Self-hosting removes the upload entirely

Running an open-weight vision model on your own hardware lets you score a photo without uploading it anywhere, which makes every retention, logging and server question moot at once. The trade is convenience and, often, accuracy against a tuned service.

What "open-weight" actually means

A closed model is a service: you send a photo, a company's server runs its model, you get a number back. An open-weight model is a file - the trained parameters themselves, published for anyone to download and run. Several general-purpose vision and vision-language models are released this way, and a scoring pipeline built on one behaves the same wherever it runs, because the model doing the scoring works the same way regardless of who owns the server. The weights do not know or care whether the machine running them belongs to a company or to you.

Open-weight is not the same as open source. The Open Source Initiative's Open Source AI Definition 1.0 asks for the parameters plus "the complete source code used to train and run the system" and detailed information about the training data. A weights-only release does not meet that bar, which matters for self-hosting only in one way: you can run the model, but you cannot see what it learned from.

This is a different question from running a model on your phone inside someone else's app, which still ships you a closed model you did not choose and cannot inspect. Self-hosting means you picked the model, you control the code around it, and you can read exactly what happens to the file at every step, because you wrote or can read that step.

What you get

The honest gain is narrow and real: nothing leaves the device. No upload, no server log of the request, no third-party API in the middle, no retention policy to read because there is no retention beyond what you choose to keep on your own disk. Every question about what happens to a photo after it is sent somewhere stops applying, not because the answers got better but because there is no longer a somewhere.

That also means no dependency on a company staying in business, changing its terms, or getting acquired by one with different terms. A model file on a drive keeps working after the service that inspired it shuts down.

What you give up

Convenience, first and most obviously. Running a model yourself means installing a runtime, downloading several gigabytes of weights, and having hardware that can run inference in a reasonable amount of time - a task a phone can sometimes manage and a laptop without a dedicated GPU often cannot without a long wait per photo. The gear question is real enough that measuremycock.com's gear coverage is a better place to work through hardware specifics than this site, which stays on the model rather than the box it runs on.

Quality is the less obvious trade, and it is real too. The best-performing scorers are usually closed, tuned on proprietary data, and updated by teams with a reason to keep improving them. An open model general enough to be published freely was rarely trained specifically for this task, so a self-hosted scorer is often a general vision-language model wearing a rubric prompt rather than a purpose-built head trained on labelled examples of exactly this kind of image. It can still produce a usable score. It is unlikely to match a system purpose-built and continuously tuned for the job, which is the whole reason a hosted product exists rather than everyone running the same open weights at home. Comparing a homemade setup against a maintained one is exactly the kind of before-and-after check penisrater.com's tools coverage walks through, and it is worth doing before trusting either.

Where the effort actually goes

Setup is the bulk of the cost: choosing a model, writing or adapting a prompt or scoring head, and testing it against enough images to know whether the numbers mean anything. None of that is exotic - it is the same rubric-design work any scoring system does, just done once, by you, instead of by a product team maintaining it across thousands of users. Once it runs, a self-hosted setup does not improve on its own the way a hosted product does with each retraining cycle, so the gap between it and a maintained service tends to widen rather than close over time.

Where this fits

Self-hosting is the correct answer to one specific worry: not trusting any company with the photo, full stop. For that worry, nothing else on this list of options actually resolves it - anonymous uploads, VPNs and short retention windows all still involve sending the file somewhere. For every other worry - reliability, calibration, wanting a second opinion - the more relevant comparison is not a model at all but a person, who reads context a model structurally cannot, or a hosted tool built and maintained for consistency rather than privacy. Running a model yourself is a real option for a narrow reason, not a general upgrade, and knowing which reason you have decides whether the setup is worth the evening it takes.