Every photo library has the same three problems. The perfect landscape with a stranger standing in the corner of it. The old family photo scanned at a resolution that turns faces into mosaics. The product shot that needs to be on white, with a shadow, at Amazon's exact pixel size, by tonight. Three different tools usually solve these — and the standard pricing model for all of them in 2026 is credits: SnapEdit's free tier runs about 10 credits a month and stamps watermarks on exports; PicWish sells credit packs. You spend your credits carefully, or you subscribe.
The interesting thing about all three problems is that none of them fundamentally need a server. Object removal is a machine-learning model (LaMa) that reconstructs background pixels; upscaling is a model (or a good classical algorithm) that resamples; product-photo compositing is canvas math. Models run in browsers now — WebGPU for the heavy lifting, an automatic fallback otherwise.
Try Object Remover — free, no sign-up.
So: a free object remover that runs the actual LaMa inpainting model in your browser tab, an image upscaler with an AI engine and a classic fallback, and a product photo studio for the marketplace workflow. No credits, no watermark, no upload of your photos to anyone's processing queue. The models download once from a CDN and run on your machine; the only network traffic is that one-time download.
This guide covers how inpainting actually works (so you can predict when it will fail), brush technique, honest upscaling expectations, and the product-photo pipeline.
The short answer
Brush over the thing you want gone in the free object remover, hit run, and the LaMa model repaints the masked area with a plausible reconstruction of what was behind it. In a real end-to-end test, the default 60 MB quantized model completed a removal in about 29 seconds and produced a downloadable result — no credits spent, no watermark burned in. A 200 MB Big-LaMa option exists for harder cases. For the resolution problem, the free upscaler offers AI upscaling (Swin2SR) or an instant classic engine — the classic path took a 320×200 image to 640×400 in roughly half a second in testing — and for the marketplace problem, the product photo studio handles background, shadows, and platform size presets in batch.
How inpainting actually works
It helps to know what the model is doing, because it calibrates your expectations of what survives the brush.
Inpainting models are trained on a simple premise: hide part of an image, and ask the model to predict the hidden pixels from everything around the mask. LaMa — the model this tool runs, shipped in ONNX form so it executes in the browser — is unusually good at the two things that matter for photo cleanup: long-range structure (the fence line that continues through where the person was, the horizon that keeps its level) and repeated texture (grass, sand, water, wallpaper). It fills the hole by extending the visual logic of its surroundings.
That is also precisely the failure map:
- Large objects on complex backgrounds. A person covering 40% of frame over a busy cafe scene leaves the model guessing a lot; expect smeared, dreamy fill.
- Thin structures crossing the mask. If a wires-or-branch line enters and exits the mask, LaMa usually reconnects it — that's its strength. If the line only enters (cut off at the frame edge), the reconstruction may drop it.
- Faces and text in the fill zone. The model generates texture, not identities. Don't expect it to synthesize a believable face or a readable sign; expect it to invent plausible wall.
The mental model: you're asking a very good pattern-completer to finish a puzzle it can mostly see. It excels when the surroundings determine the answer.
Brush technique for clean removals
Technique changes results more than model size does:
- Mask slightly past the object's edge. The model needs the boundary pixels gone too; a tight mask leaves a one-pixel halo of the old object.
- Don't mask generously. Every extra masked pixel is a pixel the model invents. Overmasking is the top cause of mushy results.
- Work in passes for crowded scenes. Removing three people? Do them one at a time — each pass gives the next pass more real surroundings to work from.
- Use the big model when the subject deserves it. The 60 MB quantized model handles most cleanup; the 200 MB Big-LaMa option buys noticeably better structure on demanding fills. It's the same one-time-download deal.
- Downscale your ambition, not your image. If a fill comes back bad, try masking the problem area in pieces rather than re-running the whole thing.
The 29-second figure from testing is the default model on a normal photo-sized mask; bigger masks and the 200 MB model take proportionally longer. There's no queue and no meter — a bad result costs a retry, not a credit.
Upscaling without the hype
Upscaling ads promise detail from nothing. Reality is more specific: upscalers resample intelligently, and the gains are dramatic exactly where the original information constrains the answer.
- Text, screenshots, comics, line art, pixel art: the biggest wins. Edges are supposed to be crisp, so a 2x or 4x pass (the AI engine here is Swin2SR) produces results that look genuinely higher-resolution. Old document scans and game-era sprites are the poster children.
- Photographs: moderate wins. Skin and foliage look cleaner at 2x than any bicubic stretch, but the model is interpolating texture, not recovering pores. A soft photo becomes a larger, softer photo.
- The classic fallback: when you don't want a model download at all — or you're batch-processing dozens of files — the Lanczos classic engine resamples with traditional interpolation. It took a 320×200 test image to 640×400 in about half a second, and for clean sources the difference from AI is small.
The "no limit" part matters more than the "4x" part: cloud upscapers ration free tiers by image count and resolution caps. Here the constraint is your machine, and there's no count.
A practical pairing order for old photos: upscale first, then inpaint. Removals look better on the larger image because the model sees more surrounding structure — and the upscale gives the inpainter finer grain to fill with. (Both tools are siblings in the same editor family as the browser photo editor, for the exposure-and-crop work that usually follows.)
The product-photo pipeline
Marketplace listings eat time in a specific way: the photo is fine, but each platform wants it different. The product photo studio compresses that loop:
- Drop the raw shots. A batch of photos from the lightbox, kitchen table, or studio.
- Background to clean white (or your brand tone) — the background-removal pass runs in-browser like everything else.
- Add the shelf presence: a soft shadow so the product doesn't float, and text or a badge where the listing needs it.
- Pick the preset: Amazon's square, eBay's, Shopify's — the exact dimensions each platform asks for, rather than a resize-and-pray.
- Batch export. All variants out in one pass.
The economics comparison is stark: product-photo SaaS (the Glorify class) runs freemium with export caps for exactly this compositing work. The browser version's trade is honest — no team features, no cloud asset library, no AI product photography generation — but the mechanical core (cut out, shadow, resize, batch) is all there, at the price of zero.
What the credit systems are selling
To be fair to the paid tools: their credits pay for GPU time on servers, which is why they can run bigger models and queue thousands of images. The browser stack inverts it — your GPU does the work, the model downloads once, and nothing about your usage is metered because nothing about it costs anyone else compute. For a photographer cleaning 50 shots a week, a credit subscription may still be rational (bulk speed, bigger models). For everyone else — the person with one photobombed vacation photo, the reseller with 12 listings, the archive of low-res scans — the free path does the same job without the meter. And since the photos never upload, subjects who didn't consent to being processed on a stranger's server (kids, clients, houses) stay that way.
Frequently Asked Questions
Is the object removal a real AI model or a blur trick?
It's the real LaMa inpainting model running locally via ONNX in your browser — the same open model family used by server-based removers. Verified end to end: a default-model run completed in about 29 seconds and returned a downloadable cleaned image.
Do I need to download anything?
The model downloads once from a CDN the first time you run a removal — about 60 MB for the default quantized LaMa, or 200 MB if you switch to Big-LaMa. After that it's cached, and the only thing that moves is your mouse.
Will it work on big objects and busy backgrounds?
Best on small-to-medium objects with extendable surroundings (grass, sky, walls, water). Large objects over busy scenes get soft, invented fill — mask in pieces and use the bigger model for hard cases.
How much can I upscale, and is there a limit on runs?
The AI engine (Swin2SR) handles 2x and 4x passes; the classic Lanczos engine resamples instantly without any model. There's no count limit or resolution gate — your machine's memory is the only ceiling.
Does it add watermarks to exports?
No. Exports from the object remover, upscaler, and product photo studio are clean PNGs — no watermark, no forced sign-up, no account.
Can it make my blurry old photos sharp?
Partially. Upscaling enlarges and cleans edges; it does not recover focus that was never captured. Text, line art, and screenshots gain the most; soft photos become larger and cleaner but stay soft. Run the upscale before inpainting for the best removal results on old scans.
