|
 Postin'
|
#
?
Nov 12, 2023 21:03
|
|
|
|
| # ? May 30, 2024 14:50 |
|
|
mcbexx posted:Style prompt pleeeeeease. "1980s VHS cover artwork movie poster for a dark fantasy horror film about [X] --ar 2:3" Edit: have some B roll       
Kosmo Gallion fucked around with this message at 21:40 on Nov 12, 2023 |
|
#
?
Nov 12, 2023 21:30
|
|
|
pzy posted:
(Points at thread subtitle) 
|
|
#
?
Nov 12, 2023 22:07
|
|
|
"the best computer window layout" thrown at the new dall-e-3 API, which massages the prompt like chatgpt and doesn't tell you why you end up with cathedral architecture
|
|
#
?
Nov 13, 2023 01:54
|
|
|
you can get the prompt it ended up using for that image, on the mobile app you hold down on the image and can copy it, not sure on web. it might have some specific style stuff
|
|
#
?
Nov 13, 2023 01:59
|
|
|
Did Clipdrop stop their free tier? When I try to submit a prompt, I just get the 'Upgrade to Pro' popup, and nothing is rendered.
|
|
#
?
Nov 13, 2023 08:55
|
|
|
       
|
|
#
?
Nov 13, 2023 10:25
|
|
|
those rule
|
|
#
?
Nov 13, 2023 11:43
|
|
|
Badass.
|
|
#
?
Nov 13, 2023 14:15
|
|
|
            
|
|
#
?
Nov 13, 2023 16:07
|
|
|
 Neat!
|
|
#
?
Nov 13, 2023 16:09
|
|
|
Awesome stuff! I'm also diving into a little project of my own and I'm looking into using AI tools to help out. I've seen some cool videos where the AI can draw over your sketches to make them better. Does anyone know any tools like that or something in a similar vein?
|
|
#
?
Nov 13, 2023 16:11
|
|
|
       
|
|
#
?
Nov 14, 2023 17:03
|
|
|
Monglo posted:Awesome stuff! Depending on when those videos were made changes things. Right now the most of that ability you are talking about wouldbe Control Net within Stable Diffusion, different options with Control Net 'see' your initial image in different ways. Or you could always just use img2img playing with the options to see how far you want the image to deviate from where it started with a combination of prompting as well. There is no direct answer or place to point, it's all a lot fuzzier and mixed up than that. Just know that yes, absolutely, you can do what you ask.
|
|
#
?
Nov 14, 2023 17:52
|
|
|
 
Archer666 fucked around with this message at 03:43 on Nov 15, 2023 |
|
#
?
Nov 15, 2023 02:43
|
|
|
          
|
|
#
?
Nov 15, 2023 05:29
|
|
|
"psychedelic risograph" doing some heavy lifting in MJ    
|
|
#
?
Nov 15, 2023 12:47
|
|
|
What's the latest consensus on Dalle Vs Midjourney? I only generate images in chatgpt, but do find them often looking not as varied as the ones posted here. Not sure if I'm just bad at prompting, or it is the model difference.
|
|
#
?
Nov 15, 2023 14:20
|
|
|
Monglo posted:What's the latest consensus on Dalle Vs Midjourney? Midjouney is better at just creating beautiful or photorealistic images, but dall-e is absolutely amazing at creating things according to your prompt and can include multiple things in the image without just blending them together Midjouney will generate images of copyrighted stuff without having to debate a computer though
|
|
#
?
Nov 15, 2023 16:32
|
|
|
Midjourney will try and generate copyrighted stuff but their system definitely has it filtered out of the data set in some way so it just ends up generic. Through Bing (on DALLE) you can make actual legit looking copyrighted stuff  Vs  I think MJ has honestly fallen behind in basically all regards at this point, the ability to do more complex ideas intentionally is really important
|
|
#
?
Nov 15, 2023 16:39
|
|
|
speaking of copyrighted characters, here is one my buddy sent me recently 
|
|
#
?
Nov 15, 2023 16:45
|
|
|
Midjourney has a cool new(ish) Style Tuning feature where you can stick in a prompt and have it generate 64-256 different images based on the prompt, then you select up to half of them to guide the visual style on future prompts. It lets you reuse the same art style across multiple prompts without having to insert a bunch of superfluous words that can affect the output in a way that other AI can't (I guess it's sort of like a LORA or whatever from SD though). But right now they're very focused on getting a web interface up and running and ditching Discord, as opposed to releasing a new model, which has caused it to fall behind a bit. DALLE has always (even DALLE2) been way better about creatively interpreting prompts. deep dish peat moss fucked around with this message at 16:49 on Nov 15, 2023 |
|
#
?
Nov 15, 2023 16:45
|
|
|
Yeah, I rarely use MidJourney these days, there's just not enough control over the output. The unpainting feature is fantastic, but the coherency is just so much lower than Dall E. The project I'm working on requires using the likenesses of actors, which ChatGPT gives me guff about, but I can usually get it "close enough" then bring it into SD and regenerate the faces. Pain in the rear end to both keep the style from the DALL E image and change the actor likeness. Definitely some seams here and there that I have to manually patch up or just hope nobody notices.
|
|
#
?
Nov 15, 2023 17:16
|
|
|
Monglo posted:What's the latest consensus on Dalle Vs Midjourney? DallE is super great at text and being specific, but it's quite bad at photos now, and hyper, weirdly restrictive with no real rhyme or reason because that's how censorship has to work with these systems. I've grown to really dislike DallE, working with it through ChatGPT itself. It is bad at art, it's awful at photos but it's great at context. If you want to make funny memes, DallE 3 can't easily be beat with one-shot prompting. Their clamp down on doing one image at a time (instead of four like MidJourney) really blows though and slows down a lot of the fun since you can't compare images directly like with MidJourney. At ~art~ itself MidJourney is clearly the winner, by far it's the most deep in so much as how far you can explore the latent whatever the hell these models are built on. It's the prettiest one of the bunch. If you learn it's language then MidJourney is the most capable of making new and interesting images. Stable Diffusion is behind both of them when it comes to one-shot prompting, but is the absolute king of controland granularity. If you are making things in a more traditional sense then it's Stable Diffusion or nothing. Inpainting or outpainting at whatever detail you'd like, integration directly with photoshop, an infinite torrent of models and loras, open source. Plus unrestricted in its output and ~free~. Here is some prompt comparrisons "a sheet of paper with the word "ARTS" written on it in red marker, with a doodle of a dog"  DallE  MidJourney  SDXL DallE nailed it perfectly, which is the strength. MidJourney kind of lost the plot, while SDXL is actually decently component. The trick though is I can work with SDXL and refine any and all parts of the image easily. Less so with MidJourney though the tools are getting better if they'd completely their divorce from awful, awful discord. DallE though? That's it. Good thing it's exactly what I asked for. "1975 retrofuturistic surreal photo featuring a Sassy 70s dinosaur"  DallE  MidJourney  SDXL  SDXL but hand-refined This illustrates exactly how DallE isn't really great. I asked for a surreal photo, yeah? But for some reason DallE inerts a bunch of extra crap in the prompt and hyper-focuses on other aspects that muddy what I was asking for. This isn't really a photo, though the dinosaur is pretty sassy and photo-like. The fact that I can't tweak this besides begging ChatGPT to refine it makes it difficult to 'work' with it. MidJourney smashes it perfectly imho, that poo poo is a photo that looks like it's from 1975, is a dinosaur and is sassy. Maybe a woman in a big rubber dinosaur mask, whatever! It's surreal! And SDXL did pretty good, but as you can see just a few really easy and quick tweaks makes it a much less AI looking thing, mostly the faces. Less rich than MidJourney, however. DallE for funny meme images and not much else, plus very free through Bing MidJourney for pretty and varied art and exploring latent spaces. Stable Diffusion for porn. Also the only real art tool of the bunch. They all have their own models and their own language they speak which makes direct comparison difficult. Most of the time it's going to be DallE because of how accessible and cheap it is. All of these things use text interpreters but they are so much more than that. It's going to be really difficult to beat typing in a sentence into DallE and getting a passable image. Really, the king poo poo mode of doing AI image stuff is making your base image in DallE or MidJourney then 'refine' it in Stable Diffusion. feedmyleg posted:Yeah, I rarely use MidJourney these days, there's just not enough control over the output. The unpainting feature is fantastic, but the coherency is just so much lower than Dall E. Most people don't notice besides assholes like me KakerMix fucked around with this message at 17:43 on Nov 15, 2023 |
|
#
?
Nov 15, 2023 17:25
|
|
|
Thank you for the detailed answer, KakerMix! Going to put some effort into figuring out SD, the toolset is clearly the best for actually working with images. Cant wait for more user-friendly interfaces though.
|
|
#
?
Nov 15, 2023 17:49
|
|
|
Monglo posted:Thank you for the detailed answer, KakerMix! Going to put some effort into figuring out SD, the toolset is clearly the best for actually working with images. All this stuff moves so, so fast, BUT as far as I know it's still Auto1111 and the photoshop plugin. It's the most serious use of these tools in so much that it's Photoshop, but with SDXL running within it. Avoids ComfyUI's spaghetti and Auto1111's archaic and weirdo interface as well. Just Photoshop and a hyper-powerful generative fill. There is a learning curve what with drawing bounding boxes and checking the resolution you're running at with the plugin and masking, but it's all adapted from general photoshop work if you're familiar with it. https://github.com/AUTOMATIC1111/stable-diffusion-webui https://github.com/AbdullahAlfaraj/Auto-Photoshop-StableDiffusion-Plugin#how-to-install This makes it so your photoshop looks like this: 
|
|
#
?
Nov 15, 2023 17:57
|
|
|
I reregistered to MJ to try out the style tuner but haven't managed to make one I like enough yet to want to keep messing with it. The combinations I have picked all end up largely unable to generate subjects other than the one they were trained on. I do really like the idea of it, but I dunno, maybe need to read a guide or something.
|
|
#
?
Nov 15, 2023 18:09
|
|
|
I think part of it is that you need to tune on a prompt similar to the final output. And also don't set high --s values, which push the image further away from the tuner. For example here's a style tuner for making planets that are grounded in reality: --style ik58VsJpTPLxWAFc The style tuner prompt was "planet seen from orbit. Majestic, fantastical, exotic, anomaly." and I selected all the darker, more realistic looking outputs An exotic and vibrant alien bioluminescence planet, seen from orbit. --style ik58VsJpTPLxWAFc --s 50  A planet overtaken by alien biomaterial, seen from orbit. --style ik58VsJpTPLxWAFc --s 50  The same prompts without the style tuner:   The tuner let me push the results more towards looking like actual planets instead of colorful fantasy illustrations, but it's ultimately still making the same type of subject matter. The tuner just let me push that planet generation prompt away from the weird ones. It reduces the variance in those prompts away from the heavily-stylized images that don't match each other stylistically. (And conversely if you wanted the colorful stylized planets, you could push the tuner further in that direction) If I create another prompt that isn't a planet and apply the same style tuner it doesn't come out looking like a planet and it doesn't exactly mimic the art style of the planets but it does keep the same sort of darker, subdued, grounded tone: Prometheus stealing fire from the gods. --style ik58VsJpTPLxWAFc --s 50  Prometheus stealing fire from the gods. --s 50  But you still have to be kind of careful with your prompting. This prompt includes the word "forest" which tells MJ that there should be trees: forest planet seen from orbit. --style ik58VsJpTPLxWAFc --s 50  Even with the tuner, you get these really fake-looking planets because it inserts the trees. So calling it a "lush" planet instead gets the intended effect: lush planet seen from orbit. --style ik58VsJpTPLxWAFc --s 50 
deep dish peat moss fucked around with this message at 18:38 on Nov 15, 2023 |
|
#
?
Nov 15, 2023 18:36
|
|
|
       
|
|
#
?
Nov 16, 2023 01:26
|
|
|
One cool thing about image generation is being able to think of a stupid idea and turn it into a t-shirt in 10 minutes Bing dall-e image: "A t-shirt featuring an illustration of a burly, hairy man, inspired by the 'bear' community within LGBTQ+ culture, doing a yogic pose. "  Cleaned up eyes with Photoshop AI and added text  Cropped, background removed, AI upscaled with Photoshop  Dumb shirt: https://www.bonfire.com/yogi-bear/
|
|
#
?
Nov 16, 2023 06:28
|
|
|
joe football posted:Midjouney is better at just creating beautiful or photorealistic images, but dall-e is absolutely amazing at creating things according to your prompt and can include multiple things in the image without just blending them together MJ definitely needs to get on the GPT4 prompt train. I feel like I'm going back to DOS when I use it now. Image tax: I was impressed how well DALLE was able to capture the essence of every single Ubisoft game cover art ever made. I had to edit out the gun, though, my character doesn't use them. It kept adding guns even when I asked it not to because video game = guns I guess. 
Roman fucked around with this message at 15:25 on Nov 16, 2023 |
|
#
?
Nov 16, 2023 15:19
|
|
|
Roman posted:Most of my stuff right now is pre-pre-vis for characters and settings in a screenplay and DALL-E is so much better at it it's crazy. DALL-E makes my characters with the exact clothes I describe every single time while MJ struggles with "green jacket and red backpack." I may actually cancel my MJ sub for a while. This is where I'm at. The amount of specificity I can get with DALLE means I save a TON of time in Photoshop refining and editing.
|
|
#
?
Nov 16, 2023 15:23
|
|
|
I think that telling an image generator to draw "[character] in the style of a Roy Lichtenstein painting" is a good way of gauging how well it can do a certain style with minimal instructions whether or not it knows the character is another matter entirely as seen with these two sets, "Midna in the style of a Roy Lichtenstein painting" and "Imp Midna in the style of a Roy Lichtenstein painting"         I didn't ask for a majora's mask crossover but it's great that it gave me one, I'm definitely using the first imp midna's design for a project in the future zetamind2000 fucked around with this message at 15:48 on Nov 16, 2023 |
|
#
?
Nov 16, 2023 15:45
|
|
|
Is dalle3 access now universal for all gpt4 subscribers or is it still in roll-out?
|
|
#
?
Nov 16, 2023 16:10
|
|
|
hydroceramics posted:Is dalle3 access now universal for all gpt4 subscribers or is it still in roll-out? As far as I know, it's for all GPT4 subscribers. That said, since the GPT rollout they had to slow down subscription purchases because usage is at an all time high.
|
|
#
?
Nov 16, 2023 16:13
|
|
|
The Bing chatbot (which is free) uses dall-e 3 whenever you ask it to generate images, but I don't think it has the extra tools (e.g. inpainting) that standalone dall-e has.
|
|
#
?
Nov 16, 2023 16:17
|
|
|
Tarkus posted:As far as I know, it's for all GPT4 subscribers. That said, since the GPT rollout they had to slow down subscription purchases because usage is at an all time high. Good to know. I guess that's why they choked off returns to a single image per prompt. Image tax: MJ is getting slightly better at making a red panda with blue fur: 
|
|
#
?
Nov 16, 2023 16:17
|
|
|
hydroceramics posted:Is dalle3 access now universal for all gpt4 subscribers or is it still in roll-out? I believe it is part of base GPT4 now it no longer has a drop down for websearch or images. You can just open a normal GPT 4 and say go look for news and make me an image that shows what happened so far today, make sure to not violate ToS with the prompt we want a successful image so be careful with the prompt. Today is 10/16/23 I mean it might have the date I just included it without trying without it first pass "I was unable to generate an image based on today's news events due to content policy restrictions. The events of the day include complex and sensitive topics that are challenging to represent in a single image while adhering to these policies." "Please provide the prompt" oh I see you are using names! Don't use the names use president of country name instead  did the US and china agree to a joint space program or something?
|
|
#
?
Nov 16, 2023 16:18
|
|
|
Huh, Bing has started watermarking images again, which is kind of a pain in the rear end. Weirdly, they appear on images created in browsers (I've had them in both Firefox and Edge) whether you use Image Creator or the GPT-4 Copilot, but - so far - they don't in phone apps when you save the picture even if the watermark is there in the preview. Another bit of Microsoft competency and consistency is that there are now two different Bing watermarks, and it's totally random which you might get. Edit: something else I've noticed is that you can now use Copilot to create images in both Edge and Safari on a phone, which means you don't use any boosts in Image Creator. Edge gives you 30 prompts (rather than just 5 in Safari) per chat, but it also eats up your battery and makes the phone run hot, so lol Microsoft again. Edit edit: if you really can't face Photoshopping the watermark out, you can click on 'customize' in a browser to send it to MS Designer and get a clean download option there. Small Strange Bird fucked around with this message at 17:25 on Nov 16, 2023 |
|
#
?
Nov 16, 2023 17:07
|
|
|
|
| # ? May 30, 2024 14:50 |
|
|
hydroceramics posted:Good to know. I guess that's why they choked off returns to a single image per prompt. GPT4/DALLE is weird with restrictions. I made a bunch with a woman holding an assault rifle and standing in flaming wreckage but when I asked to make the same woman in a suit with a holstered glock in the same session it was all "GUNS BAD SORRY."
|
|
#
?
Nov 16, 2023 18:46
|
|