait edge alignment to 77.5% at 1,422 ms median (and 1,155 ms overall across the 50-photo set).
Bug 3: Why BiRefNet Lite Fails Low-Salience Scenes More Often Than Uยฒ-Net-p
Even though birefnet-lite wins clearly on fine hair (0.9738 vs 0.8344 synthetic IoU; 75.5% vs 72.8% real hair edge alignment) and 4K edges (0.9641 vs 0.8590), across 50 real photos Uยฒ-Net-p actually achieved a higher non-empty subject retention rate (96.0% / 48 of 50) than birefnet-lite (90.0% / 45 of 50). On cluttered jewelry and thin objects, birefnet-lite occasionally flooded or dropped the foreground while u2netp kept the subject intact.
That confirmed u2netp (4.6 MB) is not just faster โ it is the safest universal default.
4. The <3 ms Client-Side "Studio HD" Advisor
Nobody wants to wait for a 94 MB download when a 4.6 MB pass in under a second already produced a clean cutout on a product photo or car.
Instead of downloading BiRefNet upfront, RemoveFree inspects the generated u2netp alpha mask right in memory (analyzeMaskForHdRecommendation, sampling 12,000 pixels in <3 ms) and triggers a 1-click "Enhance with Studio HD" suggestion only when any of four conditions holds:
-
Suspicious foreground ratio:
foregroundPct < 3%or> 94%(likely low-contrast / dropped subject) -
Semi-transparent halo / strands:
semiTransparentPct >= 7.5%(hair, fur, glass, smoke) -
High edge transition complexity:
edgeComplexityPct >= 19.5%(spokes, foliage, intricate silhouettes) -
High-resolution input:
max(width, height) >= 2000 px
Across our 50 real photos, this triggered on 48% of images (85.7% of thin_objects, 57.1% of hair, and only 25.0% of clean portraits), keeping simple cutouts under 1 second while guiding tricky images straight to BiRefNet Lite + manual Erase/Restore brushes.
5. Benchmark Summary
| Model | License & Runtime | Non-Empty Subject Rate (50 real photos) | Avg Synthetic IoU (14 scenes) | Fine Hair IoU | Median Time (50 real photos) | P95 Time |
|---|---|---|---|---|---|---|
Uยฒ-Net-p |
Apache-2.0 ยท WASM 4.6 MB
|
48 / 50 (96.0%) | 0.7796 | 0.8344 | 942.5 ms | 1,393.5 ms |
BiRefNet Lite |
MIT ยท WebGPU/WASM ~94 MB
|
45 / 50 (90.0%) | 0.8590 | 0.9738 | 11,915.5 ms | 16,463.1 ms |
MODNet |
Apache-2.0 ยท WASM q8 6.3 MB
|
44 / 50 (88.0%) | 0.3072 | 0.7065 | 1,155.0 ms | 3,197.3 ms |
Full benchmark breakdown and visual inspection notes are live at removefree.app/benchmark.
6. Honest Trade-Offs of Browser-Local AI
Client-side inference isn't magic, and there are places where cloud or server pipelines are still the right tool:
-
No REST API: Because computation happens inside your browser tab, RemoveFree cannot be called from automated headless server pipelines โ self-hosted
rembgor cloud APIs (e.g., Leonardo.ai) belong there. -
Cold cache on Studio HD: First-time WebGPU
fp16shader compilation +~94 MBdownload still takes ~10โ38 seconds depending on connection and GPU driver; repeat runs benefit from Cache API + warm pipeline reuse inside the worker. -
Difficult micro-details: Tiny jewelry chains and extreme low-contrast objects remain hard for zero-shot client models (
0.22โ0.43 IoUsynthetically), which is why interactive Erase/Restore brush refinement sits right next to the export button.
If you want to inspect it directly: open removefree.app, open your browser's Network tab (or toggle Airplane Mode once loaded), and run a cutout on a portrait or product photo.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ full credit and traffic to the original publisher.