AI Art Generators: Buying and Selling Guide

Most AI art generator listings are a thin interface over someone else's model, and the price often reflects the interface rather than what you actually acquire. The buyer's job is to establish which parts of the stack transfer to them and which remain rented.

The Five Things Sold Under One Label

A prompt library — curated prompts with example outputs. No software. Value is entirely in whether the prompts still work on the current model version, which changes.

A hosted app with no source — you get a login. You own nothing, and you depend on the seller keeping it running. Rarely worth much.

An app with source code — a wrapper around Stable Diffusion, an API, or a hosted endpoint. Now you own something, provided the dependencies are yours.

A custom-trained model — a fine-tune or LoRA producing a specific style. The genuinely differentiated product in this category.

A full SaaS platform — accounts, billing, gallery, generation queue. A real software business, priced accordingly.

The Question That Decides Value

Where does inference actually happen, and who pays for it?

Three possibilities, with very different economics:

The buyer's own GPU. Self-hosted Stable Diffusion or similar. No per-image cost, but you need the hardware — realistically 8GB+ VRAM for SDXL at reasonable speed.

A third-party API. The app calls OpenAI, Replicate, Fal, or similar. Every image costs money. Ask for the per-image cost and the current monthly bill. An app generating 10,000 images monthly at $0.04 each is a $400/month obligation you are inheriting.

The seller's infrastructure. The worst case. If generation runs through the seller's account or server, you have bought a dependency on someone who has just stopped being invested in it.

Establish this before discussing price. It is the difference between an asset and a subscription.

Licensing — Where the Real Risk Sits

The base model's licence governs everything downstream. Stable Diffusion variants ship under licences (CreativeML OpenRAIL-M and others) with use restrictions that follow the weights. A fine-tune inherits them. The seller cannot grant you broader rights than they hold.

Commercial rights to generated images are separate from the right to use the software. Confirm both.

Training-data provenance. A model fine-tuned on a named living artist's work to reproduce their style is a legal and reputational liability, whatever the listing says. Ask what it was trained on. If the answer is evasive, that is the answer.

Trademarked outputs. A model reliably producing recognisable branded characters is a problem you would be buying.

What to Verify Before Confirming

• Generate at least twenty images yourself, across varied prompts — not the seller's cherry-picked demo set. Consistency across prompts is what separates a real fine-tune from a base model with a preset.
• Confirm the model files are included, not downloaded at runtime from a URL the seller controls.
• Check hardware requirements on your actual hardware.
• Trace every API dependency and its cost.
• Read the licence chain, base model through to the listing.
• Confirm source code is complete and runs — see the standing advice on running any purchased software yourself before confirming receipt.

Realistic Pricing

• Prompt library: $20–$100
• Hosted app, no source: $100–$500 (and be sceptical)
• App with full source: $500–$3,000
• Custom-trained model or LoRA: $1,000–$5,000
• Full SaaS platform with users: $5,000–$20,000+

A platform with paying subscribers is valued on revenue, not on code. Ask for churn and the payment-processor history, and remember Stripe accounts do not transfer.

Selling These Tools Well

State plainly what is included: source, model weights, documentation, and whether inference is self-hosted or API-based. Publish hardware requirements. Give real API costs — hiding them guarantees a dispute. Provide a demo the buyer can drive themselves rather than a gallery of your best outputs. Name the base model and its licence.

Sellers who disclose the running costs up front close more deals, because the buyers who proceed have already accepted them.

Verifying Under Escrow

Escrow is particularly useful here because the failure modes are discoverable in minutes if you look. On Escrozon funds are held in escrow while you check, so generate your own images, run the code on your own hardware, and total up the API costs before releasing anything. If the tool only works through the seller's infrastructure and that was not disclosed, that is a material misrepresentation and grounds for a dispute.

Frequently Asked Questions

Can I sell images generated by a model I bought? Usually, but it depends on the base model licence and the terms of any API involved. Confirm both. Copyright status of AI-generated images also varies by jurisdiction — in the US, purely AI-generated work has generally not been granted copyright protection.

What GPU do I need to self-host? For SDXL, 8GB VRAM is a practical minimum and 12GB+ is comfortable. Quantised models run on less, with some quality cost.

Is a LoRA worth as much as a full fine-tune? Often more, practically. LoRAs are small, portable, and composable. Judge on output consistency, not file size.

The seller says it uses "a proprietary model." What should I ask? Ask what base architecture it derives from and to see the weights. Almost everything in this space is derived from a known open model. "Proprietary" with no detail usually means a base model with a prompt preset.

How do I check a fine-tune is genuinely trained and not a preset? Run the same prompts against the stated base model and compare. If outputs are indistinguishable, you are being sold a prompt template at fine-tune prices.