General dedicated to creative use of free and open source text-to-image models.Previous /ldg/ bread : >>101166584Daily Edition>Beginner UIFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studioEasyDiffusion: https://easydiffusion.github.io>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiStableSwarmUI: https://github.com/Stability-AI/StableSwarmUIInvokeAI: https://github.com/invoke-ai/InvokeAIComfyUI: https://github.com/comfyanonymous/ComfyUI>Auto1111 forksSD.Next: https://github.com/vladmandic/automaticForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeAnapnoe UX: https://github.com/anapnoe/stable-diffusion-webui-ux>Use a VAE if your images look washed outhttps://rentry.org/sdvae>Models, LoRAs & Traininghttps://aitracker.arthttps://civitai.comhttps://huggingface.cohttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiTComfy Nodes: https://github.com/city96/ComfyUI_ExtraModels*SD.Next also works with PixArt-Sigma>Animationhttps://rentry.org/AnimAnonhttps://rentry.org/AnimAnon-AnimDiffhttps://rentry.org/AnimAnon-Deforum>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>View and submit GPU performance datahttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest>Share image prompt infohttps://rentry.org/hdgcbhttps://catbox.moe>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
bread thread
>mfw
>chang woke up nice
Any good XL model for anime that gets you actual anime style? (No korean-mahwa AI style).Tried Pony XL because everyone said it was great but the result is always some western, deviantart tier, SD1.5 base model tier shitty result, pic related.
Hey I got on the OP image, nice
>>101183562>getting filtered that hardJust tag a manga artist for style. https://pastebin.com/qDRXFfBM
official pixart bigma waiting room
lumina next preprint is upLumina-Next: Making Lumina-T2X Stronger and Faster with Next-DiThttps://arxiv.org/abs/2406.18583
>>101183035
>>101183035>>101183685>>101185054nice
>>101184931I hope it runs with 12gb
>>101186180good girl
>>101186239she is
>>101185419imagine owning a 3060 and considering using local AI. kys, or but a a6000 or up povo
>>101183562Try AutismMix. PonyXL gives rather random styles without LoRAs.
>>101178800sauce anon pls! this is beautiful!!
Whoa this thing has grown, last time I checked there were a lot less rentry links and one or two UIs lolShould the sticky have something about what's the "best option"? I'm looking to update my SD1.4 (yeah, it was enough for my uses ok?)Should I go with AUTO1111 docker and SD2.1, or are you guys using something different?
>>101188403 AUTO1111 and SDXL is good ... for a start. Not so fragile it needs docker: install, download SDXL model and go.
>12 hours old What happened?
Is there any way to stop pixart sigma from blurring the background? I fucking hate blurred backgrounds and absolutely nothing I've tried has stopped it from doing that
>>101188559the spammer left, imo i think it's better when these threads are slow.
>>101188611Depth of field in the negatives, maybe.
>>101188619fast, slow, doesn't matter. as long as team psycho is not in attack mode and manufacturing drama.
>>101188611avoid using tokens that have those characteristics attached. for sigma, try some of the dalle3 tokens
>>101188441Thx fren. For more than a start? I'm ok with rentries, it's just that there's so many that I don't know where to start lol. I see one guy posting goid images with "Sigma" in filename, I'll CTRL+F for Sigma
>>101188976Check out sigma here : https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma . It's different and has different results. I can run it from command line, but ComfyUI claims to support it. I don't thin Automatic1111 (yet). Someone else may know more about UI program support for it.
>>101189080sd.next (vladmandic automatic1111 fork) and comfyui support it, yes
>>101180452few musclemen added, ty
>>101183035Set jpeg to 100% quality. Got more artifacts than a museum
>>10118968895% is usually good enough to fit these resolutions into 4chan's limits
>>101189723811 KB OP pic vs 4 MB upload limit
>>101189688Whoops. I should update my ffmpeg alias.
>>101189688pixart is not aesthetic enuf
>>101183562any number of pony merge checkpoints or style loras will solve that
>>101183562also are you using source_anime in your prompt?
>>101188157some 1.5 merge I made that is old as Jesus + test lora
>>101183562>pic related.Did you get that with pony? That's kind of impressive, catbox?
>>101188403A1111 or Comfy are both good choices, if you aren't confident, start with A1111you don't need docker, and SD2.1 is deadSDXL is good generally, PonyXL and derived models are good too but less general, exceptional at 1girl and booru stuff though, learn how to score_9 tags work if you choose that
>>101191088ngl, this one looks kinda legit
training status?
>>101191765>sports car on the roadPeople are also starting to have faces.Also bought another 16 TB hard drive so I can download Pixelprose and Coyo HD 11m. Coyo is interesting because I'm doing their short caption + search caption + booru tags. Search caption is good because it often includes proper names although typically dog shit.
>>101192148HYPE
>>101192148why is the image 256 and not 512
>>101192507Pretty sure that's the resolution this 1.3B is currently getting trained on from scratch on two plebian GPUs on some million images.
>>101192507Because a new model is better progressively trained from 256 to 512 than throwing it in the deep end, at 512 there's 4x the possible choices the model has to choose from. honestly curious if it's more efficient to do intermediate sizes like 384 so that there's less of a gap.>>101192558Even if I had access to H100s I would do the same thing (although the model would probably be 2B). The Pixart paper proved there's efficiency scaling up.
>>101192558why shit on people who try things>>101192583Intredasting. Aren't you worried that they will release some better base model before you have finished the finetune?
>>101192830>Aren't you worried that they will release some better base model before you have finished the finetune?nta, but trying new things is never a bad thing. there are few people trying to finetune a model, and even less sharing anything about it.
>>101192830I hope a better model is released and I hope it further pushes the efficiency of limited resource training. But I don't really expect to see anything before the end of the year. And at the end of the day this is just a fun hobby project I'm working on and encouraged me to get my data hoard together because it's likely in 5 years sitting on tens of millions of images will be gold.>>101192885This isn't even finetuning technically, this is a foundational model trained from scratch.
>>101192583Good idea to verify the dataset and training, sure.
>>101193359
>>101193374
>>101183489It must be painful to have>oOoO>ooOOo> OOoofor feet
>>101192830>why shit on people who try thingsI don't think I did.
does anyone know where i can find a simple regional prompter workflow for comfyui? the ones im finding are completely broken, overengineered and have a ton of features i do not care about and doubt even work with ponyxl.
Ready for the 4th
>>101196084Nice
>>101194802this was really cool btw>>101196193ty
Good night. See you soon
>>101196266gn
>>101168532>>101168594Is the liminal spaces anon still online? Catbox?
>>101197224i wish
>>101197847PS1 nostalgia
>>101198095Isnt sgm for lightning models
>>101198049>>101198073>>101198095Prompt would help to understand these grids desu
>>101198122>Prompt would help to understand these gridsI'm testing impact of steps on coherence of the sampler for visual artifacts like extra/missing limbs, or any other inconsistencies. The prompts are kept the same and irrelevant, hence no mention of it. When you're looking at a grid, chances are everything is kept the same, except for whatever is actually mentioned. At least that's how it should be.>>101198110>Isnt sgm for lightning modelsGood question. Seems to be the case, but whatever it's original purpose was, it yelds good and interesting alternatives to the basic Euler in non-lightning models.
>>101198198>At least that's how it should be. I misread the first example as euler a which, from my understanding, never fully converges. My reply was more like "did the prompt reference fishnet?" since it disappears after 25 steps but you're probably right. I'm happy to see grids either way
>>101198242>did the prompt reference fishnet?nope, and in your defence, that's actually an interesting point to keep in mind, I do now wonder how much do steps help with prompt adherence.
What's new in the /sdg/ space
>>101198383/ldg/**
>>101198383>>101198392Very slow days lately, both for this general and for the tech in general.
>>101198411Any idea why?
>>101198392hi chang
>>101198424Could just be summer, people busy irl with vacations. I'd expect things to pick up the pace a bit once something like the next Pony relases. Since SD3 was a flop, everyone is probably waiting for any breakthrough alternatives, hence the relative silence. Even /sdg/ seems to have calmed down, relatively speaking.
>>101198424https://www.youtube.com/watch?v=gcB4ay_fO1kno hurry
>>101198501Yeah it's the eternal waiting room
>>101199497better delete, middle pic is too revealing
>>101198424people are just in waiting mode. the pixart guys said they are working on a bigger model and i expect that to take a few months at minimum.
>>101188611"bokeh" in the negative?>>101191112Thank you! I already used 1111 (without docker) for a year at least, and I'm not afraid to try something different. My main use case is inspiration for drawings, or hitting the ground running with a composition to fix up and modify to make what I was planning, so that generally does not include anime/toons, although I can make do with that as well.For this reason I was liking the PixArt-Sigma generated images, although in the end yesterday I didn't manage to actually get around and install/test anything.
>>101199767I'd advise against comfy for artwork related workflows. It's good for experimenting and learning how txt2img models works, but it's abysmal to get anything actually done. Either stick with auto1111, or try something that uses comfy as a backend, StableSwarm or Metastable to name examples. As for models, just like the other anon mentioned, you probably want to go either for SDXL or PDXL/Pony. For all I know, Pixart is currently supported only by ComfyUI with that ExtraNodes addtiion, or SD.Next branch of Auto. I.. think you could also get it running with StableSwarm if you manage to apply that ExtraNodes macguffin, but I'd need someone to confirm that.
Caption dropout seems to work really well for style loras. Dropping every 3rd epoch now
>>101199933What base?
>>1011999431.5 mix
>>101199879> it's abysmal to get anything actually done.I was in fact afraid of that. I suppose I'll try SD.Next then, seems more familiar. Downloading weights currently, so I'll try to set it up
For all my tests with samplers, think I'm leaning towards these four in terms of quality and originality.What do you folks use these days?
>>101200431euler_pp
>>101200630Very nice
>>101200922ty, nice gen yourself. Love the gradients
>>101201083Thx
>>101200630very lovely image
>>101196084Prompt?
>>101201160ty>>101201242It was a happy failure"A painting of a face with smoke behind it"
>>101201300You really just got lucky on it?
>>101196084>>101201300>>101201350Yup, it put the smoke in front
I love how sigma does machinery in general, wires, coil springs etc
>>101201402wtf anon is that from the base model on spaces?
>>101201369Still nice.
>>101201402Okay yeah, this kinda slaps.
>>101201427naisu
>>101201419Yeah. I've been genning and turning them into loras>>101190654>>101191141>>101191204Here you can see some machine parts, 1.5 with said loras
>>101201448You're not using (sigmax ** (1 / 7) + y * (sigmin ** (1 / 7) - sigmax ** (1 / 7))) ** 7on your PixArt?
(sigmax ** (1 / 7) + y * (sigmin ** (1 / 7) - sigmax ** (1 / 7))) ** 7
>>101201482Only custom node is ExtraModels because I'm afraid of STD's. euler_pp @ 3CFG/50 steps is amazing. normal/simple/sgm_uniform seem to work well
>>101201516It just hits different. But I understand.
>>101201477That's pretty cool btw. Most people advise against using synthetic data, so interesting to see it working. Any downsides so far?
>>101201659>Any downsides so far?Well the base model + clip is showing age. Lack of expressions, lighting and colors are systematically wrong. Having to manually clean the dataset images can be tedious. That's pretty much it
>>101201659>>101201482https://files.catbox.moe/4a43m2.pngSome new tech
>>101200922>>101201109Cool gens
>>101202306Thanks
>>101197224Oh wow that's cool, I love it.>>101197642Aaaaaaaaaaaaaaaaaaaaaaa
Okay, so basically Euler A SGMUniform behaves like DDPM, except sliiiiightly better with small details.
>>101202835I always go back to same old 2m karras / euler a
>>101203116Yeah, the more I mess around, the more enamoured I am with Euler A. Never had luck with any Karras though. From my experiance and testing they tend to have more inconsistencies, extra limbs, fingers, weird line merging or shit going nowhere, etc.
>>101203157>Never had luck with any Karras though.Very sensitive to high cfg. I use this https://github.com/mcmonkeyprojects/sd-dynamic-thresholding with 4-5 mimic cfg
>>101201784Was hoping it would magically solve typical SD problems. Oh well>>101202095>https://files.catbox.moe/4a43m2.pngThat's a good double exposure blend! Haven't tried CFG rescale either
>>101203243>Very sensitive to high cfg.Oh! Never even considered this.
>>101204111the end of bowling
>>101204111>>101204198Bowlingelion
>komm süsser todit's owari
its time you guys gave up and came home
>>101204403less drama in the provinces
>>101204403Home is where the heart is
>>101204850>>101204932good stuff
https://www.youtube.com/watch?v=-pr-WUa8eEs
Anyone got the retro diffusion leak and Aesprite plugin? Not paying 60$ for this shit.
>>101183035>2 days agoLocal would be better together...
>>1012054202 dog days are approximately equivalent to 24 human days
>>101205420brother, we don't have to fight. They promise you freedom and offer you API's. Join us in the local holy land
>>101205474cool
>>101205420its actually much cooler
>>101205641thx
>>101205021badass and a cool style >>101205522nice
I like it here, calm cozy no drama.its slow. but its homely
>>101206326>cozy
>>101183035new to ai but I'm looking for one where I can input my gf pics and make it output pics of my gf with my prompts. which one let me do that
>>101206712You can train lora with those pics. Local model 1.5 will let you do that
>>101201482whered you find that equation?
I don't know wtf my lora is learning anymore...
>>101207700hunyan lora?
>>101207912yeah
Say I have placed a cutout character from a different picture to a new background.How can I make this character blend? Maybe light is coming from the right and I want it to shine on the body
Could someone tell me how to animate a picture? asking for a friend.PS: no complex geek shit, just a website where you can upload an image and it will animate it without artifacts and bullshit.If AI can't do this then it's shit.
>>101206521He's too close to the flame!
>>101207267https://github.com/Extraltodeus/sigmas_tools_and_the_golden_scheduler
newbie here. where do you put these settings?
>>101208297>If AI can't do this then it's shit.>no complex geek shitthis is too new, dude. give it 2 years.
eyes
>>101208876But there is a new website that is doing these crazy meme pic to videos already, what was it?
>>101199683Very nice
first time trying to generate backgrounds, got what looked like a trashed cybertruck and had a small chuckle.
>>101209627runwayml gen3 alpha? but that's not locallocal video based on just diffusing images rather than a video model is more difficult and the local video models are more basic
>>101208297lumalabs
>>101210659>lumalabs>sign up with googlehere we go. isn't there a way to do this without doxing yourself to skynet?
>>101210687Yeah wait for the local versions dumbass
>>101210739when are they released and what's the name? where do people keep track of these thingswhy would they release that? isnt it propietary
>>101208965did you prompt for a cat or just rng>>101210843ty architect
>>101199808damn, reflection of that paint in the water is such a small detail, but it's SO FUCKING GOOD, I really like how well this lora of yours cooked up, it's delicious
so what are our thoughts on the new _pp samplers?aaaautism, cfg 7, 32 steps, normal scheduler,euler, euler_a, heunleft=base, right=_pp
>>101212040Very marginal gains, but this pp seems to favour more contrast. Another sample please? Upclose bust maybe, to see how it handles eyes and facial details.
>>101212053
>>101212082Interesting. Subtle but likeable changes in detail, especially for the more angular shapes in something like the eyes. Less coherent hand and arguably too much detail in ribs though. Again a bit more contrast. Interesting to see less visual artifacts due undercooking with Heun. I guess that could imply it denoises slightly more efficiently.
>>101212082How about a test for slightly increased amount of steps? I wonder if that could smooth it out a bit.
>>10121213950 steps
16 step
>>101212264>>101212208Oh wow, that's even more curious, especially the lower step one. It won't budge even at higher ones, though it caught up on the hand at least. That fry on Euler A though, wew. This seems to imply that it indeed has a much more agressive denoising process. Heun is relatively slow, if better supposedly better quality than Euler, and having it produce an image without visual artifacts at steps this low looks promising. With all that in mind, I'd conclude there's a sweet spot to be found somewhere below that 32 step threshold, and this pp might really go in favour of slower samplers.
>>101212645
aonon
>>101212002ty man it was one of the few good ones
>>101213943what else do you have cooked up?
>>101213954some work related stuff that can't be posted anywhere
Here's a fresh one...>>101214102>>101214102>>101214102
>>101213996Nice to hear. Had the opportunity to use txt2img in work myself.
>>101214108Appreciated! Though not a bit too soon? 310 is the post limit, and 150 is the image one.
>>101214114Yeah it's pretty fun. Doesn't feel like work at all>>101214131few more images then
ton bhic son
>>101200383>>101199879I can agree with this guyI do use Comfy but I have a research interest, I get stuff done now but it did require tinkering at the start, for you A1111 sounds best, for the drawing inspirationI've used SD.Next as well and I liked it, there was a UI fork too which last time I checked they were trying to merge together (next and the UI fork) but I've not kept up with their progress
