Media creation with artificial intelligence: Difference between revisions

'''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.

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

For professional-level fine-tuning you might want look at editors like [https://www.celemony.com/en/melodyne/what-is-melodyne Melodyn].

[https://www.nvidia.com/en-us/glossary/multimodal-large-language-models/ Multimodal LLM]s 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 image editing and post-processing, dedicated graphics software such as [https://www.gimp.org/downloads/ GIMP], [https://krita.org Krita], or [https://www.adobe.com/products/photoshop.html Photoshop] is recommended. These tools allow precise control (e.g., masking, compositing, color correction) and can complement GenAI workflows.

* '''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.  

* '''Avoid quality loss''': If the GenAI degenerates the image quality because of too many iterations, try from a new start with combined text prompts.

* '''Consistency strategies''': When available, use features such as seeds, style references, or controlled variations to maintain visual coherence across multiple images.-->

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.

'''Google Veo'''

'''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 objects===

* ...

===3D animations===

* ...

* ...

* 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.