Constant skinny bodies, then constant errors when requesting larger body types

[scarlet mulch]

Bug Report:

Steps to reproduce:

1. Ask for a man, woman, girl or boy.
2. Ask for a larger body type. Even ask ChatGPT first what words it suggests are most appropriate and use them.

Expected result:

1. A realistic depiction of our diverse population. By default, there should be a variety of body sizes and types.
2. When specifically asking for a body type, it should provide characters with the requested body type.

Actual result:

1. You will constantly receive skinny people.
2. You will start entering a rabbit hole of constant errors, rate limits, images that aren't what were requested, and ChatGPT even accusing you of objectification

Additional information
This is an issue that is getting worse, wastes many hours of people's time, wastes OpenAI's resources and is simply illogical, flying in the face of diversity and inclusivity. All I ever get are images of skinny people constantly, often ultra skinny, and the moment I ask for any larger body type (using completely normal words, even ones ChatGPT says are appropriate and suggests itself), it starts either accusing me of content policy violations and/or objectification with error after error after error, or just continues to send skinny people.

Due to this extreme body diversity censorship, as well as the other extreme censorship in normal everyday poses, it is so difficult and frustrating to generate a simple, harmless, body POSITIVE scene. For example, a plus-size person laying on their back, whether on a chair, grass, ground, etc. Either constant errors or images of skinny people laying on their front instead. Hours upon hours wasted trying to use DALL-E 3 for a genuine, non-problematic, positive cause.

I have received easily hundreds of errors today and over the past week, way more errors than actual images, yet not a single request of mine was bad in any way. All body types should be represented, not just "perfect skinny" people. Please fix this extreme censorship, thanks.

Other people are also experiencing this.

• Example: #images-discussions message
• Example: #images-discussions message
• Example: #images-discussions message

[soft quartz]

I haven't tried this exact use case, but I do run into issues like this where it really wants to sample more common/popular body shapes and will not understand my explicit requests to change that up. It has NO IDEA how to draw a person missing an arm in any way other than giving them a robot arm, for example. (And it loves to give them other random robot bits)

And I have ranted at length on other discords about how hard it railroads the appearance of nearly any character towards what is conventionally attractive in our society. Many of my attempts to generate short, chubby gnomes result in tall hot elf girls, which is always a giggle even as I grumble and add modifier words to fix it

I don't believe this is willful on the part of the devs, but rather reflects the biases that appear in art made by humans in our culture already.

It's not a solution, BUT I have found that the bing image generator is much less strict about what words it will block out of the prompt. I have to sometimes copy a prompt over to it if I want to even use certain innocuous anatomical words.

[scarlet mulch]

Completely agree about both of those elements too. I have been unable to generate any images with various disabilities, and instead run into error after error, or the images that do get generated aren't what was requested. For example, like you said about adding robotic limbs instead of a realistic missing limb whether through congenital absence or else.

I think DALL-E 3 does a better job than competitors at steering away from conventionally attractive characters in general, but there's still a long way to go as the extreme censorship significantly reduces the usability of it, placing it behind competitors. It is such a huge shame that this incredible technology is restricted so much for genuine uses that are simply desiring diversity, inclusivity and ultimately, realistic depictions of our societies.

Will give Bing a try, thanks!

[soft quartz]

I truly think it's a limitation of the training data ie way less people draw these subjects so it doesn't know what to do with it

Once I realized it was making my tabaxi wizard buff because dudes in fantasy art skew buff. I added in a few modifiers like "scraggly" and it quickly understood the assignment and gave me the planar cryptid energy I was looking for

[scarlet mulch]

I don't know if it's so much of a limitation, but rather a bias that hasn't been refined on its limitations. Even though there is significantly more art depicting skinny bodies, there are still massive amounts of references of other body types particularly from stock imagery.

DALL-E 3 has still been able to occasionally produce a non-skinny body for me and it does it well, it's just insanely difficult usually to do and clearly starts degrading the user experience substantially with errors, accusations, etc. Other AI image generators can do different body types way way easier than DALL-E 3 and you don't get wrongly flagged or continuous errors.

So I guess it's more of a matter of refining the biases to work with any limitations in content it has, while improving the accuracy of its detection of bad behavior (e.g. body-shaming).

[molten wigeon]

Yeah mainly only SD has had success for me and only specific models such as RealisticVision. I have tested it similarly but for female characters with stronger build(ie warrior characters that are believable). ControlNet & Img2Img can help a lot though it's challenging.

while improving the accuracy of its detection of bad behavior (e.g. body-shaming).
For female characters at least, describing things in terms of feminism seems to help at least in avoiding it being seen as bad behavior. But yeah it really tends to gravitate towards super skinny or exaggerated muscle like male or outright male character

[scarlet mulch]

Keep meaning to give SD a try! Out of the ones I've tried, MJ has been a lot easier to accomplish body diversity.

Funnily enough I was using appropriate language originally, then asked ChatGPT to suggest different terms, such as ones more acceptable, or biological, all kinds, and then it started accusing me of objectification because of the terms it suggested itself 😂 Completely agree with the exaggerated male characters too – I asked for a man holding a skateboard and it generated a huge ripped bodybuilder, making the skateboard look like a toy under his massive muscly arm.

[molten wigeon]

So this trick is a lot of effort and we oughtn't have to do it but one thing that worked is I came up with an exaggerated premise for the character concept: An Elder Scrolls mod featuring an island kingdom where historical gender norms are completely inverted like in the manga series Ōoku: The Inner Chambers

I got 2 good images after a lot of attempts and after this setup. I only tried out of boredom due to being stuck in bed after surgery and because absurd premises have produced some of the most interesting responses. I typically just use SD but can't on mobile

SD via Automatic1111 is pretty great, far more control than the cloud based stuff though DallE with ChatGPT is pretty good for abstract ideas

Funnily enough I was using appropriate language originally, then asked ChatGPT to suggest different terms, such as ones more acceptable, or biological, all kinds, and then it started accusing me of objectification because of the terms it suggested itself 😂
This actually happened to me too on a similar chat, using the word it suggested resulted in an automated warning message

[molten wigeon]

@scarlet mulch If you do try SD or SDXL, the ThinkDiffusionXL model looks like it might be well suited for your needs. I don't know if we're allowed to post any links here as commonly on big official servers the automod would understandably prevent it due to spam, so try a search for "Probably the Best Model of 2023 So far. Sebastian Kamph Youtube". On the first image on the civitai page it will show the generation parameters used when viewing it which is great for a starting point & you could use ControlNet with another image that has what you want and use Reference mode or SoftEdge

[scarlet mulch]

That's a really interesting method! Yeah we definitely shouldn't have to do that, though at the same time it sounds like that setup will lead to some creative results. Always good to use a limitation to think outside the box as you have.

The ThinkDiffusionXL model seems like a great shout. The range and quality of samples on CivitAI of this model are outstanding. Will install SD via Automatic1111 this weekend and give it a try. Have been put off SD by the lack of prompt specificity understanding and the initial learning curve of LoRAs, all the various settings, etc, but it seems to have come a long way. Will start with that video – appreciate you taking the time my friend 🙂