'/%3E%3C/svg%3E){kind=small}
Education Center Talent.pl
01 — Introduction
Why this guide
Image generation by artificial intelligence is one of the fastest-growing fields of modern technology. In just a few years, it has gone from laboratory experiments to tools used on a mass scale by artists, designers, advertising agencies and brands around the world.
The Talent.pl platform was created for professional creators, hobbyists and companies that want to use the achievements of generative AI in an informed, safe and legally compliant way. This educational material aims to reliably explain the technical mechanisms of image generation, show the capabilities and limitations of AI models and — most importantly – discuss the licensing issues and legal frameworks applicable in Poland and the European Union.
Understanding these issues is essential for anyone who wants to professionally commission, sell or commercially use AI-generated images. A lack of this knowledge can lead to unknowingly infringing the rights of third parties or to purchasing a licence that is unsuitable for the intended use.
Who this guide is for
02 — Technology
How AI image generation works
Before we move on to legal issues, it is worth understanding how AI algorithms “create” images. This knowledge has a direct bearing on legal assessment – in particular on questions of authorship, originality and rights to training data.
2.1. Training data – the foundation of the model
Legal issue
From a legal point of view, the legality of training data is crucial. Questions arise as to whether using copyrighted images to train models constitutes permitted use (fair use / scientific quotation) or copyright infringement. Lawsuits against Stability AI, Midjourney and DeviantArt in the USA and the ongoing debate in Europe show that the issue has not yet been definitively resolved.
2.2 Diffusion Models
Every generative AI model must first go through a training phase. During this process, the algorithm is exposed to huge datasets — usually billions of (image, text description) pairs — collected from the Internet, photo databases, scans of artworks and other sources. The model does not store these images directly; instead, it learns statistical distributions: what a face looks like, how perspective works, how colours match objects.
The dominant architecture used in current systems (Stable Diffusion, DALL·E 3, Midjourney, Firefly) is latent diffusion models (Latent Diffusion Models, LDM).
How it works step by step
- The model creates a “noise image” — a random set of pixels with no visual content
- The encoder converts the image into a representation in latent space — a dense, low-dimensional feature map
- The diffusion algorithm iteratively removes noise over many steps (usually 20–50), gradually revealing a coherent image
- The key element is conditioning: at each step, an encoded representation of the text description (prompt) is added, which “pulls” the image towards the desired content
- The decoder transforms the latent representation back into a raster image at the target resolution
The whole process usually takes from a few seconds to several dozen seconds on a modern GPU. The result is a digital image that statistically matches the given description but is never a simple copy of any specific image from the training data — at least in theory.
2.3 GAN models
The previous generation of dominant models were Generative Adversarial Networks (GAN), proposed by Goodfellow et al. in 2014. GANs consist of two neural networks competing with each other:
- Generator — creates fake images, trying to “fool” the discriminator
- Discriminator — assesses whether a given image comes from the training data or from the generator
Both networks learn simultaneously in a zero-sum game until the generator starts producing images indistinguishable from real ones. GANs were a breakthrough in generating realistic faces (StyleGAN), but their training is unstable and difficult to control. Today, diffusion models have largely replaced GANs in commercial applications.
2.4 Transformer architecture and language models
In modern image generation systems, language models act as a “translator” between text and visual space:
- CLIP (Contrastive Language–Image Pretraining) by OpenAI — learns to embed text and images simultaneously in a shared vector space
- T5, BERT — text encoders used to embed prompts in many diffusion models
- GPT-4V, Claude — multimodal models capable of both understanding and generating visual content
In practice, a modern image generation system is a complex pipeline of multiple models, not a single algorithm.
2.5 Technical glossary
Prompt
A text (sometimes graphic) input description for the model.
Negative Prompt
A description of what should NOT appear in the image.
Steps / iterations
The number of diffusion steps; more = slower, but usually better quality.
CFG scale
The strength of prompt conditioning (how closely the model “listens” to the description).
Seed
A random seed; the same seed + prompt = the same result.
LoRA / Fine-tuning
Fine-tuning a base model to a specific style or brand.
Inpainting
Filling in or editing a selected part of the image by the model.
Outpainting
Extending an image beyond its original boundaries.
img2img
Generating a new image using an existing one as a starting point.
ControlNet
A module for controlling the pose of figures or the layout of objects using skeletons or edges.
03 — Practice
Capabilities and limitations
3.1 What AI can do today
- Photorealistic portraits, outdoor scenes and studio-quality product compositions
- Rapid concept iteration — a designer generates dozens of variants within minutes
- Styling — consistent styling of an entire set of graphics (branding, advertising campaign)
- Personalisation — fine-tuning on specific models, products or brand style
- Scaling without loss of quality — modern models generate images in 2K–4K resolution
- Generating variants and adaptations — quickly creating language, seasonal and format versions
3.2 Limitations and technical risks
- Number of fingers and hands — models often generate the wrong number of fingers
- Text in images — coherent, legible text embedded in an image remains a difficult problem
- Character consistency — keeping the same character across multiple shots without fine-tuning is difficult
- Conceptual hallucinations — the model may generate something logically absurd
- No deterministic control — results may differ even with the same parameters
- Risk of similarity to protected works — the model may generate images close to the style of a specific artist
The most important distinction
Buying an AI-generated image is not the same as buying the rights to that image. The distinction between the file, the licence and the copyright is fundamental for anyone planning to use AI content commercially.
04 — Licences
AI image licensing
4.1 What is a licence — the basics
A licence is a contractual permission granted by an authorised party (the licensor) to another person (the licensee) to use a given asset in a specified manner, for a specified time and scope. In the digital world, a licence does not transfer ownership or copyright — unless the agreement expressly provides otherwise.
Important distinction
BUYING A FILE ≠ BUYING A LICENCE ≠ ACQUIRING COPYRIGHT
By buying a file (download), you only obtain a data carrier. By buying a licence, you obtain the right to use the work under specified conditions. Acquiring copyright requires a separate agreement transferring the author’s economic rights (in Poland: in writing, otherwise null and void — Art. 53 of the Copyright Act).
4.2 Talent.pl three-tier licensing model
|
Feature |
Standard licence |
Exclusive rights / transfer |
|---|---|---|
|
Intended use |
Portfolio, social media, personal use |
Advertising, print, products / Full assignment of rights |
|
Exclusivity |
Non-exclusive (multiple buyers) |
Exclusive (single buyer) / Full |
|
Territory |
Poland |
Worldwide |
|
Duration |
Perpetual |
Perpetual |
|
Sublicensing |
Not permitted |
Permitted |
|
Right to modify |
Limited |
Full |
|
Removal from sale |
No |
Yes |
|
Chain of title documentation |
Basic |
Extended / Full notarised assignment |
4.3 AI model licences vs. licences for images sold
Creators working on Talent.pl must understand that an image they sell is subject to two separate sets of rules at the same time: the licence terms of the AI model (e.g. Stable Diffusion: CreativeML Open RAIL-M; Midjourney: TOS Pro; Adobe Firefly: CC) and the terms set by the creator and the Talent.pl platform. A conflict between them may mean that even the best Talent.pl licence will not be effective.
Model compatibility with commercial use
|
Status |
Model |
Licence terms |
|---|---|---|
|
YES |
Stable Diffusion (SDXL) |
CreativeML Open RAIL+SD |
|
YES |
DALL·E 3 (OpenAI API) |
Active subscription |
|
YES |
Adobe Firefly |
Creative Cloud · full indemnity |
|
YES |
Flux.1-schnell |
Apache 2.0 |
|
CONDITIONAL |
Midjourney Pro / Mega |
Basic plan limited |
|
CONDITIONAL |
Flux.1-dev |
Licence from Black Forest Labs |
|
NO |
Midjourney Free / Basic |
Commercial use prohibited |
4.4 Creative Commons licences and AI images
Some creators label their works with Creative Commons licences. For AI images, they have the following implications:
- CC BY 4.0 — any use (including commercial) provided the author is credited
- CC BY-NC 4.0 — non-commercial use only; not suitable for business users
- CC BY-SA 4.0 — commercial use permitted, but derivative works must be under the same licence (copyleft)
- CC0 (public domain) — the creator waives all rights; the broadest possible use
Note
A CC licence on an AI image does not solve the problem of rights to training data. If the model was trained on protected images, potential claims may still exist regardless of the licence granted by the creator.
05 — Law
Legal framework — Poland and the European Union
The legal status of AI-generated images is one of the hottest and fastest-evolving areas of intellectual property law in the world. Below we present the current state of play (May 2026), with a focus on the regulations in force in Poland and the EU.
5.1 Copyright — the question of who can be an author
Legal basis
Polish Act of 4 February 1994 on Copyright and Related Rights (Art. 1, Art. 8).
The Polish Copyright Act protects “any manifestation of creative activity of an individual nature, fixed in any form”. The holder of copyright can only be a human being (a natural person) or — to a limited extent — legal persons. AI is not and cannot be an author within the meaning of Polish law.
This means that a purely algorithmically generated image, without human creative input, may not be protected by copyright and fall into the public domain.
When a human can be considered the author of an AI image
- They created a detailed, original prompt expressing their individual artistic vision
- They selected and edited the results — choosing from many variants the one with the highest creative value
- They added original graphic elements or retouched the image to create a derivative work
- They fine-tuned the model on their own original training data
- They documented the entire process, demonstrating the “imprint of the human mind”
5.2. Artificial Intelligence Act (AI Act) — classification of image generation systems
Legal basis
Regulation (EU) 2024/1689 of 13 June 2024 (Artificial Intelligence Act).
The EU Artificial Intelligence Regulation (AI Act), fully applicable from August 2026, introduces a risk-based classification approach. Image generation tools are classified as general-purpose AI (GPAI) systems:
- Models such as GPT-4V, Gemini, DALL·E or Stable Diffusion XL are GPAI models (General-Purpose AI Models)
- Models with particularly high capabilities (>10²⁵ training FLOPS) are subject to additional transparency requirements
- Providers of GPAI models must disclose a summary of their training data, including information on copyrighted content (Art. 53 AI Act)
For end users, the transparency provisions are key: all synthetic content (image, video, audio) generated by AI must be appropriately labelled as AI output, especially when it could be misleading as to its authenticity.
5.3. Deepfakes and protection of image rights
- Art. 16 of the Artificial Intelligence Act — generated content depicting real people must be labelled as synthetic
- Civil law (Art. 24 of the Civil Code and Art. 81 of the Copyright Act) — protection of image; generating realistic portraits of real people without their consent may constitute an infringement
- Criminal Code (Art. 212) — defamation, including by means of fake AI-generated images
Talent.pl rule
On Talent.pl, it is prohibited to offer for sale images depicting identifiable real people without the documented consent of those people.
5.4 GDPR and personal data in the generation process
Legal basis
Regulation (EU) 2016/679 of 27 April 2016 (GDPR) — Art. 6 and 9.
AI image generation may involve the processing of personal data at several levels:
- Training data — if the model was trained on photographs of natural persons, this may give rise to GDPR obligations for the model provider
- Input prompt — if it contains a first name, surname or other data identifying a person
- Generated image — if it allows a natural person to be identified (e.g. a deepfake)
Talent.pl recommends that users avoid including personal data in prompts unless they have an appropriate legal basis.
5.5 DSA — platform liability
Legal basis
Regulation (EU) 2022/2065 of 19 October 2022 on a Single Market for Digital Services (“DSA”).
Under the Digital Services Act, Talent.pl, as an intermediary platform, is obliged to:
- Have a mechanism for reporting illegal content and removing it without undue delay
- Provide transparent terms of use describing the rules for AI content
- Keep a register of removed content for at least 6 months
The platform is not liable for content listed by creators, provided that it acted promptly upon obtaining knowledge of its illegal nature (safe harbour).
5.6 Copyright infringement
A creator using AI models trained on protected data may generate images that are too similar to a specific copyrighted work. The law provides for the following remedies:
- A demand to cease the infringement and remove the content
- Damages amounting to two or three times the appropriate remuneration (Art. 79 of the Copyright Act)
- Claims for infringement of moral rights (the right to the integrity of the work, the right to attribution of authorship)
The best protection is to use models based on licensed data (e.g. Adobe Firefly) or to have full documentation of the image creation process.
06 — Practice
Best practices
6.1 For creators — how to document the creative process
- Save the complete prompt (positive and negative) and the seed value for each final image
- Note the model, version and provider of the AI tool, as well as the subscription plan (it affects commercial rights)
- Keep logs or screenshots from the generation platform (date, time, parameters)
- If you used fine-tuning on your own data — keep documentation of the training data
- Describe your own creative contribution: selection, composition, retouching, inpainting, compositing with other elements
- Upload the documentation to the Talent.pl system — the platform stores it securely and makes it available to the buyer when exclusive licences are purchased
6.2 For buyers — how to check an image before buying
- Check which model the creator used and whether the model’s licence terms permit the intended use
- Verify the available licence options (non-exclusive / exclusive / transfer of rights)
- When buying an exclusive licence or a transfer of rights — ask to see the documentation of the creative process
- If you used fine-tuning on your own data — keep documentation of the training data
- For images used in advertising or product packaging — consult an IP lawyer
- For large campaigns, consider an indemnity clause in the agreement with the creator
6.3 Red flags — what to avoid
- An image generated on a model with a non-commercial plan (e.g. Midjourney Basic) — no right to sell
- A portrait of a recognisable person without documented consent — risk of infringing image rights
- An image with elements referring to trademarks or logos — risk of infringing IP rights
- No way to obtain documentation of the creative process from the creator
- An image labelled as generated by a model that prohibits specific uses
07 — The future
What’s next for AI and law
7.1 Technology trends
- Multimodal models — seamless transitions between text, image, video and audio in a single pipeline
- Real-time video generation — Sora, Runway Gen-3 and Kling are approaching production standards
- Precise control of style and characters — IP-Adapter, PhotoMaker, InstantID without the need for fine-tuning
- 3D and avatar generation — DreamFusion and ZeroShape create 3D assets from a text description
- AI in augmented reality — real-time AR content generation
7.2 Upcoming legal regulations
- Full entry into force of the Artificial Intelligence Act (AI Act) (August 2026)— obligations for GPAI providers, publication of training data
- DSM Directive (2019/790) — its implementation in Polish law shapes the opt-out rules for creators with regard to AI training
- A potential AI act in Poland — the Ministry of Development and Technology (MRiT) is holding consultations on a national act supplementing the AI Act
- Reforms of the copyright registration system in the EU — discussion of a register of AI works in the public domain
- Case law — rulings in Getty Images v. Stability AI and Andersen v. Midjourney will shape practice for many years
7.3 The role of Talent.pl in the ecosystem
- Regular updates to the licensing policy in response to regulatory changes
- Cooperation with law firms specialising in AI and IP law
- Creator documentation audit programme — verification of the chain of title for exclusive licences
- Documentation API — integration with the legal and asset management systems of B2B clients
08 — summary
10 rules for users
AI image generation combines advanced technological achievements with a number of legal issues that are not yet fully regulated. Below are the key rules for Talent.pl users:
- Understand the model
every AI tool has its own licence terms, which determine the rights to the images. - Document the process
keep the prompt, seed, model and version for each final image. - Match the licence to the use
do not use a non-exclusive licence for a campaign that requires exclusivity. - Be careful with people’s likenesses
generating portraits without the consent of the person portrayed creates legal risk. - Check the model’s training data
models trained on licensed data (Adobe Firefly) provide greater certainty. - AI is not an author
only a human can be an author under Polish law; document your contribution. - Label AI content
the labelling requirement is an ethical standard and will soon be a legal one (AI Act). - Do not combine incompatible licences
e.g. Creative Commons NC with commercial use. - For large campaigns
consult a lawyer and consider an indemnity clause in the agreement with the creator. - Follow legal changes
in particular, the Artificial Intelligence Act (AI Act), case law and amendments to the Copyright Act will change the rules of the game.
The Talent.pl platform is committed to building an ecosystem in which AI-assisted creativity can be both innovative and legally safe. Feel free to contact our support team and legal centre if you have any questions about a specific use case.
