Talk:Stable Diffusion

Commons files used on this page or its Wikidata item have been nominated for deletion

The following Wikimedia Commons files used on this page or its Wikidata item have been nominated for deletion:

• Algorithmically-generated landscape artwork of forest with Shinto shrine using negative prompt for green trees.png
• Algorithmically-generated landscape artwork of forest with Shinto shrine using negative prompt for round stones.png
• Algorithmically-generated landscape artwork of forest with Shinto shrine.png
• X-Y plot of algorithmically-generated AI art of European-style castle in Japan demonstrating DDIM diffusion steps.png

Participate in the deletion discussion at the nomination page. —Community Tech bot (talk) 13:37, 27 December 2022 (UTC)

Model List

Maybe we should add stable Diffusion model list to article 188.162.39.131 (talk) 14:02, 14 April 2023 (UTC)

Question regarding the "squint your eyes" image generation/diffusion

Recently, I have created a few diffusions based off the "squint your eyes" meme trend from a year or so ago, and I was wondering if I could upload them for the use of the article. I do want to consider having a "fair use" rationale, but at the same time since it's AI work it doesn't show originality and/or authorship being in the public domain. I could consider putting a CC-BY-SA 3.0 license instead, but idk what works the best here. If you think the "creation" of mine should be in the article for explanatory purposes and what license fits, please use my talk page or reply here. Thank you!

Note: The website(s) I primarily used are: https://illusiondiffusion.net/ and https://huggingface.co/spaces/AP123/IllusionDiffusion. Ｍｏｄ ｃｒｅａｔｏｒ 🏡 🗨 📝 05:17, 26 May 2024 (UTC)

This does not sound like an improvement to the article. The article already has too many unnecessary, unsourced, non-notable, user-generated images in it. Elspea756 (talk) 15:49, 26 May 2024 (UTC)

Visual compositions are subject to copyright

The "Usage and controversy section" states:

"As visual styles and compositions are not subject to copyright, it is often interpreted that users of Stable Diffusion who generate images of artworks should not be considered to be infringing upon the copyright of visually similar works."

However, according to the US Government Copyright website:

"When it comes to copyright, creativity can be demonstrated in a variety of ways and reflects artistic choices like the subject matter, composition, depiction, and the use of the elements of design."

As an example, one might employ the stable diffusion img2img script with a low denoising factor of 0.3 to modify a copyrighted work by incorporating new visual elements while retaining the original composition. The resulting image would become a derivative work thereby infringing on the original copyright.

A review of the relevant literature reveals numerous other references that demonstrate the subject of copyright is also applicable to visual compositions. In light of this, it is recommended that this section be revised to avoid potential litigation against users and Stability AI. 72.224.151.92 (talk) 18:14, 6 June 2024 (UTC)

Yes, this looks like it needs improvement. The source used for that is from Japan. It is unclear what country's copyright laws they are trying to describe. The source is "Automaton Media," and does not seem like the most reliable source and does not attribute its legal analysis to any legal expert, so I don't believe we should be using it. For now, I've switched the source to the English language version rather than the Japanese version, but this likely still needs improvement with far better sourcing than was used here. Elspea756 (talk) 18:48, 6 June 2024 (UTC)

I've just removed that sentence as being poorly sourced for legal opinions, and for being unclear what country's or countries' copyright laws they were trying give opinions about. It seems pretty clearly incorrect for U.S. copyright law, where many of Wikipedia's readers will be interested in. Thank you for pointing this out. Elspea756 (talk) 19:43, 6 June 2024 (UTC)

"First, copyright protects original works of authorship, including original pictorial, graphic, and sculptural artwork. A work is original if it is independently created and sufficiently creative." is the primer for your quote from the USCO. The full quote does not imply composition is subject to copyright but rather composition is a component of originality which is required for protected status. What you are saying would only be correct with auditory copyright as in the composition of a song.

Your example is speculation and, while it appears the sources referenced by Elspea may be inappropriate, the core idea is simply common sense. I do not wish to completely revert that change but it is incorrect.

You cannot copyright style. You cannot copyright composition. Your likeness is protected. Sojoelous (talk) 06:20, 7 June 2024 (UTC)

Should common GUIs be mentioned in the article?

I previously added this sentence to the article in the Capabilities section:

>Several graphical user interfaces exist that use Stable Diffusion as their base model to generate images, ranging from simple to use programs like Fooocus to node-based ones like ComfyUI.

It was reverted by User:MrOllie who argued that we should avoid promoting or mentioning particular UIs here. I would like some feedback from other contributes about this issue. I agree that we shouldn't pollute the SD page with too much info on GUIs, especially since there are so many, which could cause WP:UNDUE issues if we don't mention all of them. At the same time, Stable diffusion is generally not used via command line, and mentioning only the more common GUIs that have coverage by reliable sources should make the list manageable. We can also limit it to open source software to reduce the issue of WP:PROMOTION. Also, I feel that these GUIs are not fit to be mentioned anywhere else, making them orphans assuming they have their own page. It would be useful if there is some precedent for this type of situation. The closest I can think of is mentioning the individual implementations of codec specifications, like in JPEG XL#Codec implementations, although I wouldn't say that case is exactly the same. J2UDY7r00CRjH (talk) 17:51, 10 July 2024 (UTC)

Are there any GUIs which already have an article in WP? I think such GUIs could be mentioned. What about web applications (like Open WebUI for example) which support SD? They should be treated like locally installed GUIs. So my suggestion is to mention only open source GUIs and web applications which either already have an article or which are very common for this. --Stefan Weil (talk) 18:45, 10 July 2024 (UTC)

Yeah, Fooocus does. Disclaimer: I created that article recently and have zero affiliation with Fooocus. J2UDY7r00CRjH (talk) 18:52, 10 July 2024 (UTC)

ComfyUI also has a page now. I may create one for Automatic1111 as well if there is interest. These are the main UIs that people are using. J2UDY7r00CRjH (talk) 02:38, 11 July 2024 (UTC)

ComfyUI is already mentioned in the article - it's just not defined. IMO it's not worth mentioning standard UIs, which is why I gave it only half a sentence in my edit that you reverted. Michaelmalak (talk) 18:49, 10 July 2024 (UTC)

I see your point. My issue is not that you listed ComfyUI but that you put "node-based editing" as a capability of Stable Diffusion. But the node-based part is not part of Stable Diffusion. I think if we create a section it should be called "User Interfaces" and list both ComfyUI and Fooocus as well as perhaps Automatic1111 (although we will need RS demonstrating notability for that). Also, I wouldn't call Fooocus a standard UI. It is simple to use but it is not simply a frontend for SD. You can read about what it does in the article or https://github.com/lllyasviel/Fooocus?tab=readme-ov-file#tech_list. J2UDY7r00CRjH (talk) 18:58, 10 July 2024 (UTC)

@Stefan Weil @Michaelmalak I added a new section called User Interfaces and included Stability's official DreamStudio as well as included ComyUI and Fooocus. let me know what you think. J2UDY7r00CRjH (talk) 19:23, 10 July 2024 (UTC)

References

LoRAs

The section End-user fine-tuning should include LoRAs alongside embeddings and hypernetworks, since a LoRA is currently a very common way to fine-tune a stable diffusion model. Existing text about LoRA on Wikipeida can be found here: Fine-tuning (deep learning)#Low-rank adaptation . I would add the text, but I'm not qualified to write it. 74.92.140.241 (talk) 01:49, 17 September 2024 (UTC)