Transparency, in plain language
How Artist Models Work
A step-by-step guide to what you upload, what pAInter trains, what every technical term means, and what remains part of the prototype.
About 8 minutes · No technical background needed
The most important thing to know
Your uploaded artwork does not create a complete AI model by itself. It creates an artist-specific adaptation that works together with a pre-trained base model. Both influence every image that is generated.
You are adapting a model, not building one from zero
How the base model and artist LoRA work together
A base model is a large general image-generation system that already exists before pAInter training begins. The current prototype uses the FLUX base model through Replicate. pAInter did not create the base model and does not control its original training data.
When you train through pAInter, a much smaller LoRA (Low-Rank Adaptation) is created from only the artwork you select and approve. The LoRA learns visual adjustments specific to your work. It does not replace the base model — it runs on top of it.
Every generated image is influenced by all three: the base model, the LoRA, and the creator's prompt. You cannot separate their contributions after generation.
How every test image is produced
Pre-trained base model
FLUX · not created by pAInter
Artist LoRA
trained on your approved artwork
Creator prompt
text and generation settings
Test image
influenced by all three
Confirm who you are and what you are allowed to use
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.
Artist account details
Your public artist name, account email, and country or region.
Control of a public artist profile
A short authorization code added to your profile so a reviewer can confirm you control that account — not that every artwork on it is authorized for training.
Rights over the selected training artwork
A declaration of the rights you hold over each work you are submitting for training: self-created, commissioned, or with written permission.
Supporting creation evidence
One file (such as a source file or progress scan) that helps demonstrate authorship. Only the file name and size are recorded — no file contents are stored.
You choose the artwork used for the artist adaptation
Only the files you select and upload in Artwork Upload enter that training job's package. pAInter does not automatically download work from your public profile or any other source.
Artwork not selected and uploaded is not intentionally included by pAInter in the artist-specific training package.
You must have the necessary rights or permission for every file you upload. Commissioned, collaborative, licensed, third-party, and AI-generated elements must be disclosed in Verification & Ownership.
Check whether the selected artwork is ready
The readiness check scores your upload across several categories. A score of 60 or higher is required to start training.
Number of approved images
Minimum 10 images required. More variety and count tend to improve pattern learning.
Supported formats
JPG, JPEG, PNG, and WebP are accepted. Other formats are not.
Useful resolution
Recommended 512 × 512 px or larger. Very small images may limit the detail the model can learn.
Ownership and permission declarations
The Verification & Ownership section must be completed before training can proceed.
Safety confirmations
Confirmations that no uploaded images contain copyrighted characters, scraped content, or another artist's work.
Licensing setup
At least one license tier must be set in Licensing & Terms before training.
Give the model a clear setup
Before training starts, you choose a model type, trigger word, and destination. Here is what each term means.
Style model
Learns recurring visual patterns — such as color palette, texture, line quality, rendering style, and composition — from varied artworks. Best for capturing an overall aesthetic.
Subject model
Learns a repeated person, character, object, or other specific subject shown consistently across training images. Best for reproducing a specific figure or design.
Trigger word
A unique label associated with the artist adaptation during training. pAInter uses it in prompts to activate the trained concept. It is not a password, ownership certificate, or public artist name. Example: PAINTER_CHERI_V1
Destination model
The private model location where the completed adaptation and its versions are saved. Think of it as the model's folder or home. It is not another source of training artwork.
Private test model
A trained result available for artist review inside the Artist Studio before any separate publishing step. Test results are not automatically public.
What pAInter and Replicate do during training
The artist does not need to open Replicate, handle an API token, or edit developer settings. Here is what happens in order:
- 1pAInter validates the selected files — format, size, and minimum image count.
- 2pAInter packages only those approved files into a training package.
- 3The package is sent to Replicate through pAInter's server. Your API token is never exposed to the browser.
- 4Replicate may generate short captions describing the images and associate them with the trigger word.
- 5The base model remains unchanged. Only the smaller LoRA adjustment layer is trained.
- 6The LoRA weights learn adjustments connected to the visual patterns in the selected artwork.
- 7The trained LoRA and a runnable model version are saved at the destination model location.
What each training status means
Your request has been accepted and the training system is preparing the job.
The artist adaptation is actively being trained.
Training finished and a model version is available for testing.
Training stopped because an error occurred. No successful model version was created from that attempt.
The training job was intentionally stopped.
pAInter checks status automatically. Artists should not need to use a terminal or visit Replicate directly.
What you receive after training
Training ID
A tracking number for one specific training attempt. It helps pAInter recover the job after a refresh or page close.
Destination model
The model location where the trained result is saved. Not an additional source of training artwork.
Model version
The exact trained result from one completed run. Retraining may produce a separate version.
LoRA weights
The smaller file containing the artist-specific learned adjustments.
Trigger word
The label pAInter applies to prompts when using this adaptation, to activate the trained concept.
Completed time
When Replicate finished processing the training job.
The artist reviews the trained result
A model may learn some visual patterns strongly and others poorly. Testing helps you decide whether the result is acceptable before any publishing step.
- 1Open Test Outputs in the Artist Studio.
- 2Write a prompt describing what you want to see.
- 3pAInter applies the trained version and trigger word automatically.
- 4Generate a real test image from the trained model.
- 5Review the result — approve, generate another, or request changes.
- 6Approval applies only to private demo testing in this prototype.
- 7Approval does not automatically publish the model to the marketplace.
Set what people may do
Licensing & Terms is where you set what creators can and cannot do with images generated from your model.
Demo
Limited preview use. Available to free users.
Creator
Personal, social, and project use by subscription creators.
Commercial
Business or commercial use where the artist enables it.
You also set restricted categories — for example, excluding adult content, political ads, or NFTs. Creators see these terms before generating.
Keep control after training
The intended long-term controls are listed below with their current status in this prototype.
What works today — and what does not yet
- Selected artwork upload (JPG, PNG, WebP, up to 10 MB each)
- Real Replicate training run from uploaded artwork only
- Automatic training-status polling (every 12 seconds)
- Real trained model version produced by Replicate
- Real test-image generation from the trained version
- Artist test-output approval and retrain flow
- Training ID recovery and status check
- Account and portfolio-control submission (saved to this browser only)
- Artwork-rights declarations and evidence preparation (local only)
- Licensing and restriction settings (saved to this browser only)
- Readiness score calculation (local, not server-validated)
- Sample consent receipt (session only, not permanently stored)
- Publish checklist display (no live marketplace connection yet)
- Secure permanent artwork storage
- Full account authentication
- Production identity verification
- Automatic public marketplace publishing
- Real payment collection from creators
- Artist payouts
- Permanent consent receipt storage across devices
- Cross-device synchronization
- Complete production model-deletion workflow
- Automated prompt restriction enforcement
Glossary
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.
After generation
What happens after a creator generates?
Generation is one part of the creator journey on pAInter. Creators can also continue into Co-Create — a prototype commission workflow that connects the generated idea directly to you as the artist.
Generated-image rights and commissioned-artwork rights are separate.
The license attached to a generated image applies only to that image. When a creator commissions artwork, you confirm the terms, rights, and process separately.
You decide whether to accept generated references.
Co-Create lets you set your own commission availability. You choose whether to accept generated images as references and what services you offer.
The current request flow is a prototype.
Co-Create requests are stored in the creator's browser only. No request is transmitted to you or saved to a server in the current demo.