Inpaint-web (Open-Source Browser WebGPU Inpainting) · Object removal
Free open-source, login-free and ad-free on-device processing exports full size; complex bookshelf repairs show blocky texture and need inspection.
US-West (California, AS2914)
No native export attached; page screenshot evidence only
Subsequent pricing and service checks are shown separately without altering historical scores.
One run does not establish all complex scenes, untested paid maxima or long-term uptime.
Plans & historical pricing
Inpaint-web in-browser open-source object removal and upscaling tool · 2026-10-21
1 plan / billing-period records, all listed; — means that period is not recorded.
Open-source free web tool (MIT License) powered by client-side WebGPU and WebAssembly. Models run locally in the browser with no registration, paid plans, or usage limits.
Cancellation, sources & verification record
Completely free open-source project with no accounts or billing.
Inpaint-web official web workspace ↗ · 2026-10-21 Capture 1 ↗
Pricing recorded on the test date · 2026-10-16
$0 (MIT Open-Source Free Project)
100% free open-source; compute runs on user's local browser WebGPU/WASM
Historical context only; not current purchase guidance.
Delivery & access
- Free delivery record
- 2048×2048 (Full-Res Unwatermarked PNG)
- Watermark
- None (100% watermark-free)
- Free allowance
- 100% free and open-source; unlimited local browser inference and full-resolution exports
- Account requirements
- No registration or login required; works immediately in browser
- Maximum specification record
- 2048×2048 (Full-Res Unwatermarked PNG)
May include advertised capabilities; this does not establish a tested paid tier.
- Batch & API
- Supports multi-step history snapshots, hold-to-compare original, and built-in local 4x upscaling
- Manual refinement
- Not recorded or not verified
Tested result & sample
Inpaint-web runs the MIGAN inpainting model locally in the browser via ONNX Runtime Web, erasing all 3 signs directly on the 2048×2048 image and exporting a full-size PNG without uploading a single byte to any server. Smooth wood surfaces restore cleanly, while complex bookshelves show lightweight client-model texture patterns.
- Fixture
- anchor_inpaint_v1
- Recorded environment
- US-West (California, AS2914)
- Local processing record
- 13.8 s
- End-to-end reference
- 13.8 s
Network wait includes round trips, queuing and polling error, not pure inference time. Local timings depend on the device; repetition counts are not standardized in legacy records.
No native export is attached. Screenshots show the visible page, but do not independently establish export properties or file validity.
Experience, rules & trust
- Disclosure
- Completely free and open-source with zero hidden charges
- Interaction
- Minimalist open-source workspace with zero popups, ads, or email gates; automatically preserves step-by-step history thumbnails along the top bar (02_workspace.png, 03_download.png).
- Entity record
- Inpaint-web Open Source Project (lxfater)
- Origin
- Open-Source Developer Community (GitHub: lxfater/inpaint-web)
- Domain history record
- 2023–Present (Flagship WebGPU client-side AI vision project)
- Third-party tracking record
- Not recorded or not verified
- External information
- Not recorded or not verified
- Processing
- Pure browser-side ONNX Runtime Web (WebGPU / WASM) local inference (~13.8s per stroke)
- Retention statement
- Zero server transmission; image data stays 100% in local browser memory and is destroyed when the tab closes
A recorded statement does not verify deletion or all data-processing behavior.
Verifiable audit trail & evidence
Four SHA-256-verified stages link the benchmark fixture, same-run interface screenshots, native exported artifact and archived pricing sources.
anchor_inpaint_v1 · Inspect fixture & file inspector →
Captures entry rules, result view and export disclosure.
Page screenshot evidence only
Screenshots, native output and official claims support different conclusions. Evidence grades are not product safety certifications.
Evaluation basis & full scores · Overall 82.6/100
Overall = honesty & reputation × 30% + capability × 70%. Component maxima follow this report’s rubric. List scores are rounded for display; original precision is retained here.
Scoring version 2.3.1 · Test fixture anchor_inpaint_v1
Honesty & reputation · 91/100 · 30%
| Criterion | Score | Rationale | Basis |
|---|---|---|---|
| Pricing transparency | 34 / 35 | 100% free open-source tool (01_landing.png): zero registration, zero login, unlimited runs, and no subscription paywalls. Clicking Download exports a native 2048×2048 unwatermarked PNG (03_download.png). | Evidence A → |
| Interaction restraint | 29 / 30 | Minimalist open-source workspace with zero popups, ads, or email gates; automatically preserves step-by-step history thumbnails along the top bar (02_workspace.png, 03_download.png). | Evidence A → |
| Background & reputation | 14 / 20 | Popular GitHub open-source project lxfater/inpaint-web powered by ONNX Runtime Web / WebGPU with fully auditable MIT-licensed source code. | Evidence B → |
| Engineering & privacy | 14 / 15 | 100% local browser-side WebGPU/WASM execution: once the ONNX model is cached in IndexedDB, all pixel processing happens inside local browser memory with zero server uploads. | Evidence A → |
Capability · 79/100 · 70%
| Criterion | Score | Rationale | Basis |
|---|---|---|---|
| Task quality | 49.5 / 65 | Masked GT Reconstruction (21.5/30: structural PSNR=14.63dB · SSIM=0.4166) · Unmasked Fidelity (17.5/20: zero pixel alteration outside brush strokes) · Texture Continuity (10.5/15: smooth wood table restoration, though complex bookshelf spines show stylized MIGAN block patterns) | Evidence B → |
| Deliverable specifications | 20 / 20 | Exports native 2048×2048 full-resolution unwatermarked PNG without login (20/20) | Evidence B → |
| Processing speed | 9.5 / 15 | Client-side ONNX/WASM 2048×2048 local inference takes ~13.8s per stroke with 0.0s network upload overhead | Evidence B → |
Legacy scores may include advertised maximum capabilities. Read tested output separately from paid capabilities. Records without matching tiers and scope are not ranked together.
Revisions & corrections
- 2026-10-16
Initial real-machine benchmark using Mode A synthetic occlusion ground-truth anchor (anchor_inpaint_v1: 3 occluding signs -> 2048px GT).