Trying My Hand at AI Art *Advice Welcome*

[Granelex]

I know very strange. Sometimes I get the opposite of what I expecting. For instance I tried this prompt: "Brunette woman in an apartment next to a giant couch. Scientist is pointing a shrink ray at her which is shooting a beam of light. Woman is very small. Scientist is normal sized. Surprised expression. Woman is wearing white capris, a blue blouse, and cork wedge sandals. Brown eyes and very detailed face." Then I got this output:

[foreverlurk]

It seems weirdly hard to get AI to render a classic King Kong handheld -- a giant closed fist clutching a struggling girl. Either it says "Unsafe Content" or some nonsense, or it spits up palm holds and hand-adjacent stuff. Still tinkering.

I've managed to create a few, but it's very hard - the filters identify this grab/pose as "putting someone in distress" or a dangerous situation. You have to carefully craft your prompt to make it way less threatening that it sounds.

[rscholar]

I know very strange. Sometimes I get the opposite of what I expecting. For instance I tried this prompt: "Brunette woman in an apartment next to a giant couch. Scientist is pointing a shrink ray at her which is shooting a beam of light. Woman is very small. Scientist is normal sized. Surprised expression. Woman is wearing white capris, a blue blouse, and cork wedge sandals. Brown eyes and very detailed face." Then I got this output:

It's definitely tricky. I think the key may be to be specific about both the woman and the surroundings - I usually use combos of "tiny" "shrunken" "extremely small" "only a few inches tall", followed by a little description of the surroundings: "enormous chair" "blades of grass towering over her." Also helps to give directions, like where the ray gun is pointing.

None of it is foolproof and I've made way more images than I ended up saving. I'm especially getting these weird results where "tiny woman" more than a few times gives me childlike Pixar characters. I will say I seem to have more luck with outdoor scenes than inside. Describing furniture especially seems to combine the character with the size - woman near giant couch means the woman is giant too (I guess).

Sometimes lightning strikes and I get results like the first one which I'm really happy with. I used this prompt:

"zoomed in side view of a photorealistic shrunken extremely small woman in her 40's, only a few inches tall, wearing flipflops, magical blue mist swirling around her body, standing at the edge of a driveway, looking up in confusion, a ray gun above shooting blue energy down at her chest, nearby blades of grass rising high above"














I monkied around with your prompt and hit on something that kind've works - I'm also getting those ones where she's giant for some reason.

"closeup view of a shrunken extremely small Brunette woman with brown eyes and very detailed face, only a few inches tall, wearing white capris, a blue blouse, and cork wedge sandals, standing inside an apartment next to a couch that towers over her, looking up in surprise at a giant male scientist pointing a ray gun down at an angle shooting a beam of faint yellow energy at her chest."

[Pacific1982]

I decided I would give it try using your prompt. Here are the results using Bing DALL-E3. The thing I had the most difficulty with was getting the prompt to fit inside the prompt box. Seems there is a character limit.

[Granelex]

Thanks for working with my prompt. I appreciate it . I try to use shrinking instead of shrunken in my prompt.

by all means feel free to share prompts

"closeup view of a shrinking extremely small attractive Brunette woman with brown eyes and very detailed face, only a few inches tall, wearing white capris, a blue blouse, and cork wedge sandals, standing inside an apartment next to a couch that towers over her, looking up in surprise at a giant male scientist pointing a ray gun down at an angle shooting a beam of faint green energy at her chest."

[Prof Sai]

These are quite good. Are these post-edited, or is this how the prompts render? I can't get two distinct characters at once in SD, without cutting them together in a photo editor. But of course I couldn't deal with the restrictions of an online service either.

[Granelex]

[Pacific1982]

I try to use shrinking instead of shrinking in my prompt.

Shrinking instead of shrinking? Did you mean to say something else?

[Granelex]

Yes let me correct that. I use shrinking in my prompts instead of shrunken. Shrunken makes the ladies look rather chibi-ish.

[Granelex]

I want to thank everyone for their help with the prompts. Here is some of the work I got from the prompt: "Extremely small attractive Brunette woman in a giant laboratory, on a giant lab desk, shrunk by an enormous shrink ray, magical teal mist surrounding her, shrink ray is held by a giant male above her scientist pointing it down at her, Surprised expression, wearing white capris, a blue blouse, and cork wedge sandals. Dark Brown eyes and very detailed face." Enjoy and leave any suggestions you would like.

[rscholar]

I just tried "Imagine with Meta". I feel like it's an overall more powerful tool than Bing and is actually infinitely reusable with no credits and averages about 10 seconds for every prompt. However, it seems much trickier to use and doesn't follow Bing's "language" with prompts so it takes a lot of redos to get anything decent. That said:

[foreverlurk]

I just tried "Imagine with Meta". I feel like it's an overall more powerful tool than Bing and is actually infinitely reusable with no credits and averages about 10 seconds for every prompt. However, it seems much trickier to use and doesn't follow Bing's "language" with prompts so it takes a lot of redos to get anything decent. That said:

Wow. Meta looks very good, but their watermark is kind of annoying.

(....annnd of course, it's not available in Canada.)

[Granelex]

I have been getting some better prompts using Dall-E. What have you found as worked for you guys when using Dall-E? When playing around its hard to get a decent model that doesn't look distorted or her eyes seeming awkward. Here is arompt I have been using. Feel free to tweak it as you like: "close up of an attractive Brunette woman standing in inside ontop of a giant mousepad on a giant table in a giant labratory, magical teal mist surrounding her, raygun is directing the mist towards her, Surprised expression, wearing white capris, a blue blouse, and cork wedge sandals. Dark Brown eyes and very detailed face."

I have gotten result like this:






also check my deviant art page and tell me what you think:
https://www.deviantart.com/granelex

[Prof Sai]

I tried the same prompts and got this:

[Pacific1982]

I have been getting some better prompts using Dall-E. What have you found as worked for you guys when using Dall-E? When playing around its hard to get a decent model that doesn't look distorted or her eyes seeming awkward. Here is arompt I have been using. Feel free to tweak it as you like: "close up of an attractive Brunette woman standing in inside ontop of a giant mousepad on a giant table in a giant labratory, magical teal mist surrounding her, raygun is directing the mist towards her, Surprised expression, wearing white capris, a blue blouse, and cork wedge sandals. Dark Brown eyes and very detailed face."

Here is the result I got with your prompt using DALL-E3. The first image was OK but the second one produced an enlarged head on the girl and also had her holding the ray gun as opposed to being shot by one.

[foreverlurk]

Using a mix of DALL-E and Stable Diffusion... using "detailed face" in DALL-E seems to give a LOT of weight to that token, thus making the head/face much bigger than it should be.

[Granelex]

Yes I get those results often as well. I typically have to try a few times to get a good result. Would you know of another prompt that could be used to the results come out less like a bobble head? Thanks for the replies Pacific1982 and Foreverlurk!

[Pacific1982]

I liked this one particular image, so I used it in a collage to see her squeezed by a hand.

I smooshed her legs closer together (since squeezing her would probably do that). I think her expression matches the circumstances she is in at the moment. I chose to have her arms dangling freely outside of the Fingers.

[Granelex]

Simply amazing Pacific! A stellar image.

[foreverlurk]

Would you know of another prompt that could be used to the results come out less like a bobble head?

If you're working with DALL-E I think you should avoid mentionning "detailed face". It depends on the scene, if the SW is taking up more than 50% of the render than you could keep it, but beyond a threshold it's better to fix it later with SD. Otherwise the AI will try to give you a detailed face anyway, which might be what is leading to the bobbleheads (just a theory).

I liked this one particular image, so I used it in a collage to see her squeezed by a hand.

Wow, amazing! I bow to your skills.

[yahedon]

Hi, I am a newbie with this AI stuff. I have seen that there are LORA packages to achieve specific effects/objects/poses. I wonder if it will be possible to train one of this LORA to produce SW stuff (this would be amazing!). The question is, in order to create one, we must train it with set of images, I wonder if the images created by this community should be enough to the AI to get the "point"

[foreverlurk]

Hi, I am a newbie with this AI stuff. I have seen that there are LORA packages to achieve specific effects/objects/poses. I wonder if it will be possible to train one of this LORA to produce SW stuff (this would be amazing!). The question is, in order to create one, we must train it with set of images, I wonder if the images created by this community should be enough to the AI to get the "point"

I've only had experience with training models on Stable Diffusion 1.5 using Dreambooth. It's very good at adding new tokens to an existing model, or at overriding existing tokens and replacing them with your own. For instance you can train it to you own face and body and it will give pretty accurate results. I've tried using it for "tiny" or "shrunken" concepts and it didn't work at all, there was maybe too much variation in my dataset for proper training.

LoRA is a different technique than Dreambooth, it's like a leaner, much faster version that works on a different layer. It seems easier to train it on concepts rather than just models, and there's at least one LyCORIS (a more advanced version of LoRA) available right now :
https://civitai.com/models/151689/shrunk-xl-or-concept

Sadly I have no experience with training LORAs yet. Maybe reach out and ask the author of ShrunkXL what he used? I'm sure there will eventually be a fine-tuned model purely for size content, but right now there's nothing out there.