Vinetas runs entirely on open-source AI models, on your own device. We think you have a right to know which models those are, what’s publicly known about how they were built, and where we stand on the ethics of training data. This page is our honest answer — including the parts the model creators haven’t disclosed, and the parts we’re still working to improve.
We’d rather tell you exactly where things stand than make a claim we can’t back up.
The two models Vinetas ships
Vinetas does not train any models of its own. It runs two existing open-source models, both released publicly with open weights and open licenses you can read for yourself.
PixArt-Σ (free tier)
- What it is: PixArt-Σ (“PixArt-Sigma”), a compact (~0.6B-parameter) diffusion-transformer text-to-image model. It’s the model every Vinetas user starts with — fast, capable, and runnable on 8 GB Apple Silicon.
- License: CreativeML Open RAIL++-M — an open license that permits commercial use, subject to use-based restrictions intended to prevent harmful applications.
- What’s public about its training data: PixArt is unusually transparent for an image model. Its published research describes training in part on images from Meta’s Segment Anything (SAM) dataset, re-captioned by a vision-language model (LLaVA). The resulting caption set, SAM-LLaVA-Captions10M, is openly published. PixArt also uses additional internal data that is not fully itemized.
FLUX.2 [klein] 4B (Vinetas Pro)
- What it is: FLUX.2 [klein] 4B, a 4-billion-parameter model distilled from Black Forest Labs’ larger FLUX.2 base model, for sharper, more prompt-faithful results. It’s unlocked by the one-time Vinetas Pro purchase and runs on 16 GB+ Apple Silicon.
- License: Apache 2.0 — a fully permissive open-source license that allows commercial use, royalty-free. (Note: the larger FLUX.2 [klein] 9B and FLUX.2 [dev] models carry a non-commercial license and are not used in Vinetas.)
- What’s public about its training data: Black Forest Labs has not published the sources of FLUX.2’s training images. What they do document is safety filtering: the pre-training data was filtered for not-safe-for-work content and for known child sexual abuse material (CSAM), the latter in partnership with the Internet Watch Foundation. That’s a meaningful harm-reduction step — but it speaks to what was removed, not to the copyright or consent status of what remains.
Where we are honest with you
We want to be precise about what “ethically sourced” can and can’t mean today:
- Neither model creator has published the full provenance of its training images. Like most large image models available today, both were trained on large-scale collections of images whose individual copyright and consent status is not disclosed. We are not in a position to claim that every image used to train these models was licensed or contributed with consent — and we won’t pretend otherwise.
- Open weights and open licenses are a real, verifiable improvement over closed, proprietary models: you can inspect the licenses, read the research, and run everything locally without anyone’s API in the loop. PixArt goes a step further by publishing a documented, openly available portion of its training data.
We see open-source models as a start — a step in the right direction — not the finish line.
What is fully ethical today: how your work is used
There’s an important distinction between how a model was trained and how it’s used. We can make an unconditional promise about the second:
- Nothing you create in Vinetas is ever used to train any model — ours or anyone else’s. Generation happens entirely on your device. Your prompts, characters, and finished boards never leave your Mac or iPad and are never collected, sold, or fed back into a training pipeline.
In an industry where “free” usually means you’re the training data, this is the part we can promise without an asterisk. See our Privacy Policy for the full detail.
Our commitment
As the open-source ecosystem matures, models trained on demonstrably licensed, public-domain, or consent-based data are beginning to appear. Our commitment is straightforward:
- Prefer provenance. When a model of comparable quality offers cleaner, better-documented training-data provenance, we will favor it.
- Stay transparent. We will keep this page current with what we actually know about every model we ship — including what we don’t know.
- Never overclaim. We will not describe our models as “ethically sourced” until that is a claim we can fully stand behind.
If you have questions about the models or their data, or you know of a higher-provenance model we should evaluate, we’d genuinely like to hear from you: support@vinetas.app.