AI Art: It is criminal to not post your prompt

The Something Awful Forums > Main > General Bullshit > AI Art: It is criminal to not post your prompt

rio: Mar 20, 2008

I was hoping to get some help with a prompt issue. I keep coming back to this every few days but SD never seems to do what I am looking for despite the varied prompts I try.

I was trying to get an image of a girl in silhouette standing in front of a ruined city, but zoomed out so the girl doesn�t appear very large in the frame.

1) if I only type �silhouette� or �girl silhouette� then it works fine. However if I direct any specifics like �wearing long boots� or �with robot arms� it will not show the subject in silhouette. Is there any way to make the person totally dark (an actual silhouette)while still directing specifics about what they are wearing and things like that?

2) I can�t get the camera as zoomed out as I�d like. I tried a lot of different prompts for this but the person is always too big, the smallest seeming to be full body composition but height at about half of the height of the image. I�ve also tried adding �closeup� and things like that in the negative prompt. When I add wide angle instructions like �24mm lens� it will give a more panoramic view of the background but make the subject look like and actual giant who is too big for the scene. At least I think that is what causes it - as I add more prompts it is often making the girl not just too big in frame but also physically too large.

I feel like this one has a straightforward solution but trying a lot of different prompts I can�t seem to find it, nor through searching online.

# ? Jul 20, 2023 19:28

Adbot: ADBOT LOVES YOU

# ? Jun 11, 2024 00:02

Megazver: Jan 13, 2006

rio posted:

I was hoping to get some help with a prompt issue. I keep coming back to this every few days but SD never seems to do what I am looking for despite the varied prompts I try.

I was trying to get an image of a girl in silhouette standing in front of a ruined city, but zoomed out so the girl doesn�t appear very large in the frame.

1) if I only type �silhouette� or �girl silhouette� then it works fine. However if I direct any specifics like �wearing long boots� or �with robot arms� it will not show the subject in silhouette. Is there any way to make the person totally dark (an actual silhouette)while still directing specifics about what they are wearing and things like that?

Some things are easier to just learn to do in photoshop than ask the AI to do and this is probably one of them.

quote:

2) I can�t get the camera as zoomed out as I�d like. I tried a lot of different prompts for this but the person is always too big, the smallest seeming to be full body composition but height at about half of the height of the image. I�ve also tried adding �closeup� and things like that in the negative prompt. When I add wide angle instructions like �24mm lens� it will give a more panoramic view of the background but make the subject look like and actual giant who is too big for the scene. At least I think that is what causes it - as I add more prompts it is often making the girl not just too big in frame but also physically too large.

You should either do outpainting in SD or use the new-fangled Zoom function in Midjourney. I must admit I am not up-to-date on how to do it in SD, but now that you know the term, you should be able to find it. As for MJ, I tried to generate something basic as an experiment to show you:

"a concept art painting for a postapocalyptic videogame, a silhouette of a girl survivor seen from her back, standing in front of a ruined city --ar 3:2"

Then I upscaled the one I liked and just Zoomed it x2:

# ? Jul 20, 2023 21:47

rio: Mar 20, 2008

Megazver posted:

Some things are easier to just learn to do in photoshop than ask the AI to do and this is probably one of them.

You should either do outpainting in SD or use the new-fangled Zoom function in Midjourney. I must admit I am not up-to-date on how to do it in SD, but now that you know the term, you should be able to find it. As for MJ, I tried to generate something basic as an experiment to show you:

"a concept art painting for a postapocalyptic videogame, a silhouette of a girl survivor seen from her back, standing in front of a ruined city --ar 3:2"

Then I upscaled the one I liked and just Zoomed it x2:

Thanks! I have been trying to avoid post processing to see if I can get what I want purely out of SD but can just pop into an image editor to burn shadows for a silhouette. If that�s the way to go then I�ll go that route if things like that are a problem when generating an image.

Outpainting to get a smaller subject makes sense - I have only dealt a bit with outpainting so far so didn�t think of using it for this purpose. Appreciate the help!

# ? Jul 20, 2023 22:32

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

With SD you can also use control net if you know the full pose you want to also get the size how you want at whatever and do an upscale instead of outpainting.

# ? Jul 20, 2023 22:40

Hadlock: Nov 9, 2004

These guys are claiming 3 hours to generate a 512x512px image in stable diffusion using a raspberry pi with ~260mb memory using all sorts of tricks :hmmyes:

https://github.com/vitoplantamura/OnnxStream

Swagman posted:

What was the prompt for this one

# ? Jul 20, 2023 22:47

Sedgr: Sep 16, 2007; Neat!

# ? Jul 20, 2023 23:38

RIP Syndrome: Feb 24, 2016

rio posted:

I was hoping to get some help with a prompt issue. I keep coming back to this every few days but SD never seems to do what I am looking for despite the varied prompts I try.

I was trying to get an image of a girl in silhouette standing in front of a ruined city, but zoomed out so the girl doesn�t appear very large in the frame.

1) if I only type �silhouette� or �girl silhouette� then it works fine. However if I direct any specifics like �wearing long boots� or �with robot arms� it will not show the subject in silhouette. Is there any way to make the person totally dark (an actual silhouette)while still directing specifics about what they are wearing and things like that?

2) I can�t get the camera as zoomed out as I�d like. I tried a lot of different prompts for this but the person is always too big, the smallest seeming to be full body composition but height at about half of the height of the image. I�ve also tried adding �closeup� and things like that in the negative prompt. When I add wide angle instructions like �24mm lens� it will give a more panoramic view of the background but make the subject look like and actual giant who is too big for the scene. At least I think that is what causes it - as I add more prompts it is often making the girl not just too big in frame but also physically too large.

I feel like this one has a straightforward solution but trying a lot of different prompts I can�t seem to find it, nor through searching online.

Megazver did a really nice job with MJ. I think it's a little harder to get nice-looking pictures with SD, but here's one way to do it:

If you're using Auto1111, you could look into generating just the person first, preferably with a blank background. Then send that to ControlNet, select the segmentation ("Seg") preprocessor and hit the little 💥 button. It should convert it into a segmentation map (looks like a silhouette with funky colors). Edit that in an image editor, for instance you could make the canvas bigger and place the outline where you like. Something like this:

You could generate a normal map instead, maybe you'll get better clothing detail that way.

Adjust the "ending control step" to 0.8 (higher means the silhouette will influence the final image more). Make sure you tick the "enable" box under the ControlNet panel, then change your prompt to read "girl silhouette, ruined city background" and generate a bunch of images until you get one you like. It helps to have a bunch of models installed and using the x/y/z script for exploring different combinations of model, CFG and negative embeddings.

I got this one from the juggernaut_final model.

To zoom out, send the image to img2img and use outpainting. Under "scripts" you'll find a couple of options; "outpainting mk2" works for me with a bit of tweaking (denoise at 0.68 or close to it). If it's a cityscape it may be enough to outpaint left/right to widen the horizon.

Then send to extras and upscale as you like.

# ? Jul 20, 2023 23:55

Swagman: Jun 10, 2003; Yes...all was once again peaceful in River City.

Hadlock posted:

What was the prompt for this one

its been a minute but something fairly simple like mech-guineapig or guineapig-mech possibly with a 'deformed' in the negative prompt
had to go through a lot of generic (but still beautiful) pigs until I got that one

# ? Jul 21, 2023 00:35

hydroceramics: Jan 8, 2014

Hadlock posted:

These guys are claiming 3 hours to generate a 512x512px image in stable diffusion using a raspberry pi with ~260mb memory using all sorts of tricks

https://github.com/vitoplantamura/OnnxStream

What was the prompt for this one

Guinea pig mech suit:

Guinea pig T79 tank cosplay:

# ? Jul 21, 2023 01:14

Sedgr: Sep 16, 2007; Neat!

# ? Jul 21, 2023 01:32

RIP Syndrome: Feb 24, 2016

# ? Jul 21, 2023 02:18

Humbug Scoolbus: Apr 25, 2008; The scarlet letter was her passport into regions where other women dared not tread. Shame, Despair, Solitude! These had been her teachers, stern and wild ones, and they had made her strong, but taught her much amiss.; Clapping Larry

# ? Jul 21, 2023 02:42

KakerMix: Apr 8, 2004; 8.2 M.P.G.

RIP Syndrome posted:

Megazver did a really nice job with MJ. I think it's a little harder to get nice-looking pictures with SD, but here's one way to do it:

If you're using Auto1111, you could look into generating just the person first, preferably with a blank background. Then send that to ControlNet, select the segmentation ("Seg") preprocessor and hit the little 💥 button. It should convert it into a segmentation map (looks like a silhouette with funky colors). Edit that in an image editor, for instance you could make the canvas bigger and place the outline where you like. Something like this:

You could generate a normal map instead, maybe you'll get better clothing detail that way.

Adjust the "ending control step" to 0.8 (higher means the silhouette will influence the final image more). Make sure you tick the "enable" box under the ControlNet panel, then change your prompt to read "girl silhouette, ruined city background" and generate a bunch of images until you get one you like. It helps to have a bunch of models installed and using the x/y/z script for exploring different combinations of model, CFG and negative embeddings.

I got this one from the juggernaut_final model.

To zoom out, send the image to img2img and use outpainting. Under "scripts" you'll find a couple of options; "outpainting mk2" works for me with a bit of tweaking (denoise at 0.68 or close to it). If it's a cityscape it may be enough to outpaint left/right to widen the horizon.

Then send to extras and upscale as you like.

This is killer advice, thank you. I've been lazily loving with MidJourney so much that Stable Diffusion has slipped away from me too much that I miss now that ControlNet has a segmentation function now.

Really though I'm waiting for the official release of Stable Diffusion XL.

# ? Jul 21, 2023 03:38

RIP Syndrome: Feb 24, 2016

Aren't we all. I haven't actually tried the leaked model, but I noticed Auto1111 was adding support for it. Probably won't bother until there's something finalized.

# ? Jul 21, 2023 04:06

crazysim: May 23, 2004; I AM SOOOOO GAY

I don't even know about leaked. It's so easy to get if you just put in some info. There's no human checking of the form. It's instant.

# ? Jul 21, 2023 04:12

RIP Syndrome: Feb 24, 2016

crazysim posted:

I don't even know about leaked. It's so easy to get if you just put in some info. There's no human checking of the form. It's instant.

Yeah, I guess I'm mainly just lazy, and I'd have to deal with beta a1111 integration and then update it again later anyway

# ? Jul 21, 2023 04:19

Hadlock: Nov 9, 2004

RIP Syndrome posted:

Yeah, I guess I'm mainly just lazy, and I'd have to deal with beta a1111 integration and then update it again later anyway

Is this hampster dance

# ? Jul 21, 2023 08:45

SniperWoreConverse: Mar 20, 2010; Gun Saliva

bing is extremely displeased at the idea of a hamster military parade

# ? Jul 21, 2023 11:38

RIP Syndrome: Feb 24, 2016

Hadlock posted:

Is this hampster dance

full body view of one anthro (hamster dancing:1.2) at a rave, holding glowsticks, laughing with closed eyes, singing, confetti, music, isometric
Negative prompt: bad-picture-chill-1v Steps: 150, Sampler: DPM++ 2M SDE Karras, CFG scale: 5, Seed: 2081595673, Size: 1536x1536, Model hash: 61e23e57ea, Model: mixProV4_v4, Denoising strength: 0.51, Mirror Mode: 1, Mirror Style: 3, Mirroring Max Step Fraction: 0.05, X Pan: 0.01, Y Pan: 0.01, Hires upscale: 1.5, Hires upscaler: Latent, Version: v1.3.2

I've been experimenting with getting SD to make images much, much bigger than the 512x512 it was trained on. The results are usually not great because your motif gets repeated all over the place and there's no real cohesion. But what if you just roll with it?

You have to tune the parameters just right (and use a simple prompt asking for one instance of something). I used Latent Mirroring (thanks KwegiboHB!) to introduce some macro structure, having it rotate and pan a little bit between each step.

The result is 2304x2304 pixels (sorry) generated in one shot (well, technically two, since it does a pass with latent upscaling). You could use SwinIR on that and make a 4k or 8k wallpaper that doesn't look completely awful, although tbh I'm not sure I've gotten anything wallpaper worthy this way yet. You've set the bar high in this thread.

# ? Jul 21, 2023 15:35

Megazver: Jan 13, 2006

Megazver posted:

I figured it out:

Turns out MJ has a soft censorship mode where it does this to prompts that it thinks might be NSFW, but not enough to deny it altogether. In my particular case, the heinous crime I was asking it to commit was "color bleed". Very annoying.

You can pop them back into the channel with /show.

Ugh, I think they also retroactively hid all the older pics that didn't pass through their filter, like a number of slightly booby ones I generated for my D&D game:

I relied on generating stuff in different channels on my server to keep track of stuff I generated, and now it turns they can just delete all that willy-nilly. Ughhhhhhhhhh.

Megazver fucked around with this message at 15:44 on Jul 21, 2023

# ? Jul 21, 2023 15:42

Serotoning: Sep 14, 2010; D&D: HASBARA SQUAD
HANG 'EM HIGH

We're fighting human animals and we act accordingly

RIP Syndrome posted:

The result is 2304x2304 pixels (sorry) generated in one shot (well, technically two, since it does a pass with latent upscaling). You could use SwinIR on that and make a 4k or 8k wallpaper that doesn't look completely awful, although tbh I'm not sure I've gotten anything wallpaper worthy this way yet. You've set the bar high in this thread.

How much RAM does that require?

# ? Jul 21, 2023 16:41

Roman: Aug 8, 2002

Going to make an art book out of the sci-fi thing I'm doing in the style of a "The Making of the Movie" book. These are my cover options so far, I think I'm going with the first one.

# ? Jul 21, 2023 17:30

RIP Syndrome: Feb 24, 2016

Pvt. Parts posted:

How much RAM does that require?

I dunno if this is the best way to measure it, but nvidia-smi says 7135MiB VRAM in use during the first pass, and then 9301MiB on the latent upscale. When it's about to finish, it briefly goes to 24GiB. That's where it bombs out if I try to go any bigger. Main memory usage seems to be under 6GiB. I use xformers, maybe that helps.

a1111 says this:

Time taken: 6m 17.41s Torch active/reserved: 20484/39086 MiB, Sys VRAM: 24564/24564 MiB (100.0%)

Dropping it into SwinIR and 4x'ing it to 9216x9216 I get:

Time taken: 1m 25.62s Torch active/reserved: 3863/4986 MiB, Sys VRAM: 6646/24564 MiB (27.06%)

SwinIR is tile-based, so it's not all that resource hungry. Web browser's really struggling at those sizes, though.

# ? Jul 21, 2023 17:41

Roman: Aug 8, 2002

RunwayML Gen-2 now allows image only prompts and, uh, it's very good.

https://i.imgur.com/G6HTO8W.mp4

# ? Jul 21, 2023 17:49

LifeSunDeath: Jan 4, 2007; still gay rights and smoke weed every day

I do see cars morphing but yeah it looks awesome otherwise.

# ? Jul 21, 2023 17:53

Roman: Aug 8, 2002

LifeSunDeath posted:

I do see cars morphing but yeah it looks awesome otherwise.

It was just a quick test. But being able to actually use the input image as the first frame is a pretty big deal for me.

https://twitter.com/WilliamLamkin/status/1682409471887265792

Roman fucked around with this message at 18:07 on Jul 21, 2023

# ? Jul 21, 2023 18:03

LifeSunDeath: Jan 4, 2007; still gay rights and smoke weed every day

Roman posted:

It was just a quick test. But being able to actually use the input image as the first frame is a pretty big deal for me.

https://twitter.com/WilliamLamkin/status/1682409471887265792

this poo poo is nuts. my buddy I sent this too was asking about "Roop," said it was for face fakes, dunno what do people use for deep fakes now?

# ? Jul 21, 2023 18:18

Cabbages and VHS: Aug 25, 2004; Listen, I've been around a bit, you know, and I thought I'd seen some creepy things go on in the movie business, but I really have to say this is the most disgusting thing that's ever happened to me.

crazysim posted:

I don't even know about leaked. It's so easy to get if you just put in some info. There's no human checking of the form. It's instant.

I didn't even lie about any of the info I put in; on the other hand I legitimately work for a decent sized ML place, so even though this is personal interest more than work, "loving whatever, man"

# ? Jul 21, 2023 18:24

Roman: Aug 8, 2002

LifeSunDeath posted:

this poo poo is nuts. my buddy I sent this too was asking about "Roop," said it was for face fakes, dunno what do people use for deep fakes now?

for still images I use InsightFace Swap which runs through Discord and is free (works great with Midjourney workflow). I haven't seen any non-open source DIY deepfake software that isn't paid.

# ? Jul 21, 2023 18:30

Doctor Zero: Sep 21, 2002; Would you like a jelly baby?
It's been in my pocket through 4 regenerations,
but it's still good.

Cool! Hamsters!

# ? Jul 21, 2023 18:42

Doctor Zero: Sep 21, 2002; Would you like a jelly baby?
It's been in my pocket through 4 regenerations,
but it's still good.

Roman posted:

RunwayML Gen-2 now allows image only prompts and, uh, it's very good.

https://i.imgur.com/G6HTO8W.mp4

Holy poo poo, they made a HUGE leap forward from a couple weeks ago.

# ? Jul 21, 2023 18:44

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

Made this for an MS paint thread then went how will SD handle calling this dime store doctor doom (quote from the show)

calling him Baron Underbheit

Maybe I should be picking up MS paint skills again and combining it the results are fun. Don't know why it got rid of the metal jaw (I think I did a poor job defining it in Gimp)

# ? Jul 21, 2023 18:49

KakerMix: Apr 8, 2004; 8.2 M.P.G.

LifeSunDeath posted:

this poo poo is nuts. my buddy I sent this too was asking about "Roop," said it was for face fakes, dunno what do people use for deep fakes now?

roop is two things, a video deepfake thing and a stable diffusion plugin that does the same thing but with static images. Open source, completely free:

KakerMix posted:

https://github.com/s0md3v/roop
https://github.com/s0md3v/sd-webui-roop

roop, both a video deep-fake generator (and all the uncomfortable things that come with it) and a stable diffusion plugin allowing the same function on images.

What this allows you to do is what people have been asking for a long time now: Generate a single character and then bring them back up.

This native american CEO boss lady I used MidJourney to create, she isn't anyone in particular, but distinctive:

Using roop in stable diffusion she's now:

An 80s stylish business woman!

A race car driver!

Shopping!

Making cookies!

A man!

Younger!

A space marine or cosplayer, your choice!

LARPing!

A scientist!

Boxing!

At the club!

Scaring the children!

All done as quick as any other image in Stable Diffusion, with roop running just afterwards to put the face on, it looks like this:

That single image imposed that face on all the images I just posted. One picture, seconds of time, a default character you can summon up.

most people use it to do all the poo poo you'd expect, but the other part is you can make a person that doesn't exist and keep them around, without having to invoke a celebrity that's in the data.

# ? Jul 21, 2023 18:55