Media creation with artificial intelligence

Sub pages:

• Examples gallery

Copyright and fair use

To understand the full picture of copyright, it is necessary to look at its real-world implementation.

Game companies have a strong interest in not upsetting their fan base, especially organized structures such as gaming communities.

Game communities function as voluntary support structures that provide services companies would otherwise need to fund themselves. Many of these contributions can be understood as direct or indirect sales promotion. These communities typically offer:

• First contact and general help for newbies.
• The creation of guides, tips and tricks, and even complete playthroughs.
• A space for like-minded people to socialize and exchange and develop ideas.
  • Such spaces were forums in the past, nowadays they are real-time communication platform like a Discord channel which is thematically decorated and helps to keep the product in general awareness over long time frames.
  • Communities pose a potential source for researching future products.
• Support in organizing competitions and other promotional events, including creative activities such as fan art. Cosplay contributions can increase a company's visibility at events like Gamescom.
• An additional channel for updates (general information and content announcements).
• Bug reporting and, in some cases, bug fixing. In rare instances, community members may even contribute to maintaining source code.
• Mods that improve replay value and thus increase overall customer satisfaction.
• Increased likelihood that fans will purchase other games and products (merchandise) from the company.
• A pool of trusted and engaged players who can be recruited as beta testers for new releases.

As a result, most companies also employ community managers. At least for active cash cows.

In practice, companies often tolerate limited uses of their intellectual property because they benefit from these activities. A strictly enforced copyright regime could suppress creative community contributions, reduce engagement, and ultimately harm the company itself.

However, this does not mean that copyright is overridden. In some cases, fan works may fall under fair use (depending on jurisdiction), but more often they exist within a space of informal tolerance or explicit licensing policies.

Game modifications (mods) often remain short of becoming independent games:

• Typically, they add optional 2D and 3D content, sometimes in large quantities. However, more fundamental changes - such as new game mechanics - often require access to or modification of the game engine and are therefore limited.

Generative AI (GenAI) has the potential to disrupt this cost–benefit balance:

• An automated large-scale production of new content based on existing assets may conflict with the company’s interests.

Considering the artists and technicians involved in creating the original work, mass-produced fan content generated at little or no cost could threaten established creative professions.

• An overabundance of derivative content may dilute attention and reduce consumer motivation to engage with official products.

The goal should be a form of symbiotic coexistence:

• Gaming communities should avoid creating direct competition with official products.
• In contrast to official expansion packs and DLCs, mods should generally not be placed behind paywalls.

As of 2026, conflicts arising from GenAI-driven mods remain largely hypothetical, but their relevance is likely to increase. In the long term, cooperative development models between companies and communities are conceivable, though this remains uncharted territory.

Possibilities and limitations

Generative artificial intelligence (GenAI) can ease and accelerate content creation. The difficulty of using or creating your own setups will continue to decrease with each newly released commercial forerunner model.

You can use GenAI via websites, desktop clients, or dedicated programs that may run fully locally.

To build your own programmatic solutions, you will either need downloadable AI models or API keys to access cloud services that perform the heavy computation remotely.

With sufficient expertise, you can even build agentic AIs such as Open Claw that use your existing tools and carry out tasks automatically. However, caution is advised: Probabilistic AIs can hallucinate and may pose a risk to your system. Internally, MCP are used for more safety but these don't compensate to have backup means. Sandboxes offer an additional layer of safety, but they can limit the usefulness of agents and introduce extra complexity, which may offset the time savings you intended to achieve.

GenAI systems operate probabilistically. Do not expect identical results when repeating prompts with the same inputs. The same text prompts may produce similar, but not identical, outputs. Therefore, in some scenarios, it can be beneficial to generate multiple results and select the most suitable candidates for your intermediate or final goal.

Sounds

For natural voices you may want to look for emotional text-to-speech.

Voice cloners and generators

Elevenlabs

• https://elevenlabs.io/

Music generators

Suno

• https://suno.com/create

Lyria (Google Gemini)

• https://gemini.google/overview/music-generation/

Editors

For fine-tuning on professional level you might want look at editors like Melodyn.

Images

Techniques

Creating new content.

• Text-only prompts
• Text prompts with one or multiple references
  • As default, the first image serves as main canvas that gets modified by the other references. The order of references can be overridden by pieces in the text prompt. You have to be careful in describing what reference is used for what if there are multiple.

Changing existing content

• Expanding
• Inpainting (replacement of subsections)
• Style transfers

Select best tool per use-case

Overview

Right now, Gemini seems to perform best in most use-cases.

(Add table.)

Upscaling

Very often you will prefer a generated style of one tool over the style of other tools. However, that shouldn't stop you from trying out other tools for other tasks. In the end a combination of generators can get you closer to the result you had in mind.

For instance, if you like ChatGPT for the style then maybe you still want to scale up a draft or reference image in Gemini first.

Gemini

• Free Users: Generally capped at 1K resolution.
• AI Plus/Pro Subscribers: Can access 2K resolution.
• AI Ultra Subscribers: Have full access to the 4K resolution toggle and downloads.

Workflows

Multimodal LLMs and plugins-using LLMs such as ChatGPT, Copilot, Gemini, Grok and Meta AI provide image generation capabilities.

• In addition, there are specialized tools (e.g., diffusion-based systems) that offer more control and customization. However, most users already have access to at least one of these platforms and can begin generating images immediately.
• For high-volume generation, a paid subscription or plan is typically required.

For image editing and post-processing, dedicated graphics software such as GIMP, Krita, or Photoshop is recommended. These tools allow precise control (e.g., masking, compositing, color correction) and can complement GenAI workflows.

In general you will always want to takes these steps: Generate, refine via text prompts, select final candidates, refine via graphic tools.

Beginner workflows

In context of its limitations, chatbot-based workflows are still improvements over a pure manual workflows: They speed up prototyping and let you explore different creative directions for drafts.

When you want to work with chatbots, you effectively have to learn prompt engineering: You basically learn how to write good prompts.

ChatGPT

• Iterative prompting: Describe the desired result as clearly as possible: Motive, perspective, colors, lights, shadows, art style. Refine the prompt step by step based on undesired aspects rather than expecting a perfect result on the first attempt. Negative prompts: You can also explicitly write what you don't want.
  • Caution: As for March 2026, you have explicitly tell ChateGPT to not crop the image in some scenarios. In that case you have to give it context like where the image should end and tell it how much "extra" space you want. Saying "don't crop the image" is to not specific enough.
• Avoid quality loss: If the GenAI degenerates the image quality because of too many iterations, try from a new start with combined text prompts.
• Reference images: When supported, provide one or more images to guide style, composition, or subject consistency. This is often more reliable than text-only prompting.
• Context management: If previous prompts begin to overly influence results, start a new prompt and explicitly restate the desired outcome. This prevents unintended bias from earlier context.
• Merging / composition: Supplying multiple images in a single prompt can help combine elements. However, repeated re-editing of generated outputs may degrade detail or introduce artifacts.
• Batch generation: Since outputs are probabilistic, generate multiple final candidates and select the best.
• Post-processing workflow: Combine the best elements using external tools (e.g., masking in Photoshop or Krita). This hybrid approach often yields higher-quality results than relying on a single generation.

(Add examples here.)

Paint (Cocreator)

With Windows 11 and 40 TOPS minimum you can use Microsoft Paint with its Cocreator module. The Cocreator is sometimes also named Image Creator. (The Tera Operations Per Second is usually referring to INT8 operations on AI accelerator hardware, NPUs.) Windows PCs that have the naming tag Copilot+ are safe to assume to have that feature.

• You write a prompt, optionally select a style and then draw a draft that gets updated almost in real-time in a secondary panel.

Photoshop (Adobe Firefly)

Photoshop has a build-in image generator which can use Adobe's own Firefly model as well as other ones.

In Photoshop you can prompt images and immediately start editing them. Or you chose expand them first or do partial replacements.

Advanced workflows

AUTOMATIC1111 (aka Stable diffusion)

ComfyUI

Expert workflows

This would include to train own models. The idea is to let the models have a neural representation of objects that equals to screenshot-taking from 3D so that prompts will output almost never hallucinate details. That way artist can reduce post-editing as the generated outputs also include his own style.

Own models once more boost rapid prototyping because they reduce the necessity to have a more complexer, combined 3D-2D-workflow.

Videos

Sora (OpenAI)

• In 2026, Sora was announced to be discontinued. It will be probably just paused for a few years until future AI chips have lowered computation costs.

3D content

There exists content generators that turn 2D data into 3D data by calculating plausible assumptions for the missing dimension.

3D objects

• ...

3D animations

• ...

World maps

• ...
• World generator inside Unreal Engine 5
  • As for 2026, this is technically speaking still "procedural" but it is plausible to expect an LLM-driven approach in the future. On the internet you can find already LLM-driven experiments.