General dedicated to the discussion and development of local text-to-image models.Previous /ldg/ bread : >>101046579Incremental Edition >Beginner UIFooocus: https://github.com/lllyasviel/fooocusEasyDiffusion: https://easydiffusion.github.ioMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIStableSwamUI: https://github.com/Stability-AI/StableSwarmUIInvokeAI: https://github.com/invoke-ai/InvokeAI>Auto1111 forksSD.Next: https://github.com/vladmandic/automaticAnapnoe UX: https://github.com/anapnoe/stable-diffusion-webui-ux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiTComfy Nodes: https://github.com/city96/ComfyUI_ExtraModels>Use a VAE if your images look washed outhttps://rentry.org/sdvae>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>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-restsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>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
>mfw
Blessings upon you anonMay you enjoy this daily bread
>>101061949very cute anon
>>101061929that dang ol dog has five legs
>>101061846>how quickly I trainedHow long?
good morning/aftnun/evening/night anonsWherever youre from!
>>101061999cool texture
I've been out of the loop for a few weeks. SD3 is a stillbirth, /ldg/ won? Any good Sigma finetunes yet? What is that Lumina thing? HYDiT hao ma? Xiexie.
>ritual out of the loop post
hey guys, I've never posted before in my life. I was wondering why ldg is the best and sdg smells like farts
ldg is where all the cool kids hang.
looks like the chang gang has arrived
>>101061929Can you run any of these CPU-only without waiting for Chinese New Year?
>>101062761this is our lastest hopehttps://arxiv.org/abs/2406.11831>Benefiting from the inherent ability of the LLMs and our innovative designs, the prompt understanding performance of LI-DiT easily surpasses state-of-the-art open-source models as well as mainstream closed-source commercial models including Stable Diffusion 3, DALL-E 3, and Midjourney V6. The powerful LI-DiT-10B will be available after further optimization and security checks.
>>101062836Check out stable-diffusion.cpp, only supports SD1.5 and SDXL though
>>101062875>10BWith or without the text encoder? Either way, sounds promising.
>>101062906>With or without the text encoder?So that's basically a 10b imagegen + 7b LLM, I don't think a lot of people will be able to run this shit
>>101062886Very cool, thanks.
>>101062918We're getting into /lmg/ VRAM requirements, but fellow VRAMlets can generally cope by splitting a model between CPU and GPU at the cost of slower generation. I assume similar will be possible here. 7b LLM textgen 100% on CPU is actually not introlerably slow.
>>101062950I'm still not sure if a 10b imagegen model will fit on a 24gb vram card though
>>101062493On a 3090, 30 mins for 2k steps
>>101063126
>>101062988With LLMs, a model size is very roughly equivalent to how much VRAM you need to run it (and then you need more for longer context) , how does imagegen scale?
>>101063126Btw I have tested that it's 100% less uncucked than the base model. That is if I ask the base model to gen me a girl bending over a couch it wouldn't do it, but this does it perfectly (also keeping in mind I have like one or two photos of girl showing their butt and seems to have undone cucking). Also more likely to give topless without even asking it for it, and topless was also not in the training.
>>101062875Is that 10B just the model or including the text encoders?
>>101063413>>101062918
>>101063364>how does imagegen scale?SDXL is a 3.5b and it's asking for ~10gb of vram
>>101062875>this is our lastest hopeHonestly, Hunyuan has spoiled me. Prompt understanding is impressive, sure, but look at SD3. What's more impressive to me is coherency. I do not think we need 10x the parameter size to reach Dalle level prompt understanding, in fact Hunyuan as you can tell from their paper is close enough in both coherency and understanding. What it needs is just an update or very good tune to improve its understanding. Large param sizes will always be better for a model's overall knowledge though. Now, if that 10B model is truly coherent and powerful, chances are it's really on par with Dalle/MJ and it needs to be open sourced, but I don't we'll be able to run so easily. Though at 10B it's possible to run the model at Q4 with ease. I think the issue with the models we have and low quants is that they are too small, but just like with LLMs are a larger model should be better at q4.
>>101061502I would try including several closeup shots of human skin in the dataset. Perhaps Hunyuan can apply it as a texture>On a 3090, 30 mins for 2k stepsnot bad at all
>>101063461fug
>>101063522Interestingly, the effect of the LoRA seems to break or be less potent if I simply ask for a photo.The way I trained is >Candid style, a woman ...But when I prompt >Candid style, photo of a woman ...I'm likely to get fake skin againBut, based on previous experiments, I twist this to>Candid style, iPhone photo of a woman...This is meant to increase the quality, otherwise my LoRA could really use photo keywords to improve quality of results.And it works (though not in base model). This confirms to me that the word iPhone has an an effect on amateurish look in this model and is important for breaking the skin filter.In the base model, the only way I found of breaking this filter is by asking "selfie" or "iPhone selfie" in this portrait orientation.
>>101063461>>101062988>>101062918I think more people should look into 1.99 bits weight quantizationthese guys managed to reduce the model size of SD 1.5 from 1.7GB to just 219MBhttps://snap-research.github.io/BitsFusion/
Is it possible to create a diff LoRa from a finetune compared to its base model?
>>101061929Blessed thread of comradery.
Speaking of Hunyuanhttps://github.com/Tencent/HunyuanDiT/blob/main/controlnet/README.mdControlnet just officially released
>>101063571Good to know. That kind of systematic testing is very interesting. Have you tried prompting for different camera models? Like>a photo taken with Leica, full color film, focused,
>>101063573that looks interesting, on the LLM ecosystem, the bigger the model is, the more resiliant it is to quantizations, so basically, it would be even better for that 10b model (SD1.5 is a 0.75b model)
>>101063602>Is it possible to create a diff LoRa from a finetune compared to its base model?Yes. Some say it's better to train from a base, due to being supposedly more flexible in what it'll work with. I'm in the camp that it's better to train a lora on a model that is better familiar with aspects of the concept being trained.
>>101063736I've tried "Cinematic photograph", tends to work well, but it's bokeh'd and not the amateurish look I'm going for.
>>101062203>that dang ol dog has five legsit's a nuclear wasteland, the extra mutated legs is for immersion
>>101063834I rather meant creating a LoRa that encapsulates the diff(erences) between a finetune and it's base model, so that if you apply the LoRa to the base model, the output resembles the finetune.
>>101063934Oh, that's a bit more confusing to anwser. I'd say yes. For example, there's been plenty Pony/XL loras trained on output from SD 1.5 mixes, since people liked some of the styles it gave, and the loras indeed do a good job imitating their characteristic artstyles.
>>101063934I think you can do this by doing add difference merge and then turning it into a lora. It should be finetune minus base model. I did lots of this with 1.5 models, I have to check my notes
>>101062875Those comparison examples don't look better than DE3.I believe it when I see it. Currently the gap between local and DE3 is massive.Not only in understanding the prompt, but handling anatomy and complicated poses from many different viewing angles.
>>101064693China is a very cucked country (no porn allowed there) but I'm hyped because it's still less censored than SAI, that's insane when you think about it, Emad has the chance to live in a country that allows freedom and he still decides to shoot himself into the foot
dead fake worthless general
>>101065047fuck off bag to your discord bitchboy
the ride never ends
>>101065723it never ends
>>101065756>total eta 200 daysthat's something else
>>101065723>>101065756Hype!!
botted trash generalcringe
>>101065936CHINA general
>>101065904i have a bad feeling about this
>>101066025I use the original trainer w/ a fixed learning rate of 1e-7 (auto_lr = None) and CAME
>>101066150I'm converting old loras into v-pred. Using EasyTrainingScripts with prodigy + d_coef = 0.5. Slows learning, seems ok with bigger datasets
is there anyway to make 1.5 embeds work on XL and/or Pony?
>>101065936jealous discordtards from sdg
>>101066539>sdshit>>>/g/sdg/
>>101066570this place is for all models INCLUDING stable diff
Two or three weeks ago, /ldg/ was thriving and /sdg/ was seemingly just a handful of people. Now SD3 released, it's dogshit just like feared/predicted, and yet now /sdg/ is thriving again while /ldg/ is dead. What happened?
>>101066587sdg is five autists spamming 1girlsyou are one of those people
>>101066587/ldg/ was a fake general to begin with.
>>101066584Eh? No, this is the PixART/Hunyuan/Etc thread, all cucked SAI shit goes in /sdg/.
>>101066621>hehehee what if I troll this way?What, is the trani drama in sdg too much/
>>101066621youre wrong, its for all models
>>101066631>>101066633sai shills out.
>>101066631???>>101066633Newp, SD3, SDXL and all that SAI ass licking shit belongs in /sdg/
>>101066642You're obviously an avatarfag
>>101066641I dont care enough about SAI to shill, fuck them..I just want to know if there are ways to make my trained 1.5 embeds work on PD/XL
>>101066648Would a shill say he's a shill kek? Yeah I think you should fuck off to your sai shithole.
Good morning /ldg/
>>101066675ok sweety, build walls around this place to save yourself
>>101066675chang mad
>>101066714>>101066711>>101066648>>101066727Troll somewhere else
Hello fellow proooompters
/sdg/ sleeper agents trying to sow discourse between the 1girl posters and the chinks, do not let them get to you, /ldg/ was made to unite us no matter your sexual orientation
>>101066761>sow discourse
>>101066761>sow discourseOkay esl changKEK
>>101066779>>101066788fug...
Chang is one of the good ones, his heart is in the right place. It might come off as shill-ish but its just because he's passionate.
>>101066811you could say that about anyone whose passionate. what if youre a passionate israeli?
>>101066744>good morning>YOUR A TROLL
>>101066831I'm not engaging with poltardism
Just got a 7900XTX, how to use ROCm and SD on Windows 11?
>>101066886anon why not an RTX? AMD is a pain to gen with
>>101066831I'm a passionate zionist, yes.
>>101066916Because it was 200US$ cheaper than the cheapest 4080S, plus I might switch to Linux in the future and novideo is a pain in the ass on there.
>>101066986understood.for me the prices hardly matter, coz theres barely any difference btw them lmao
SAARS PLEASE
>>101066886install linux
>>101067181You aren't wrong. It would be helpful to know what anon uses to gen (a1111/comfy/sdnext/whatever) to kind of help guide him to the right area. Not sure how much of a pain in the ass it is to use rocm on windows but there's got to be a way.
>pixart team just added two members from nvidia >plans to release a new more powerful model we eating good bros
>>101067199???https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html
>>101067290sauce? Chang gang is coming up!
>>101067290Sounds good. Where did you see these news? I'm wondering why these nvidia people changed their oranization... if they did. Or are they still working for Nvidia?
>>101067318?Installing it is one thing getting it to work with the tools is another.
>>101067326it came in his sleep
>Our next version of the T2I model will try best to integrate the community's popular requirements, including text rendering, smaller model and faster inference, with best generation quality!we about to eat so good
>>101067290>>101067353>we about to eat so goodyes
>>101061949>frameare you animating, son?
>>101067353 >>101067383Sounds based in terms of goals. If it doesn't come with something like a non-FOSS license change: Godspeed to him and Pixart.
>>101067290its the opposite, two of them joined nvidia, looks to me like nvidia is gonna give them compute
>>101067473>looks to me like nvidia is gonna give them computeHUGE if true
>>101067467Copyleft makes it difficult unless they spin up from scratch. But if they did copyleft in the first place they're probably committed to the cause more than the money.
>>101066886
>>101067353>smaller model and faster inferenceI hope they don't try to please only the GPUlet crowd, my 48gb setup is begging for a high param model
>>101067532The thing about focusing on efficiency is efficiency scales.
>>101067550I hope their first model is lightweight and dethrones 1.5 for good
>>101067517They already managed to put together a model that trains on commodity hardware. Do they need to spin up much?
>>101067618That means they need to write the new training code from scratch which is work. It's not just the model, it's the code, that would also include work on stuff like the KV Compress.
>>101067550I am ok with them making a smaller model, I just hope they don't stop there and makes something larger like a 5b~8b model that is so good we may not need another model and arch in a long time
>>101067532It sounds like they will do a GPUlet model tho. But I'm sure you could train a non-GPUlet model based on the tech if it's open source again.Also Li-Dit / Hunyuan / Lumina and so on perhaps target much fatter models.
>>101067643I'd rather work be spent as much as possible on efficiency because that code could theoretically last us for years whereas if they start wasting time on large models there's more risk for them to quit or be regulated to death. I already think Pixart SIgma's architecture is enough to eat good for a couple years (we can scale models up ourselves). So the sooner we get to the true holy grail architecture even if it's just with a tiny model the better. I mean what if they figure out how to truly democratize training by sharding layers and making it possible to start linking up a bunch of cheap 3060s to train a 4B model?
>>101067633I'm not sure I get you. It already is copyleft. "new training code"?
>>101067684>Li-DitThat was confirmed not being open source, it's API only
>>101067716Oh. Damn. Well, then some future thing like it perhaps. Various entities will simply be tempted to make fat models on that enterprise nvidia hardware.
>>101067709Copyleft is a virus license, once you use it it's impossible to get rid of unless you start everything from scratch and clean room program your new codebase and model. That means if they do another Pixart model even for Nvidia it's going to be stuck on copyleft which means open weights and code.
>>101067751And how is this problematic? The only thing I see is it could hinder collabs from corpos which would want to make their own version
>>101067834> If it doesn't come with something like a non-FOSS license change:I'm addressing this worry, It's a good thing for us.
>>101067751But that's the point? It ensures the software is maintained under the same open conditions.It doesn't prevent use-in-work BTW. The output isn't copylefted because the tool is.
>>101067876I'm not saying it's a bad thing, I'm saying that we'll be eating good from the Pixart devs and not seeing a betrayal like SAI.
How much longer until pixart can do text goddamnit
>>101061929On
>>101067930further up the thread it did sound like the intent to train this was maybe announced 1-2h ago
What's some useful A1111 plugins that you use? Any recommendations?
>>101067930>>101068028Use the txtgen script
>>101068334me likey, aesthetics of good
>>101068334Wow cool
>>101068291dynamic promptincantationshires fix tweaks autocomplete can't remember others I used before migrating to comfy
>>101068291controlnet, adetailer, supermerger, neutral prompt, DynamicThresholding (CFG-Fix), multidiffusion, wildcards,
>>101068334>>101068548Trying to get that final shot of A.I. movie
>>101067353Where did you find that?
look guys i aint gonna lie to you allim just a retarded pervertand i want to make realistic fake nudes of women i know from workhow can i do this? what model or website allows me to do this shit?thanks in advance
>>101068953any UI that has a faceswap
>>101068953>and i want to make realistic fake nudes of women i know from workHope you get caught and fired
>>101068953I'm sorry Sergeant Johnson, I'm not gonna give you that information
>>101068953if you want to go the inpainting route, check out https://civitai.com/models/403361/juggernaut-xl-inpaintingits the best one
>>101068953this is why we can't have nice things
Anyone giving this faggot >>101068953 advice other than to fuck off also gets fired
>>101068953there is no turning back once you go down this path anonstay away from the darkness
>>101067716Nah they have plans of releasing it after they are done cucking it apparently.
>>101067383>lawrence-CWho is this guy (other than a new member of the pixart team)
>>101069186me
how much vram am I saving if I run the GPU headlessly? Note that I tend to have a lot of UI applications open like discord, programming IDEs and stuff
>>101063126>On a 3090, 30 mins for 2k stepsthat is quick
>>101069242linux
>>101069290Yes, I'm using Linux
>>101069366just open nvidia x server settings or the amd equivelant and see how much vram is in use when all ai related content is closed
>>101069193hello
>>101069143Source?
>>101069193Why work on VAE and not on the model?
>>101069825Because they're hoping someone coming along with a great dataset and resources to gen a better model.
was that b8?
>>101069825i hv autism
>>101069066>>101069082>>101069109>>101069110>>101069115look at this reddit tiered moralfaggotryfor shame
>>101069115not you, you got caught up in the net of dissapointment >>101069974
it's the usual freedumb pedotard
>>101069820From the paper>The powerful LI-DiT-10B will be available after further optimization and security checks.>The potential negative social impact is that images may contain misleading or false information. We will conduct extensive efforts in data processing to deal with the issue.Also keeping in mind they wouldn't release technical details at all if this were to be an API service.
>>101070014"will be available" can also mean "the API will be available"If they were going to release the weights, you would see references to the code being on Github like it usually happens
>>101070014yeah my hopes aren't that up anon, it's a fucking giant model, probably cost millions, dunno why the chinks would give it to us for free
>>101070014Even if they don't release it we have nothing to lose though. 10B is grossly inefficient for their results. From their paper it is clear that the model at best performs on par with SD3 on coherency, D3 still seems to be significantly better at it.
I have switched from 1.5 to SDXL and pony, but I can't seem to get the inpaint working properly. 1.5 used special inpainting checkpoints. If I just slam an SDXL checkpoint into SDXL, the inpainted section will be very obvious with colors off instead of being properly blended into the image. Someone told me to use the ControlNet inpaint feature, but I haven't tried that before and have no idea where to look for proper SDXL ControlNet models for it. Googling didn't get me far for now. I'd really appreciate any pointers or info on this topic.
>>101070334Take a look at this image for instance, I much prefer the D3 result for girl rowing boat and bird looking at phone. Tencent describes these kinds of inconsistencies in their Hunyuan paper as "AI artifacts" so they are aware of it as a problem, but I see no such indication in this paper. Wouldn't surprise me if Hunyuan simply is better than this 10B model.
So if everyone started training Hunyuan models, I think we might be saved from SD3.
>>101070424I disagree, the mood of the picture is actually determined by "terrifying". For importance the images should be defined by that.
>>101070492Yes but can you quickly look at the image and determine that it's AI? What about the D3 side? That's what AI artifacts are. The bad hands and oars are a telltale sign. With the bird, take a its legs and look at the phone, its relative size. Also whether or not the girl approaching the boat in the D3 pic has a terrifying atmosphere or not is subjective.
>>101070559No one cares about aesthetic quality if the AI doesn't generate what you're asking for. If you want a Booru random image generator you can just use SD 1.5. It's the wrong way to determine which model is best. DiT 10B is actually doing the best at actually following the prompt and with prompt engineering is probably capable of way more. The crab wearing a suit prompt, for example, is the one best following the spirit and intention of the prompt.
>>101070559Relevant pic. Though anyone who's tested the model has found this to mostly line up, and this wasn't even against 2B SD3 which would perform laughably in this matter.
>>101070623Aesthetic quality has nothing to do with coherency. Anon this model is 10B. Very few here can run it, and even fewer can attempt finetuning it to unfuck things such as hands. If this model gets those wrong out of the box, it's dead on arrival.
>>101070675Now you're just moving the goalposts because you can't stand losing an argument. If you want a random image generator use SD 1.5
>>101070694If I wanted a model that could also follow prompts I could just use Pixart or the slightly more efficient and better version of it coming soon. How am I moving goalposts? SD3 2B is also impressive at following prompts. Following prompts on its own is not enough.
>>101070743It's completely irrelevant to say "well you can't run it". Well what if I have two 3090s, or a A6000, or can rent a Runpod? IT IS STUPID. I do not care about poorfags not being able to run it. It's irrelevant. Your argument about it being not as good is objectively false, it's the only model actually following the spirit of the prompts, DE3 is really just a random image generator for half of these because it just picks and chooses which parts of the prompt it will follow. The point is, now, is you can't stand to be wrong so you'll change the direction of the debate to anything else. Which is okay, you doing this just means you concede.Guess what anon, you can still run SD 1.5 on your 2060. Whether or not a new local model running on your computer is irrelevant. The question for any of these models will always be is: are the weights open and available to be downloaded.And finally it's always retarded because fags like you bitch about the models being large while constantly talking about DE3. They will *not* ever be small again and compete against DE3. The fact is the future is 2B+ models. You have unreasonable, insane demands and expectations.
>>101070844>I do not care about poorfags not being able to run it. It's irrelevant.Well you should, unless you want to frequent an even emptier general.
>>101068953>picyou're gonna have her do illegal shit aren't you
>>101070925I don't give a shit, my enjoyment of AI models is not contingent on if 16 year olds can run it on their mom's laptop.
>>101070925You'll never breakthrough the elitist mentality Let him enjoy his walled garden
>>101070982>wahhh it isn't as good as DE3>wahhh it needs a 4090 to run
>>101070994case and point
>>101071015You're right, you're a child that wants to have his cake and eat it too. You have an insane demand that every new model be able to run on your 2060 while also complaining that it doesn't compete against DE3. Sorry anon, we're moving on from your Windows XP computer. Can't afford the magic box that generates anything you can imagine? GET A FUCKING JOB.
>>101070925>emptier general.Everyone has this open they are simply unable to contribute much outside of trolling since we deal with lower level concepts
>>101071046I'm NTA btw. Local will never compete with SaaS and that's simply a fact but neglecting consumer users is a fools errand.
>>101071077Wrong because SaaS tries to be everything for everyone which has exponential demands. There's definitely a sweet spot, for LLMs that 30B.
>>101070844You seem to be missing the point of my argument. Those are nitpicked images and prompts, anyone can nitpick and make their model appear great. For comparison, here's SD3 2B. As you can see, it follows the prompts quite well, and with certain prompts you can make SD3 2B look like the greatest model in existence. Doesn't tell us anything about the capabilities of the model though. A huge model following prompts better on its own is not impressive, what is impressive is both following prompts better and reducing AI artifacts. Prompt following is something that can be improved at a later point, such as by finetuning or attaching to a better text encoder, but image coherency is not something you can easily improve.
>>101071217Just another example of you moving goalposts.
>>101071217And my main point is that the model isn't as impressive as the authors are claiming, not that if it's impressive it's useless. But at 10B it should be doing much better than what we've seen. The bar is now much higher because lower param models have proven that they can in fact catch up to that prompt following ability.
>>101070963I've probably argued with you before on this matter, so I'll make my point clear AGAIN. Just as you don't care about accesibility, we don't care about a very arguably better model that's not available to the average anon. Come back when GPUs are beefier and cheaper while at it, or when models become more efficient. Just as the other anon mentioned, local will never be on the level of best SaaS, because free and open source is bound to work on a budget. It can however be very good within it's limits.I'd rather make collages of a thriving community that makes the best of what little it has, than waste my time in an elitist circlejerk for a hobby that barely anyone can engage with. I can understand that there will always be a minimum for entry due physical limitations, but that's why I am first and foremost concerned with accesibility, since democratization of art was precisely the original goal behind all this. The more people get into it, especially for a free and open source project, the quicker and better it will develop with the amount of community effort behind it.You cannot have a hermetic environment around open source projects, because it's bound to stagnate with lack of support from lacking interest.
>>101070955Well, you see, she does happen to be...
>>101071461hunyuan can run on 6gb vram. It just needs more finetuning to look cleaner. but does anatomy and prompt following way better than SD by default.
>>101071461Oh no the 16 year old won't be able to run the model, what a compelling reason not to do anything. I better not live in a house because you can't afford one.
>>101071461Nah, I recognize your arguments. You were probably arguing with me. That anon is a retard.
>>101071461>>101071566I still want bigger models and believe that TerDiT, quantization or flash attention will make them accessible on lower VRAM pcs. I just don't think that this 10B model is good enough to justify its size.
>>101071500That's irrelevant. We're discussing viability of resource intensive models within limits of an open source community, not hunyuan in particular. The point is completly flying over this >>101071535 anon's head.>>101071566Oh. Well.. I'm slightly confused now, but I'm glad to still see you around then.>>101071637There's probably a sweet parameter spot for local models, but time will tell where it is. Can't say if quantization is viable in case of txt2img, as compared to LLMs, but there's only one way to find out.
>>101071701Oh no the 16 year old won't be able to run the model, what a compelling reason not to do anything. I better not live in a house because you can't afford one.Anon, you don't contribute to open source, stop appealing as if you do.
>>101071727>contribute to open source>implying I should was my time doing so
>>101071737>maybe someone who isn't me will contribute to these modelsCompelling argument.
>>101071727lets see you train a 10B model
>>101070424>LI-DiT>Two prompts in the demo focuses on the little girl>Pedo model confirmed.
>>101071771I don't need to because the mere fact the model exists means someone cares to train it.
>open /ldg/>it's nothing but shitflinging and arguing>close thread
see you later, anon
>>101071787yeah have fun with dogs wearing sunglasses in space
>>101072018have fun on SDXL
Meme anon isnt here to post b8 img or else that whole thread of replies would likely have not happened :d
It's almost as if some anons didn't want to have interesting discussions on a mongolian basketweaving forum. I only take issue when it goes pointlessly in cricles, or when it's utterly deranged banter.>>101072063I'm here, don't you worry.
>>101072401Not much to discuss since there's just a terminal window showing training progress and another for Matt and Shane's Secret Podcast
>>101072401You may very well be fighting a loosing battle as many anon are more concerned with petty drama as opposed to image generation, but I'm a retard myself. >>101072507>terminal window showing training progress and another for Matt and Shane's Secret PodcastBased.
>>101072532It's kind of stagnant right now. Any of the new models in the end, if you want something decent outside of anime/art, you have to i2i.
>>101072565Slow and steady she goes.>>101072532Whatever the case might be, it's all about making the best of what little we have. It applies to collages, memes, machine learning, life in general, you name it.
>>101070424
HunyuanDiT confirmed to release a 0.7B "S" model soon, XL portion of announcement was mistake.
>>101073329Also integration with Kohya coming.
>>101073329 >>101073362neat. feels like a good amount of choice
>>101069186main guy behind pixart>>10106982516ch VAE would increase model input and output quality
>>101074103>16ch VAE would increase model input and output qualitydidn't notice an increase of image quality on sd3 compared to sdxl though, looks overhyped
>>101073341Is this what the trolls look like IRL
>>101074225Every isle at walmart
anyone have any good OL gens?
>>101061929What's the best software mix when your hardware is only mediocre at 16GB RAM and 8 VRAM?
Last from me, gn yall
>>101074894No clue even if I knew what you mean by OL.>>101074916nite
>>101074894
>>101075054neat
>>101075061thanks>>101074905I'd recommend Metastable, StableSwarm, or https://github.com/lllyasviel/stable-diffusion-webui-forgeI'm working with a 16ram 8vram laptop myself, and Forge gave me the boost up to Pony/SDXL models, so these should work nicely for you.
>>101075158>I'm working with a 16ram 8vram laptop myself, and Forge gave me the boost up to Pony/SDXL models,How long do gens take for you?
>>101075513Relatively fast, depends on the resolution. I stick around 1024 and usually don't hirezfix more than 1.25x, but I can go up to ~1.5x at reasonable speed I guess. ~45 seconds on initial load of a model. ~23 seconds on the average 1024 gen. ~17 seconds for a 1024 inpaint.
>>101075327Tuff
It's anime night btw
>>101076100oooo
i saw some pyramids posted here have another
>>101076100>Anime tits!
>>101076169
ok my turn to spam the rest of the thread. my goal: the remaining 62 gens are mine and mine only
sd3 only below this line---------------------------------------------------------
all old models are deprecated please delete them
>>101076331
asslicker
C-C-C-C-Combo Breaker!
>>101076410ouch
Once pixart does text out of the box without some script it's game over
>>101076468never ever, never ever ever never
>>101076468*and they finish the 16 channel VAE
>>101076468
>>101076477Confirmed not over. Keep posting
>>101076465you could not have posted a worse pixart gen kek
>>101076518
Pixart bros its over maybe
keep in mind the difference between ALL of these prompts is one word or one phrase. Text outputs = one sampler, image outputs = second sampler
Where is 4B Liecon
>>101076520Don't like oil paintings? Have another bad gen
>>101076673
>>101076697Catbox for any of these? Is it the new scheduler/sampler combo?
>>101076746these are a mix of euler sgm_uniform and dpmpp2m sgm_uniform.you can have the whole mega from the first set or a catbox if you don't trust me
>>101076765lol testing now. ty anon who has received little faith from others
>>101076765Not bad. Now how many steps. Spill the beans anon
>>101076746>>101076826>>101076862https://mega.nz/file/lXhBDRzB#FIOO-UAUuqoPBHd6gdVIBKJuZ9EOFpm8obYt1hHSz90the catbox uses some different stuff like advanced perturbed attention guidancehttps://files.catbox.moe/24k0bq.pngyou will need to use this wildcardhttps://files.catbox.moe/p8n8gj.txtremember me in valhalla
word on the street is that there is comprehension of other languages as well. ALSO, I am using only T5 for some of these, with blank text fed to CLIP_G and CLIP_L
>>101066758quite unique if generated, none of that SD slopface
>>101076906shut up chang
SD3 is the people's model
>>101076899>word on the street is that there is comprehension of other languagesHave you tried it yourself?
>>101076746The new scheduler and sampler combo is apparently SOTA according to their papers and their Github which links the papers that make these claims. Their code is in ComfyUI as of yesterday. It claims at lower steps and on convergence, you are getting comparable quality as other samplers and schedulers at higher step counts. The leader on this was DDIM but now it seems like this has displaced this. Align Your Steps was apparently almost state of the art but got displaced with the work in the paper.https://github.com/zju-pi/diff-samplerhttps://arxiv.org/pdf/2405.11326
>>101076941nah chatgpt resisted fully translating my list of naughty words so I gave up for now, just letting the epic wildcard ride
>>101076944Sorry, I realized I misinterpreted and was in the wrong thread. Regardless, I think this still should apply to other image diffusion models other than Stable Diffusion, it just might take other values and some tweaking to get to the same result.
a local model is a local model. godspeed anon.
>>101076984
>>101076971this one is actually it's own prompt and I kind of hate it, made like 30-40 and they all are too saturated. full body characters are going to take a lot of development in understanding how to use this thing
>>101076591What was the prompt for this gen? I like the scanlines,glitchy analog tv look
>>101076875>you will need to use this wildcard>https://files.catbox.moe/p8n8gj.txt>zoophiliaUhhh th..thanks anon>remember me in valhallaFor valhalla!Will look into PAG more though ty!
>>101077173that was when I was tuning the perturbed attention guidance settings and other thingshttps://files.catbox.moe/e2czji.png "(fuckme:1.5)" was the only difference>>101077206lol it's from a github of naughty words. The whole point was seeing how sd3 behaves with bad wordsthis one is coprophilia for example. The model clearly has learned some really extreme concepts. Despite all the safety training
>>101077305(coprophilia:1.5)
>>101077305>that was when I was tuning the perturbed attention guidance settings and other thingsah alright, thanks for the catbox
>>101077333
did a few with the pcm lora, the results were very interestinghttps://huggingface.co/wangfuyun/PCM_Weights/tree/main/sd3
>rundiffusion working on training sigma Someone more knowledgeable than me tell me how I should feel about this
gonna end the thread with some of a specific set I did after this one
>>101077560>fug it happened again
fin
>>101077611See a psychiatrist.
>>101077526>>101077543>>101077560>>101077575>>101077611bruh
tfw no new bake
>>101077660we (the discord) decided its time to end the ldg experiment. since we were the ones that kept it going this whole time, we're the ones to decide its over
Bake?
baking
>you must wit 300 seconds to post
>>101077823I got you senpai >>101077830>>101077830>>101077830
