I gave up hand-cutting product photos somewhere around number thirty

The photographer sent back a batch shot against whatever was available. Some studio grey, some on a desk, a few with props still in frame. The marketplace wanted white backgrounds, so the images had to be cut out one at a time.

Around the thirtieth I stopped and did the arithmetic. At that pace the batch would take three days. So I moved to batch processing. Here is what works and where it still breaks.

Three routes, and what actually separates them

Route	Time per image	Edge quality	Suits
Manual pen path	3 to 10 minutes	Most controllable	Small numbers where the edge must be exact
Classic algorithms, magic wand or colour range	Seconds	Fine on hard edges, poor on hair	Plain backgrounds with clear boundaries
AI matting	Seconds	Noticeably better on hair and semi-transparency	Large batches, cluttered backgrounds

TangyuanAI lists background removal among its capabilities (product overview, checked 6 August 2026), which puts it on the third route.

Where it breaks

This list cost me rework to assemble.

Fur and fuzzy edges. A ring of the edge gets eaten. It looks fine at normal size and shows at full zoom. I now budget rework for these.

Semi-transparent materials. Mesh, sheer fabric, frosted plastic. The edge goes grey after cutting and needs a second pass.

Subject and background the same colour. Dark grey product on dark grey background gives the algorithm nothing to separate. The cheapest fix is upstream: ask the photographer for a background with more contrast.

Shadows you want to keep. Cutting removes the shadow along with the background. Keeping it means handling it separately.

The order that works

Run ten representative images first and inspect the edges. If those ten are not acceptable, the remaining hundreds will not be either.

Group by background type next. Process images with the same background together so the parameters are comparable. Mixing them means you cannot tell which group caused a problem.

Fix the output spec before you start. White level, frame size, margin ratio, all decided up front rather than adjusted image by image afterwards.

Review side by side at the end, not one at a time. Batch problems only surface in comparison, for instance a few images where the subject sits slightly off centre.

One check people skip

After cutting to white, the subject edge often keeps a colour fringe from the original background. It shows most on a light product cut from a dark background, where the edge carries a grey halo.

Zoom to 200 percent and look at the edge. If there is a fringe, run a second pass. Finding it after the listing is live costs more than the rework.

When it is not worth it

Under about twenty images, cutting by hand is faster overall, because you skip building the pipeline, tuning parameters and reviewing in a grid.

The value of batching is in the scale, and twenty is not scale.

Sources

Capability description comes from the TangyuanAI product overview, https://tangyuanai.vip/en/about , checked 6 August 2026. The failure list and the ordering come from my own rework notes. Output depends on the uploaded material; nothing here promises a cutting quality.