A second opinion for AI photo descriptions. Several vision models, running on your own computer, describe the same photo. Crosscheck tells you what they all agree on, and, first and most clearly, where they disagree.
Built for blind and low-vision people, who use AI descriptions for labels, medicine, mail and screens but have no easy way to tell when a description is wrong.
A single AI description sounds equally confident when it is right and when it is wrong. A study at ASSETS 2025, Surfacing Variations to Calibrate Perceived Reliability of MLLM-generated Image Descriptions (Meng Chen, Akhil Iyer, Amy Pavel), found that when blind and low-vision participants were shown where several models' descriptions varied, they spotted 4.9 times as many unreliable claims as with a single description (2.62 against 0.53 per image). 11 of 15 participants ranked that presentation as their favourite, and all of them wanted to use it in the future.
Crosscheck brings that idea to models that run locally through Ollama: free, offline, and the photo never leaves your computer.
-
Each model lists what it sees as short, separate facts, and quotes any text it can read.
-
Facts that say the same thing are merged. Facts that differ in a number or quoted text ("Ibuprofen 200 mg" against "400 mg") are kept apart, because that is where errors hide.
-
Every model looks at the photo again and judges each fact the others stated: yes, no, or unsure.
-
If you asked a question, each model also answers it in a few words, and the models judge each other's answers the same way. "Cannot tell" counts as not answering, not as disagreeing.
-
You hear a short summary first, for example:
A pill bottle on a wooden table. The 3 models agree on 6 of 9 details. They disagree about: The label says "Ibuprofen 200 mg"; The label says "Ibuprofen 400 mg".
Then, on demand: what every model agrees on, where they disagree (and which model said what), and details not every model could judge.
Agreement is not proof. Models can make the same mistake. For medicine, money or safety, confirm with a person or another trusted source. Crosscheck never says "verified".
-
Install Ollama and pull three vision models from different families (different families make different mistakes):
ollama pull qwen3-vl:4b-instruct ollama pull gemma4:e4b ollama pull ministral-3:3b
-
Install and start Crosscheck (Python 3.10+, no other dependencies):
uv tool install crosscheck-vision # or: pipx install crosscheck-vision crosscheck serveIt opens http://127.0.0.1:8765 in your browser.
Choose other models with --models a,b,c (or CROSSCHECK_MODELS). With a 12–16 GB GPU, expect
about 30–90 seconds per photo while Ollama swaps the models in and out. A CPU works, slowly.
- Web app: choose a photo, optionally ask a question ("What does the label say?"), press Check photo. Progress is announced, and when the result is ready, focus moves to it. Everything works with the keyboard and a screen reader. A checkbox can also read the result aloud with the browser's own voice.
- Terminal:
crosscheck describe photo.jpg --question "What does the label say?", or--jsonfor the full result.
The web app shrinks the photo in your browser before sending it, which also removes metadata such as GPS location.
crosscheck serve --lanCrosscheck then also listens on your home network and prints an address such as
http://192.168.1.20:8765/ and a 6-digit pairing code. Open the address on your phone (same Wi-Fi),
enter the code once, and use Take a photo to open the camera. A link with the code,
http://192.168.1.20:8765/?code=123456, pairs in one step.
- Every device except the computer running Crosscheck must enter the code. After five wrong codes a new one is shown and pairing pauses for a minute, so guessing would take months. Paired devices get a session cookie for 30 days.
- The page uses the phone's normal camera picker, which works without HTTPS. Photos then travel
over your Wi-Fi unencrypted, so use
--lanonly on a network you trust, or add--tls-cert cert.pem --tls-key key.pemfor HTTPS (a self-signed certificate works after your phone's warning).
eval/modal_eval.py runs Crosscheck on 100 photos taken by blind people, from the VizWiz-VQA
validation set (CC BY 4.0), each with the question its photographer asked and ten crowd answers.
eval/rescore.py then grades every answer against the crowd answers with a separate text model
(Qwen3-8B), because word matching misjudges free-form answers. Results on one NVIDIA A10, all three
default models loaded (eval/results/vizwiz-judged.json):
| Result | |
|---|---|
| Correct, single model (Qwen3-VL 4B / Gemma 4 E4B / Ministral 3 3B) | 56% / 55% / 45% |
| Correct, Crosscheck's combined answer | 63% |
| Wrong when the models disagree | 45% |
| Wrong when all models agree | 33% |
| Wrong answers that were flagged as a disagreement | 38% |
| Time per photo (median / 90th percentile) | 13 s / 42 s |
What this means:
- Combining three small models gives a better answer than any one of them.
- Disagreement is a real warning sign (errors are more common), but agreement does not mean the answer is right: a third of agreed answers were still wrong. VizWiz photos are often blurry, dark or badly framed, which is exactly when blind users most need help and when small models are weakest.
- An earlier version compared answers by their words and used Granite 3.2 Vision, which refused
38% of photos. It flagged almost everything and told users nothing. Its results are kept in
eval/results/vizwiz-v1-granite-lexical.json.
The ASSETS '25 study used large cloud models. We have not yet measured how much bigger local models (for example on a 24 GB GPU) improve these numbers.
- Runs on 127.0.0.1 only unless you pass
--lan. Photos stay in memory and go only to your Ollama. Nothing is logged. - With
--lan, other devices must pair with a code shown on the computer. The exemption for the computer itself is decided by the connection's address, never by a header a client controls. - Requests must come from the app itself: a
Hostthat is not localhost or an IP address (DNS rebinding) or a foreignOrigin(another website) is refused. A strict Content-Security-Policy blocks outside content. - Model output is shown as text, never as HTML.
- Phones work over your home Wi-Fi with
--lan(see above). Tested with Chrome's phone emulation. Reports from real iPhones and Android phones with VoiceOver or TalkBack are welcome. - Merging facts across models uses word overlap, so two phrasings of one fact are sometimes listed separately, but each still gets judged.
- Small local models are weaker than the cloud models in the study, and their agreement is a weaker signal (see Evaluation). Treat "the models agree" as encouraging, never as proof.
Feedback from blind and low-vision users matters most: please open an issue or email admin@rushlabs.ai.
python -m unittest discover -s tests -vTests run against a fake Ollama: merging and judging, a failing model, the web app's security checks, the job queue and the CLI.
Apache-2.0. Built by Rush Labs, an applied AI lab. The models keep their own licences (all three defaults are Apache-2.0).