How Artist Models Work

Before training, pAInter asks artists to document four things. This creates a pilot review record. It does not independently guarantee copyright ownership of every artwork shown on any profile.

The readiness check scores your upload across several categories. A score of 60 or higher is required to start training.

Before training starts, you choose a model type, trigger word, and destination. Here is what each term means.

The artist does not need to open Replicate, handle an API token, or edit developer settings. Here is what happens in order:

1. 1pAInter validates the selected files — format, size, and minimum image count.
2. 2pAInter packages only those approved files into a training package.
3. 3The package is sent to Replicate through pAInter's server. Your API token is never exposed to the browser.
4. 4Replicate may generate short captions describing the images and associate them with the trigger word.
5. 5The base model remains unchanged. Only the smaller LoRA adjustment layer is trained.
6. 6The LoRA weights learn adjustments connected to the visual patterns in the selected artwork.
7. 7The trained LoRA and a runnable model version are saved at the destination model location.

pAInter checks status automatically. Artists should not need to use a terminal or visit Replicate directly.

A model may learn some visual patterns strongly and others poorly. Testing helps you decide whether the result is acceptable before any publishing step.

1. 1Open Test Outputs in the Artist Studio.
2. 2Write a prompt describing what you want to see.
3. 3pAInter applies the trained version and trigger word automatically.
4. 4Generate a real test image from the trained model.
5. 5Review the result — approve, generate another, or request changes.
6. 6Approval applies only to private demo testing in this prototype.
7. 7Approval does not automatically publish the model to the marketplace.

The intended long-term controls are listed below with their current status in this prototype.

Plain-language definitions for every term used in this guide.

Approved artwork

The specific files the artist selects, uploads, and confirms for use in one training job. Only approved artwork enters that job's training package.

Base model

A large pre-trained image-generation system that already exists before pAInter training begins. The current prototype uses the FLUX base model through Replicate. The base model is not created by pAInter and is not built solely from any one artist's work.

Fine-tuning

Adapting an existing model using a smaller, focused dataset. pAInter's LoRA training is a form of fine-tuning — the base model is not retrained from scratch.

LoRA

Low-Rank Adaptation. A technique that trains a smaller set of learned adjustments on top of a base model, without rebuilding the full model from scratch. LoRA stands for Low-Rank Adaptation.

Trigger word

A unique label applied during training to associate the artist adaptation with a specific text cue. pAInter uses it automatically in prompts to activate the trained concept. It is not a password or proof of ownership.

Destination model

The private location where the completed artist adaptation and its versions are saved, similar to a folder or repository. It is not an additional source of training data.

Training job

A single training run — one request to Replicate to train an adaptation from a specific set of approved files.

Training ID

A tracking number assigned to one specific training attempt, used to recover and check job status if you close or refresh the page.

Training status

The current state of a training job: Starting, Processing, Succeeded, Failed, or Canceled.

Model version

The exact trained result produced by one completed training run. Retraining creates a separate version.

Weights

The numerical values inside a model that encode what it has learned. LoRA weights are the smaller file containing the artist-specific adjustments.

Prediction

A single image-generation request made using a trained model version and a text prompt.

Test output

An image generated from a trained artist adaptation during the artist's private review period, before any marketplace publishing step.

License

The terms that define what a creator may do with images generated from an artist's model — for example, Demo, Creator, or Commercial use.

Consent receipt

A record attached to each generated image that identifies the artist, model, license type, and usage terms. In this prototype, receipts are stored in the browser session only.