I'm going to ignore one of your rules >recent I haven't really generated anything in over a year, good luck finding the other 3 SFW genners among the refugees

>>6430 >recent Its no problem. Mine are from the last month or so as well. Just want to keep this board going. I kind of like it being separate from /g/ /t/ for now. Could you tell me a bit about your workflow in your 4th picture? The color and depth of information is really spectacular. <Pic I set up my comfyUI a while back with a for loop to automatically take outputs and apply a number of different preset color grading on them. Saves me a ton of time in post-production or image screening.

>>6463 I can post the workflow if you like but it's old, schizo spaghetti, and relies on a bunch of defunct custom nodes. The color part is from generating based on colored noise produced by the workflow with variable dithering between the colors and arrangements in terms of positioning, pics related a few more made with that same functionality.

4chan, if it comes back, should get an AI board. Even if I only use once in a while, when there is something new that I want to try on /ldg/. I am only interested in classical fine art styles.

I like nonhuman girls...

Really, really like 'em.

Also women of unusual size. RIP /gdg/.

Why let me select more than five images if that's the posting limit?

Fruits of today's genning.

>>6487 I do like this also, but it's hard to stay loyal when psychedelic-body horror-scifi gens are just a few words away. I guess that's how the compulsive porn genners feel too

It's honestly impressive what can be done with native image gen. Future is bright.

>>6734 >Necklace Giwtwm >>6832 >Centaurs WOULD What model and Lora are you using for such realistic images? Or is it a cloud service?

>>6516 >>6627 >>6630 >>6734 nice seeing you here brother, really loved your gens back in /gdg/

Posting only two since genning seems broken for me right now. Foreverially stuck on "preparing". >>6872 >What model "Native Image Generation" via Sora. Not sure it's still technically 4o since the release of o4. >>6886 Thanks, fren. Glad you like them. Maybe we can get a /gdg/ that does both anime and western out of all this chaos.

Sad image with a girl who lost her arm in war.

Everything is fire. >Connection failed.

>>7225 OUT OF PLACERS MENTIONED RAAAAAAAGHHHH How the hell did you prompt that? I'm really impressed!

>>7158 Your gens are pretty impressive. 4o is powerful. I can see how this tech could be dangerous.

I messed around a bunch with learning how to use video models. The outputs leave a lot to be desired, I haven't yet found a way to standardize colors when there is so much noise change in each frame. If I really wanted to I could just go in with aseprite and manually edit them, but I'd rather just use pixel composer to handle it all nicely. I'll update if I find a one size fits all solution. Tomorrow I'll try and get I2V working. Ultimately I want to be able to make dynamic animations from higher-quality sources that I can generate in advance by the thousands. Video is so fucking slow.

Made these with Sora earlier today.

Also made with Sora.

>>7232 Pure, unadulterated luck. I picked a semi-decent reference image and gave it the moral equivalent of "make real plox" to test the waters. Literally >Here's a reference image of a baxxid. Could you generate a photograph of it? Despite the misidentification of an eye as a nostril and an arm as a horn, that first result was the best. No further rerolls, prompt elaborations, remixes, or additional references managed to get anything better. I tried to get it to make a reference anatomy but that ended up as a disaster, too. GPT has a really, really hard time understanding baxxid it seems. Hopefully when/if they manage to get remixes to be more targeted, the lucky image can be fixed up. If anyone's got gemini imagen, it'd be nice to see how it fares with a fix-up request. Reference image attached as spoiler. >>7345 >>7346 lmao

Refugee here from Ai Broad Focus /aibf/, this looks like the nearest to our old site, and the title blurb implies Ai made stuff, so we mainly genned anthro stuff, typically SFW stuff 'cause it was coming from Dall-e, so wondering if the locals will have us...

>>7519 May or may not be best to start a new thread. I think most anons are fine with furry girls, but people who actually identify as furries tend to have obscure tastes, and the art generated by them tends to be even more niche. Like, what you posted would be mostly fine, but the exaggerated, cartoonish proportions and pixar stylisation dip into the uncanny disgust territory. It's a bit like those foot fetishists who post pictures of beanmouth cartoon characters with disproportionate size 200 feet shoved up in the viewport to fill the whole picture. The aura of mental illness ruins any possible enjoyment that might've been had, even if you usually enjoy /co/ and feet.

4o isn't terribly usable with references yet. I tried to give it a few Doom sprites and draw a composition schematic, but it still needs heavy prompt wrangling and step-by-step iterations on references. Controlnets are unironically more precise.

4o isn't terribly usable with references yet. I tried to give it a few Doom sprites and draw a composition schematic, but it still needs heavy prompt wrangling and step-by-step iterations on references. Controlnets are unironically more precise.

On the topic of furries, what does anon think of these? Are they fuckable? How would you change them to be more attractive? Or does furry only work in 2D? It's an interesting challenge. >>7563 Yeah. The issue with controlnets, though, is it feels like they're either too strict or too lax. Attached as a spoiler is a depth map from a Flux project, iteratively derived from a drawing I like. The depth map is inconsistent, and the finger anatomy is subtly incorrect. While the controlnets I used managed to understand the broken depth map, they really suffered with the broken anatomy. Lowering the strength just resulted in new kinds of broken anatomy as Flux struggled to understand the pose. But maybe it's a skill issue on my part.

What I found to be working well with 4o+references: >describe each reference in the prompt. It's not necessarily good at picking up all relevant features, but it's significantly better when you reinforce references with words >restyle each reference to be as close to the target style as possible, with a separate call >describe each object in the reference in the final composition >don't have more than 10 references, preferably no more than 5 if possible, it goes haywire Some guy reverse engineered 4o a bit and found out it's physically unable to track more than 25 objects in the scene (the 512x512 tile separation shouldn't matter as it's not resolution dependent). The manual says 10-20 max. Actual number is lower if you want precision. >>7576 For good depth maps I'm using Blender with a rig like this https://toyxyz.gumroad.com/l/ciojz or a similar one, but you can't fix the model itself it it works poorly with spatial hints (it might be a particular controlnet issue, or rather the dataset it's been trained on)

>>7584 >>7576 Also, looking at your depth map, it might be not just broken anatomy, but the fact that the fingers are blocked by the girl, and it perceives it as unrelated objects

>>7522 >'Furries' I hear what you're saying, and on the whole, the group I was with were not really impressed with 'fur suits' when we got them. We were typically more into Zootopia, Watershed Down, Duncan Wood, Red Wall etc, and trying to trick Dall-e into genning as risqué as possible, or emotive... >>7576 >Are they fuckable? Only in our dreams anon, and usually kept to ourselves (somewhat). Oh yes, I have a character or two...

>>7801 Hey fren, glad to see you're safe

Recently I've been using ChatGPT's image generator to generate realistic IRL versions of cartoon characters, which it seems really good at doing. Here are some of the examples I've recently got it to make (Nia from I Can't Sleep, Luz from The Owl House, and Andrea and Libby from The Ghost and Molly McGee)

>>7957 So someone mentioned another site, maybe with some /aibf/ refugees, but I'm damned if I can find it \oo/ Have some more Pip, in her assorted guises...

>>6430 >>6467 Love those

I feel like Sora recognizes a lot of celebrities.

I want to open, /medivgen/ - Medieval Tehemed AI General. Do you think it will catch on this board?

Today I am going to be testing Hi-Dream. I want to see if it understands art styles mainly. I am also open to suggestions on what to try.

>>8779 But before going for styles, let's make some adjustments to the defaults Comfy provided. I am trying to get a full body picture, and it won't follow: Bad. Trying to get a more natural "candid" photograph: also bad. It's still set in airbrushed studio photography.

I am going nowhere trying to work with the default prompt. Dump everything and work something else. I've tried this type of prompt before: one figure standing and another figure floating in the water. This isn't bad. The style isn't there (I was going for an old time polaroid).

Here are Flux and ChatGPT's results for the same prompt. Colors look better in both, but Flux struggles with concepts, even if the composition is interesting. Unsurprisingly Chatgpt got it right, the guy floating face down is funny.

>>7957 >>7979 I'm still new to this layout, but have stumbled upon a new(ish) /Anthro/ page, that will probably suit any /aibf/ (SFW) refugees; https://8chan.moe/anthro/res/1.html I've noticed when you're in a thread, if you look up top, you can see which board you are on (/ais/ here), but I can't tell where the above anthro thread is based \oo/ So if you're into Bing, come on over...

>>8787 Now for styles: this looks promising. First is Edouard Manet and Claude Monet, Second is pre-raphaelite Dante Gabriel Rossetti, third is arts and crafts Arthur Rackham plus Watercolour and ink. You can see there is understanding out of the box, even if the results are a bit bland. Flux needs Loras to do artist styles and ChatGPT does very poor imitations of fine arts styles. Give this a finetune ASAP.

>>8787 where are girl's legs, though? >Colors look better in both 4o has the distinctive yellow tint everywhere, your prompt has nothing to do with it. It's clear that this model is finetuned by AI bros who can't tell if/why something looks good (see the linework). Which is typical for OAI, Dall-E was always like this.

>>8797 what model bro? can it do pose?

>>8826 forgot to post the reference image for that pic

>>8813 I am aware of 4o piss tint, but I know you can correct it by talking to it. None got the exact polaroid colour, but Flux and 4o did better. >>8826 HI-Dream. Its prompt understanding is quite good, I think it might be slightly better than Flux,

Don't have disc space left atm, I will try it later, looks interesting

>>8917 How good is the model for inpainting?

>>8930 I am downloading the full model, atm. While running more tests. I can't make sense of how styles work. The model is very tight in composition (once it settles, small changes or just different seeds barely affect it), but style varies a lot between seeds. Maybe having negatives will help to settle style First Pic is chatGPT unrelated. Inpainting works.

>>8941 does it work for making things nude with the model? can you show your ComfyUI setup or your .json project for inpainting?

>>8946 >nude It called the Chinese police on me, thank you very much. Inpainting doesn't have a lot of secrets anyway. It takes two nodes. I set noise reduction to 0.7

>>8948 yeah that looks like my usual setup, shame it can do nude inpanting, because you genned nudes here. >>8797

>>8949 *cant

>>6487 Frequent visitor of the DALL-E thread here and agreed.

Had a slump. Lots of ideas not really panning out. Please enjoy these at least. >>7584 >What I found to be working well with Yeah, reinforcing the image with words seems to be key. Clever idea about restyling. >For good depth maps I'm using Blender with a rig like this Thanks, I'll keep this bookmarked for if I ever get back into local models. Hopefully the controlnets recognize the finger bones. But it might be a limitation of the model as you suggest. Like most, it seems to do poorly in these kinds of extreme circumstances. >>7585 That definitely wasn't helping things, but I couldn't find a way around it. Trying to prompt ex-nihilo for that particular style of grip was fruitless. >>7801 >350f422a7b6de78bfb0859319f65cf0415c4afc4dcdc5dd2b50b7927cf547bfa.jpg Sexo. >>7970 Don't know about the others, but IMO managed to capture Luz's dorky vibes. Good job, anon. >>8635 /ais/ is still much too slow. Just look at this thread.

>>8415 I wonder if Sora's looseness with celebrities is going to last. I know that Dall-e went into censorship overdrive because of Taylor Swift.

>>7998 Would you mind sharing your workflow/model used? Those are very impressive sprites, I always find it difficult wrangling open source models into not just genning women. right now I'm testing this new vae for SDXL pixelart. so far it's done a fantastic job at keeping "pixels" in a perfect grid of 8x8 true pixels. Saves me a ton of post-processing work.

>>9278 You need to use Flux, Image Pixelate, prompt for a 2x2 grid for smaller items or emphasize a focus/central image for larger items (e.g. armor). Then it's just a lot of lottery and lora mixing. Here's an older one I did for simpler RPG item icons to get started, the lora's I used should all be available on civitai but if not, I'd recommend just recommend trying a bunch with different weights until you find one that hits your particular style. https://files.catbox.moe/skogif.png

So I've got some OCs I've made LoRAs for on Civitai, but I'd like to see what other non-SD systems can do with them. I've made plenty of reference gens but don't know where to start. Any advice?

>>9618 Aren't Loras specific to models?

>>9684 Yes, they are but I'm asking whether any of these non-SD systems can take a set of images.

>>9690 Here's one I've played with; https://www.krea.ai Gives you something like 10x4 gens per day free, after signup. Has an interesting feature of being able to upload a sample image, and choose up to 5 different styles. The fun bit is that with ALL the styles and the uploaded image, you can click on the 'buttons', and slide the amount of influence each button has... I've seen this on a couple oof other sites as well. >>9690

Made with Sora.

>>9954 Yesterday ChatGPT threw a fit because I named Bill Waterson in a prompt, the inconsistency between Sora and ChatGPT is really annoying.

Tradwife Stocking Tradwife Stocking

>>9957 Yes, ChatGPT is really frustrating and inconsistent. Before the 4chan hack, I was feeding it old LucasArts game screenshots and posting the results in /v/, but for some reason this one particular screen was setting off some kind of flag and it refused to do anything with it. It said it was because it's from a copyright game, but so was every screenshot I had given it previously.

>>10609 >last pic Very based

https://openai.com/index/image-generation-api/ they added imagegen to the API

>>10993 >https://openai.com/index/image-generation-api/ Don't know if I've played with it in the past, but it's not letting me try the free trial \oo/

An assortment of random pixel art gens I made today. Tried a few different models and loras. So far nothing has beat illustrious for my use-case.

>>12759 Could you catbox the one of the moon? They're very nice

>>12760 Here ya go anon: https://files.catbox.moe/lfux23.png A few things: 1. You'll need this node pack for the palette swap and pixel crunch portion since I use a for loop to make my palette sheets. There are probably better ways to do it, but fuck it, it works. (You can also just disable it and do it yourself using my pass muting selector): https://github.com/BadCafeCode/execution-inversion-demo-comfyui/tree/main 2. In return, I'd appreciate if you or any other anons in this thread can give advice on my workflow and where it could be improved. I don't know anyone else who uses ComfyUI, so I just pick up what I can from tutorials and "fuck it we ball" prompt alchemy. 3. Did they fucking ever fix the frontend changes with "get/set" nodes? My workflow was way prettier before, but I had to rewire everything a few days ago because the latest frontend patch bricked rgthree and useEverywhere.

>>12762 Awesome, thanks! The palette swaps look really good too, will try and incorporate some of that into my own. I'm not much of a Comfy expert desu, I just take others' workflows and tweak stuff, I'll take a look after I go to bed. I try to never patch Comfy once I've gotten a decent flow going, not worth the hassle of things breaking.

Some new stuff I made today. I really love this dapper motherfucker with the top hat. I hope anons continue to keep this board alive, even if its slow. I dislike 4ch/g/, too fast for my tastes.

>>13146 >I hope anons continue to keep this board alive Ok, going to post some things i did today. I originally wanted to make "the sun being eclipsed by a ship" but got this horrors instead. So i decided to expand this idea with the prompt "orange desert stuff". It is late now and i am starting to regret my decisions kek.

4o can't do THIS

>>13146 Is there a way to prevent bitmixing in AI generated pixel art?

>>13374 oh vey that's racist!

check out the workflow on this. this pic was genned in 10 steps at CFG 1 and has a working negative prompt.

>>13393 This site does not save metadata. Upload it to catbox or something.

>>13394 oh shit ok https://files.catbox.moe/2zpp82.png https://files.catbox.moe/vovrw4.png

>>13380 There are three steps I apply nowadays to help ensure more consistent pixel art. The first step is fairly new, but I've been in love with it. I use this vae: https://civitai.com/models/1492329/pixelate-x8-vae-for-sdxl it really helps constrain generations into an 8x8 grid. If you don't use sdxl models (or illustrious), you can still get good outputs by sending your generations through down scaling and a palette-swap. For example, the first image here is taken straight from the engine, the second is after I've applied 0.125x nearest neighbor down scaling and a selection of handpicked colors I enjoy working with. Its all done within comfyui itself so its a simple one-click process. You can also use a k-means algorithm for better downscaling, but I find that it doesn't produce significantly better results compared to the increased CPU load and time. The only thing I'd say not to use are any dedicated pixel art models. I've tried most of them, and they all fall short of providing quality outputs.

I didn't invent a new fetish with these did I? Is there a name for this?

>>13540 Oil drilling.

>sfw Sorry, I only generate porn. Anything else isn't worth my time or I just try to make based on my own skill.

>>13540 "Humanoid food" is your best bet. I'm getting addicted to these, honestly. It's been 3 months, but hear me out...what did you use to gen this?