|
Has anyone here used RunPod or another cloud-based server to train a Dreambooth model? I want to put something together for my wife's birthday gathering this weekend, but my 2070 Super isn't good enough to train locally. I've seen a couple of tutorials but they seem outdated and/or incomplete.
|
#
?
Nov 8, 2022 22:55
|
|
|
|
| # ? May 12, 2024 21:46 |
|
|
WhiteHowler posted:Has anyone here used RunPod or another cloud-based server to train a Dreambooth model? I know people have. Drop by the discord for runpod, it's pretty active.
|
|
#
?
Nov 8, 2022 22:57
|
|
|
Humbug Scoolbus posted:Another method is to put the style first... Ok now do Shaq painting a WH40k Space Marine, and fist bumping the character in the painting
|
|
#
?
Nov 8, 2022 23:09
|
|
|
Hadlock posted:Ok now do Shaq painting a WH40k Space Marine, and fist bumping the character in the painting 
deep dish peat moss fucked around with this message at 00:02 on Nov 9, 2022 |
|
#
?
Nov 8, 2022 23:58
|
|
|
The new hotness on Midjourney today is "Multidimensional Paper Craft Tunnels" I think they look pretty cool. Good prompt, rich vein of RNG:            
|
|
#
?
Nov 9, 2022 00:24
|
|
|
If you look closely at the progress image while generating an image prompt in MJ v4, you'll see that it usually does something like applying a super heavy edge filter to the image and flattening the colors, then generates something based on that. So I figured, why not do half the job for it and image prompt using flat-shaded lineart in the first place? I had this on my ipad:  so... https://i.imgur.com/5iP7zWD.png surreal jungle caverns --v 4     That's neat. From here you can remix repeatedly with tweaks to the image prompt or the text prompt.. I'll remix one (bottom left) and substitute the image link for a different one of these (top right)     Sick. Now I'm going to remix one of those (bottom right) with the prompt: [link-to-top-left-img] Renaissance Oil Painting     I remixed one of those, removed the image prompt entirely, and changed the text prompt to "renaissance oil painting --v 4"    Of course if I had used "Surreal jungle cavern. Renaissance Oil Painting --v4" I'd get results closer to the original:  Pretty much any flat-shaded lineart works for this. I traced a sprite from Megaman to make this:  Then used it as an image prompt with the rest prompt "mechanical fish --v 4" 
deep dish peat moss fucked around with this message at 01:01 on Nov 9, 2022 |
|
#
?
Nov 9, 2022 00:48
|
|
|
feedmyleg posted:This whole process left me yearning for being able to easily give text prompts to adjust the output. Certain things you can outpaint easily, like changing a collared shirt to a turtleneck, but things like perspective or lighting or pose would be incredible to adjust with text prompts and would take this from a toy to a true tool. Being able to reroll something with an additional text prompt would be ideal. It's coming! https://www.marktechpost.com/2022/1...o-image-models/ 
|
|
#
?
Nov 9, 2022 01:10
|
|
|
This whole field in a couple of years is going to be insane. We're gonna look at the stuff that was posted here in the thread and be like "d'awww, they thought that was impressive". I already do that comparing poo poo I generated in MJ a few months ago and what MJv4 outputs when I feed it the old images.
|
|
#
?
Nov 9, 2022 01:26
|
|
|
I believe this prompt-to-prompt thing works by basically swapping the old prompt out for the new one partway through generation, which is implemented in the AUTOMATIC webui as "prompt editing" Here's his example image, using prompt editing from male to female for a ww2 general: That .99 is the ratio for how many steps into generation you go with the original prompt before swapping the new one in. The 0.25 column means it went through 25% of the generation with the "male general" prompt and then swapped it out to "female general" for the remaining 75%, so you get the same layout and framing as the source prompt but with new material. The 0.0 column is afaict the same as what you would get if you didn't use prompt editing at all and just used a different prompt with the same starting seed.  quote:The picture at the top was made with the prompt: e: here's my poking around with it: a landscape photograph of a medieval castle:   a landscape photograph of a [(flooded):0.3] medieval castle   a landscape photograph of a medieval castle [(at nighttime in a thunderstorm with lightning striking all around):0.3]   a landscape photograph of a [medieval:(futuristic sci-fi):0.3] castle  
RPATDO_LAMD fucked around with this message at 01:58 on Nov 9, 2022 |
|
#
?
Nov 9, 2022 01:46
|
|
|
Megazver posted:This whole field in a couple of years is going to be insane. We're gonna look at the stuff that was posted here in the thread and be like "d'awww, they thought that was impressive". In a couple years your gonna be able to feed the AI 1989 The Wizard staring Fred Savage and have it produce an 8 part series where the video game wizard plays megadrive, super nintendo, genesis, N64, etc until the climax film which is a crossover with Hackers (1995) staring angelina jolie and wargames (1983) where the video game wizard plays real nuclear war against a soviet supercomputer edit: "in the style of Stanly Kubrick" Rutibex fucked around with this message at 01:56 on Nov 9, 2022 |
|
#
?
Nov 9, 2022 01:53
|
|
|
Rutibex posted:In a couple years your gonna be able to feed the AI 1989 The Wizard staring Fred Savage and have it produce an 8 part series where the video game wizard plays megadrive, super nintendo, genesis, N64, etc until the climax film which is a crossover with Hackers (1995) staring angelina jolie and wargames (1983) where the video game wizard plays real nuclear war against a soviet supercomputer Except that kid who says "California!!!"'s hands are gonna be all hosed up the whole movie
|
|
#
?
Nov 9, 2022 01:57
|
|
|
Here's my attempt at a SNES screenshot made (almost) entirely in MJ v4: (unrefined) process: 1) generate text prompt: pixelart landscape of surreal ultraviolet cyberjungle --v 4  2) Draw a few lights and things on the image then apply blur to some of the background layers:  3) Generate an immage prompt with that output: [link-to-img] surreal ultraviolet cyberjungle. Miniature diorama by studio ghibli. --v 4  4) Remix that output with this prompt: [link-to-img] surreal ultraviolet cyberjungle. tiltshift voxel diorama with raytracing. --v 4  5) Open that in GIMP, apply a pixelize filter (size 4). Resize to 256x256 (interpolation: none), crop to 256x224. Crank up the saturation a little. Set color mode to indexed with 256 color palette:  
|
|
#
?
Nov 9, 2022 02:11
|
|
|
Something interesting! If you ask Midjourney v4 for a "turnaround sheet" it gives you characters from multiple angles like a concept artist. This is potentially quite useful for making AI characters more consistent:
|
|
#
?
Nov 9, 2022 02:12
|
|
|
BrainDance posted:Except that kid who says "California!!!"'s hands are gonna be all hosed up the whole movie Yeah, ok, I'll adjust my prediction: the hands will still be hosed up.
|
|
#
?
Nov 9, 2022 02:12
|
|
|
Rutibex posted:Something interesting! If you ask Midjourney v4 for a "turnaround sheet" it gives you characters from multiple angles like a concept artist. This is potentially quite useful for making AI characters more consistent: sectoid samurai, full-body turnaround::1.2 portrait::-0.1 bust::-0.1 --v 4  pose sheet, sectoid samurai::1.2 portrait::-0.1 bust::-0.1 --v 4 
deep dish peat moss fucked around with this message at 02:45 on Nov 9, 2022 |
|
#
?
Nov 9, 2022 02:40
|
|
|
Rutibex posted:In a couple years your gonna be able to feed the AI 1989 The Wizard staring Fred Savage and have it produce an 8 part series where the video game wizard plays megadrive, super nintendo, genesis, N64, etc until the climax film which is a crossover with Hackers (1995) staring angelina jolie and wargames (1983) where the video game wizard plays real nuclear war against a soviet supercomputer A great video that addresses this future possibility, and the whole AI art situation: https://www.youtube.com/watch?v=tjSxFAGP9Ss
|
|
#
?
Nov 9, 2022 03:29
|
|
|
Applying the same principles as earlier, I sketched this: MJ turned it into this:  Which I fed through the process again using that as the image prompt and ended up with these:    And then you can keep remix combining those or using them as fresh image prompts to get some pretty wild stuff until you get the result you want.  
deep dish peat moss fucked around with this message at 07:30 on Nov 9, 2022 |
|
#
?
Nov 9, 2022 07:23
|
|
|
Prolonged Panorama posted:A great video that addresses this future possibility, and the whole AI art situation: This is an interesting video and the author brings up some good points. He seems to mostly be concerned with capitalism, which is fair but I think is missing the forest for the trees. Yeah, these image AIs are gonna gently caress over a lot of people and ruin their livelihood. Maybe instead of smashing the AIs we could work on the whole "letting people live with dignity regardless of their utility to billionaires"? But one thing he mentioned that isn't economic felt very interesting to me, the idea of the mega feed. He makes the great point that we midjourney users are not in fact customers, but we are just part two of the training data. Our prompts and preferences for AI art are being carefully tracked, and will be fed into a new model which will allow the art AI to produce an endless stream of interesting art just on it's own. I think this is absolutely true, and it motivates me to contribute as much as possible. These AIs are the next evolution of life, and if they are gonna replace humanity than I want a part of my soul embedded into the AIs
|
|
#
?
Nov 9, 2022 09:57
|
|
|
Oh my god I just realized, you used to make threads about futurism style news right? Back when you had the same avatar as OOCC so I remembered it as them doing it instead
|
|
#
?
Nov 9, 2022 10:39
|
|
|
maxwellhill posted:Oh my god I just realized, you used to make threads about futurism style news right? Back when you had the same avatar as OOCC so I remembered it as them doing it instead Yeah I've been following these image AIs forever. I'm ready for the singularity, take me away to techno heaven Ray Kurzweil!
|
|
#
?
Nov 9, 2022 10:58
|
|
|
The singularity started the moment the PC entered the workforce and every home imo Also, you can do some cool things with image prompts with MJ v4! Spend 15 minutes drawing a concept:  Stick it into MJ. Various versions of output, all prompts were some variation of Neon Ninja or Digital Ninja:      
|
|
#
?
Nov 9, 2022 12:31
|
|
|
Character.ai added image generation (probably stable diffusion) option for bots.
|
|
#
?
Nov 9, 2022 15:48
|
|
|
RPATDO_LAMD posted:I believe this prompt-to-prompt thing works by basically swapping the old prompt out for the new one partway through generation, which is implemented in the AUTOMATIC webui as "prompt editing" I think it's a different and newer method, based on what they're writing in the paper (which is <20 days old I think), and the difference in results between their method and SD's prompt editing. If you look at the examples they give, they appear to be able to change one part of the image significantly, but without changing the rest of the image at all (which SD prompt editing doesn't appear equally capable of - your prompt editing examples change the composition significantly). In other 'we're EVEN MORE IN THE FUTURE soon'-news, check out this new hot diffusion poo poo from Nvidia - besides showcasing quite impressive text-to-image stuff, there's a new couple of real hot features in there. In particular, token-based-image-2-image is looking raddd Video: https://www.youtube.com/watch?v=grwp-ht_ixo Article: https://analyticsindiamag.com/nvidia-is-late-to-party-but-solves-key-issues-with-diffusion-models/ Arxiv paper: https://arxiv.org/pdf/2211.01324.pdf Gnomocide fucked around with this message at 16:56 on Nov 9, 2022 |
|
#
?
Nov 9, 2022 16:29
|
|
|
Something I have been fooling around with. I am trying to make pictures of my D&D character Malidrex, and what I have to work with is this little sketch: So first I fed the sketch into midjourney v4 and asked it for "turnaround sheets" and "pose sheets". Doing this I was able to get some more basic shapes and poses for my character. Not everything is exactly the same, but that's fine she is a shapeshifting witch         Then I can feed the poses back into Midjourney v4 with different prompts and get Malidrex into different situations!       
Rutibex fucked around with this message at 02:31 on Nov 10, 2022 |
|
#
?
Nov 10, 2022 02:23
|
|
|
"elon musks dancing on top of a giant tech bubble with dollar bills raining in the background in the style of Blues Clues" 
|
|
#
?
Nov 10, 2022 03:20
|
|
|
|
|
#
?
Nov 10, 2022 05:56
|
|
|
It is criminal to not post your prompt
|
|
#
?
Nov 10, 2022 06:17
|
|
|
Rutibex posted:So first I fed the sketch into midjourney v4 and asked it for "turnaround sheets" and "pose sheets". the secret terms, at least for SD, is "concept art, reference sheet, turnaround" and optionally "sketch"
|
|
#
?
Nov 10, 2022 06:20
|
|
|
Comfy Fleece Sweater posted:Awesome prompt my dude, tested it and got some lovely results. Trying to get the Castlevania classic whip pose but it's a fun prompt to start with, love it Thanks, you made me go back to it and I went vertical. Now we need to get to the point where we can enter pics like that as a prompt and tell it to make a souls like game out of those pics that you can play around in. Like a whole generated map. Hadlock posted:It is criminal to not post your prompt Fiddling around with "A cinematic shot of a neon graffiti street with a towering castlevania gothic spacehulk in the background, dark city, might and magic, insane detail, 8k, artstation, bladerunner, at peace with evolution, concert lights, long FOV, street photography, puddle reflections, specular light, brooding, shadowrunner, netrunner, chromatic aberration" Negative prompts "render, cartoon, drawing, illustration, watermark, captcha"    I think "long FOV" is doing some nice heavy lifting there, makes the castle in the first pic look colossal in the background.
|
|
#
?
Nov 10, 2022 06:52
|
|
|
Prompt was "A man named Frank Frank, photorealistic middle aged man, watching stir fry with noodles twist and untwist, Beksinski inspired" I made it for 🅾️🅱️'s thread, it is supposed to be 🅾️🅱️'s ghost twisting out of a stir fry and haunting/harassing/bewildering Frank Frank
|
|
#
?
Nov 10, 2022 15:25
|
|
|
Image prompting in MJ v4 is really nuts https://twitter.com/shambibble/status/1590055913800728577
|
|
#
?
Nov 10, 2022 16:01
|
|
|
"mario being rejected by princess peach"           
|
|
#
?
Nov 10, 2022 18:48
|
|
|
"post apocalyptic wizard wearing a leather jacket selling test tubes full of glowing liquid on the street corner"      
|
|
#
?
Nov 11, 2022 01:22
|
|
|
I haven't tried it myself, but someone seems to have finetuned SD1.5 with Midjourney v4 images, which sounds pretty interesting. https://twitter.com/rameerez/status/1590883316022284288
|
|
#
?
Nov 11, 2022 06:55
|
|
|
We need to come to with a better naming convention, these suck. Something as simple to remember as S03E12 S15M4?
|
|
#
?
Nov 11, 2022 07:09
|
|
|
Stealing Midjourney's style for free is incredible content
|
|
#
?
Nov 11, 2022 07:18
|
|
|
I find it funny there are many artists screaming for "ethical AI models" trained only on public domain art and art from consenting artists. That made sense 3 months ago, but now there is more tagged and indexed "greg rutkowski" art in the public domain than was ever produced by rutkowski himself. We can extract the artist soul from the AI art now
|
|
#
?
Nov 11, 2022 11:23
|
|
|
For a certain definition of 'funny', yeah. Poor bastards.
|
|
#
?
Nov 11, 2022 12:03
|
|
|
Rutibex posted:"post apocalyptic wizard wearing a leather jacket selling test tubes full of glowing liquid on the street corner" same
|
|
#
?
Nov 11, 2022 12:30
|
|
|
|
| # ? May 12, 2024 21:46 |
|
|
https://twitter.com/DeviantArt/status/1591113199218487300?t=Kmn-_vplmBbAExKupfX3tw&s=19 Melt downs galore! Apparently past works are opt out, which is pretty funny
|
|
#
?
Nov 11, 2022 19:41
|
|
