While I do understand AI art has some good uses, creating a Scratch-appropriate image dataset on par with Llama or Dall-E (which themselves are only ok at best) performance is much easier said than done. You will still have to address copyright issues from large companies and individual artists to get any high quality art.

Gamer_Logan819 wrote:

Support. Ai art is objectively good

i_eat_coffee wrote:

If I wanted to make a platformer but horribly failed at making art for, for example, the character, an AI image generation tool would simply make everything more easy since I can just focus on the code part to make it as smooth and enjoyable as possible.

There are many dedicated people in the Collaborations and Requests forums that can help make art that a) likely better suits your prompts, and b) lets you use a feedback loop to gradually improve the end result (faster than AI image generators). If this isn't advertised enough, this should be made more advertised, and if this isn't good enough, well that's a shame because Scratch is unlikely to implement AI in this decade anyways.

i_eat_coffee wrote:

A-MARIO-PLAYER wrote:

I've even tested an AI for myself and the results are AWFUL.

I asked it to generate Scratch Cat and this is the result.

what AI model was that?
And that is actually a good thing, because it isn't infringing on Scratch's trademark.. if it knew what Scratch Cat would look like, that would be quite a lawsuit on copyrighted content

That would only be an issue if it copied every feature down to the right shade of orange, iirc.

A-MARIO-PLAYER wrote:

Generatiing AI art is lazy. Just learn how to draw, or use sprites out of the sprite library, instead.

Not everyone is entirely capable of drawing or has the resources to draw, let alone draw digital art. AI art is useful in this regard.

Edit: I meant to add this earlier but I couldn't find the post, but this is a well-written suggestion, brings up counterarguments and seems to account for TOLORS.

Not support. Now, to be fair, I don't know that much about what goes on behind the scenes in AI

i_eat_coffee wrote:

Scratch can simply finetune a stable diffusion model with their own images, which they can gather from copyright-free datasets or websites.
Resolves the two issues: inappropriate images and copyright infringement

Datasets are very big. Stable Diffusion has 14 million images and still has room for improvement. Humans would need to vet all those images. And a lot of the language part of it (how it knows what a prompt means, what it knows it can and can't combine) would have to be rewritten to be appropriate and suitable to use in all the countries and languages that use scratch.

i_eat_coffee wrote:

For the third one:
“It takes away creativity”
If I wanted to make a platformer but horribly failed at making art for, for example, the character, an AI image generation tool would simply make everything more easy since I can just focus on the code part to make it as smooth and enjoyable as possible.

I think 2 bigger issues than creativity would be:
A. It would add a lot to the electricity bill and server costs, possibly more than cloud variables. AI image generators use up to 0.29 KWH per image, if we assume the average would be about 0.25 KWH, each one would use the same amount of electricity (depending on measurement) as storing about 40 MB or 40 million letters of text in a txt file.
B. It would be controversial, because a lot of people are concerned with the ethics of AI image generation. And that's regardless of if we argue about ethics now, because we wouldn't be able to convince all the people who would be turned off Scratch by AI image generation.

i_eat_coffee wrote:

The design and overall aesthetic is way more important than the engine, generally, because of the project's first impressions.

Maybe to some, but Scratch is an educational coding website first and foremost. The social aspect is great and a vital asset, and I understand wanting a project to look good of course, but the first impression of a Scratch game outside of scratch will generally be “it's not made in a text-based language like most high-quality games”, and some things are too much work just for giving a stepping stone toward popularity on the Scratch website.

i_eat_coffee wrote:

The only other option would be to just take a character from someone else's project… but, that is literally the second main issue that everyone lists.

People who have permission from their legal guardians & abide by the T&C and copyright agreement of an AI program already have access to another option, and people who can't do that can use the Scratch library (I understand that they aren't as personalized, but it is a workaround at the base level).
The process of getting a character would be complicated too, because most AI image generation only gives rectangles with no transparency.

edward789121 wrote:

Datasets are very big. Stable Diffusion has 14 million images and still has room for improvement. Humans would need to vet all those images. And a lot of the language part of it (how it knows what a prompt means, what it knows it can and can't combine) would have to be rewritten to be appropriate and suitable to use in all the countries and languages that use scratch.

Stable diffusion is meant to generate a whole load of diverse things. It knows what a cat looks like, what a brown cat looks like, what a white cat looks like, what a TV looks like, what a house looks like, what a mushroom looks like, etc.
An AI for Scratch would only need to know how to generate characters for the project.

For the “language part of it”, no, it wouldn't – AI image generators understand prompts simply by transforming text into computer-friendly data. For different languages, etc., why can't there be a translator embedded into the process?

edward789121 wrote:

The process of getting a character would be complicated too, because most AI image generation only gives rectangles with no transparency.

No, an AI would be specifically trained to create characters with transparent backgrounds.

i_eat_coffee wrote:

It knows what a cat looks like, what a brown cat looks like, what a white cat looks like, what a TV looks like, what a house looks like, what a mushroom looks like, etc.
An AI for Scratch would only need to know how to generate characters for the project.

For the “language part of it”, no, it wouldn't – AI image generators understand prompts simply by transforming text into computer-friendly data. For different languages, etc., why can't there be a translator embedded into the process?

What I meant is a kid friendly AI image generator would have to not only use 100% clean images, but decide what words are associated with specific objects or other words in a way that suits their needs. For example, even without pictures of something like FNAF (which is not allowed on Scratch), if you enter “Five Nights Game” it might try and generate a picture of an animatronic with similar features because it knows a text description of FNAF and connects those words to eachother and then to animatronics, or if you entered something like “holes on surface” meaning surface of the earth, and it made the wrong guess about what kind of surface you meant, it could, despite not knowing what a repulsive image looks like, show a scary or repulsive image because it would still know what “holes” means and the prompt didn't have any bad words that would trigger a “unable to fulfill request” response.

i_eat_coffee wrote:

Scratch can simply finetune a stable diffusion model with their own images, which they can gather from copyright-free datasets or websites.
Resolves the two issues: inappropriate images and copyright infringement

For the third one:
“It takes away creativity”
If I wanted to make a platformer but horribly failed at making art for, for example, the character, an AI image generation tool would simply make everything more easy since I can just focus on the code part to make it as smooth and enjoyable as possible.
The design and overall aesthetic is way more important than the engine, generally, because of the project's first impressions.

The only other option would be to just take a character from someone else's project… but, that is literally the second main issue that everyone lists.

Scratch has already rejected the idea of AI art.

Unfortunately, I don't think your “solution to AI art” addresses all of the pitfalls with AI image generation, so this suggestion is still rejected.

From The Official List of Rejected Suggestions:

Za-Chary wrote:

2.10 AI image generation in Scratch
The Scratch Team has experimented with AI image generation for Scratch in the past. There are benefits for enhancing creativity, including making it easier for Scratchers to make the costumes they want for their projects, as well as inspiring Scratchers to make new stories based on the images they generate. However, there are also many downsides, including ethical concerns arising from AI images using copyrighted material, biases present in AI which reinforce harmful stereotypes, the ease of creating inappropriate content, and the difficulty of accessing AI. For these reasons, the Scratch Team has stated that they will not be adding AI image generation to Scratch anytime soon. For more information, see this article.

However, the use of AI as a virtual assistant is NOT rejected. You can discuss it on this topic. In general, AI in Scratch is not entirely rejected; it depends on the specific use of the AI.

Please read Inside Scratch Lab: AI Image Generation for more details.