A vision model reads a product photo the way a careful stranger would. It sees the object, the wear, the connectors, and any text printed on the thing. That last part does most of the work.
So the highest-value photo for an AI eBay listing tool is not the three-quarter hero shot. It is a flat, glare-free close-up of the model plate. On an S19k Pro, the end-panel sticker. On a Dell server, the pull-out service tag. I shoot that frame first on every unit that leaves here, before the pretty one.
What it reads reliably
From one decent photo, expect these back correct:
- Printed text. Model numbers, OEM part numbers, hashrate stickers, service tags. Identification comes from here.
- Connector and port layout, often the only way to tell one generation from the next when the case is identical.
- Visible cosmetic condition. Scuffs, dents, rack rash, corrosion, missing fascia.
What it cannot read, ever
A photo records light, not function. These six are not in the image:
- Whether it powers on. The most consequential fact about a used machine is invisible.
- Internal configuration. The chassis is the same at 8 GB or 64 GB.
- Battery health. A pristine laptop can hold twenty minutes.
- Firmware and tuning. Stock or VNish, boosted or rated, or someone else’s overclock.
- Hours and history. Two years in a dusty shed and two months on a rack look identical.
- Completeness. Rails, PSU, caddies, original adapter.
Those six are your job on every listing, forever. A tool that fills one in confidently from a photo alone is guessing, and you find out through a not-as-described case.
What a misread actually costs
I wanted a number, not an opinion, so I went to our own store of completed eBay sales. Method: every sold listing naming Antminer, Whatsminer, Avalon or Bitmain, 15 August 2025 to 14 August 2026. That is 4,829 sales. Lots excluded, because a lot price is not a unit price. Medians, with the sample size on every figure.
29% of those sales are not a machine at all. 1,389 are a hashboard, a power supply, a control board or a fan. And 2,349, very nearly half, do not name a model precisely enough to bucket automatically, even with a rules engine written for this exact hardware.
A hashboard and the miner it came out of photograph almost the same: a slab of aluminium with a Bitmain sticker. The plate separates them:
| Model | Whole used machine, median | One hashboard, median |
|---|---|---|
| Antminer S19 Pro | $299 (n=27) | $63 (n=16) |
| Antminer S19j Pro | $146 (n=65) | $72 (n=24) |
| Antminer S19 XP | $372 (n=64) | $110 (n=11) |
| Antminer S21 | $1,000 (n=44) | $400 (n=15) |
| All models pooled | $385 (n=498) | $100 (n=184) |
So one word of identification is worth double to five times, on machines that mostly sell under $400. Wrong in your favour is a return and a defect. Wrong against yourself and you gave away a working miner for hashboard money. Full per-model figures are on what used ASICs actually sold for.
Shoot the label, not the item
Most sellers photograph the item beautifully and the label badly. Flip it. Four rules fix nearly every misread:
- One frame dedicated to the spec plate, square-on and filling the frame.
- Kill the glare. Glossy labels blow out under a ceiling light or a phone flash, and a blown-out label is unreadable text. Shoot at a slight angle, or diffuse it. This causes more bad identifications than everything else combined. Same reason: never shoot through shrink wrap, an anti-static bag or a display case.
- One item per photo. Two miners in frame gets you a hedged description of neither.
- Include the port side. Generation differences live in the connectors, not the case.
Sometimes typing beats photographing
The part vendors do not advertise: for a mainstream catalogued product, typing the model number beats any photo. If I already know it is a ThinkPad T14 Gen 3, typing that gives an unambiguous identity and removes the vision step entirely. That is what AI Quick-Fill is for.
Photo-first wins when you don’t know what you have. Pallet buys, estate lots, the box of adapters out of a decommissioned rack. So the rule is photo for the unknown, typed model number for the known, and most days I use both.
Why a blank field is a good sign
An empty field is the design working. Given an ambiguous item, a poorly constrained model returns a confident, plausible, wrong value. A well constrained one returns nothing.
Our own software is not immune, and neither was my query above. Pulling those medians I checked the power-supply bucket for the S19k Pro and got $624, more than the whole machine. It was wrong. Eighteen listings matched and sixteen were complete miners whose titles said “include PSU.” Two were actually power supplies. A keyword saw the letters PSU and called it a part. That is the same failure a vision model makes, in text instead of pixels. I read the underlying titles for every figure in the table above, which is why there is no PSU column in it.
Blank costs ten seconds. Wrong costs a return, shipping both ways on a 30 lb machine, and a defect on the account. So when you compare tools, on what each one actually costs per listing as much as accuracy, feed each an ambiguous item and see which admits it does not know.
The sixty-second review
Before publishing anything drafted from a photo:
- Model number, read off the actual item, not off the draft.
- Condition claim. Did you test it? If not, do not say “tested.”
- The six invisibles above. Fill or disclaim each.
- Category. A wrong one is among the reasons an eBay listing never shows up in search.
- Price against your cost basis, not the comp median, which is the market’s number and not your floor.
Deciding what to source rather than what to list is the Opportunity Finder‘s job.
Try it on the next awkward item you cannot identify. Drop a photo of the spec plate in and see what comes back. The ING Listing Engine is free during the public beta, for Windows, no paywall. Download the ING Listing Engine.
Common questions
How many photos does it need?
One produces a draft. Two, a clear overall shot plus a square-on shot of the spec plate, produce a better one. Beyond three or four you are adding buyer-facing photos, not identification signal.
Does photo quality matter more than camera quality?
Yes, by a wide margin. A modern phone in diffuse light beats an expensive camera under a glaring ceiling bulb. Sharpness on the printed text is the only thing that matters.
What about an item still in its sealed box?
It reads the box label, which usually identifies the product, but says nothing about the contents. Describe it as sealed and unverified, because that is what you know.
Filed under AI Listing · Tagged Photos & vision AI, Dirty data · Browse all guides by topic