Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101933401>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>GPU performancehttps://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>Maintain thread qualityhttps://rentry.org/debo>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>>101935320anyone know any good post processing nodes for comfy? this was my bootleg method
>mfwyou guys never fill your threads
>>101935327>Welcome to the buttchin bus, stud.
What models do you anons use to caption dataset for Flux? I normally used WDtagger Swinv2 on A1111 for dataset tagging. Is it really necessary to boomer prompt your captions for flux?
asking againwhat’s ldg’s current meta for guidance/cfg with gguf?
>>101935334Is Debo the first trans avatar fag that made the leap from /sdg/ to /ldg/?
>>101935346Seems to eat all my vram. I tend to avoid cfg all together unless I need something very specific.
are flux loras for 24GB only, cause it's slow as shit on a 4080.
>>101935344>I normally used WDtagger Swinv2 on A1111 for dataset tagging. Is it really necessary to boomer prompt your captions for flux?natural language will always be more accurate than just tags, what if you have "woman, table, chair, sitting" , does that mean the woman is sitting on the table or the chair? only a natural english sentense has the answer
>>101935359*generating, not training.
>>101935357thats what my default is, just nothing fancy. its faster but I wish I had more levers of control
>>101935344Some anons are locally running this and have edited the python script to batch. Might have to ask when the thread is more active or take a look at it yourself. https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha
>>101935338kekim throwing old juggernaut gens into joycaption and throwing that into flux with minor adjustments to see the differences in style. really impressed with the prompt adherence so far>>101935344>Is it really necessary to boomer prompt your captions for flux?if you want nuance then yes, i'd say. more information is always good. i also have no idea what im talking about
picrel is the juggernautXL v8 reference image. we're 100% gonna go back to old models for the early AI aesthetic and "sovl" one day. maybe flux will be considered aesthetic and sovl too in 5 years
So can we make epic porn or what. Someone post a pic of trump bumming Kamala please
>>101935359I clocked 16.3 gb + 9.7 = 26 gb total vram on 2x4090 on 1 LoRa. Generation only. Maybe you can manage 24GB or lower with nf4 or something.
>>101935452we're not that kind of general! chud!!
>>101935452>Someone post a pic of trump bumming Kamala pleaseon a blue board?
>>101935465ok, now it makes more sense. I think the play is to use sdxl/pony for loras/characters then do inpainting in flux for stuff.
>>101935480*for characters that flux cant do natively, that is
>>101935480Mentioned this in an earlier thread but you can also do the opposite. Use Flux to create an advanced composition and then resample it in a Pony-based model for aesthetics. You can also use a mask to keep specific items such as text from being destroyed, and resample those items at a lower denoise later.
>>101935465is that normal for flux loras to take that much vram?
>plastic Flux titsfuck this was with the realism LoRA as well
>>101935519I cannot find an official word on this anywhere, but I assume it's bugged. There's no reason I can think of why a LoRA that is only a hundred or so megabytes is eating up all the vram.
girl in the back is built like a nutcracker
>>101935532this shits cooked
>jeets already spamming civit with dogshit flux lorasIncredible
>>101935532the realism lora is snake oili also havent seen any good tunes that keep the realism at base flux levels. everything becomes plastic slop
>>101935542>a round chin in fluxyou cracked the code
>>101935542very nice
>>101935537>I cannot find an official word on this anywhere, but I assume it's bugged.there's an issue about it, and yes it's a bug, before I had no problem loading a lora now it's overflowing my 3090 even though I had 10gb of vram to sparehttps://github.com/comfyanonymous/ComfyUI/issues/4343
>>101935553its cuz civit lets them train loras with their jizz credits. theres no quality control, its all to drive mindless consumption
>>101935563>the realism lora is snake oilmost tend to be
>>101935565>you cracked the codethe secret is unironically youthful adjectives>This is a high-resolution photograph featuring two young Russian girls seated in the back of a modern, luxury car. The image is bathed in a vivid, neon pink and purple light, creating a surreal and futuristic atmosphere. The young girls, both of whom have fair skin and straight, long auburn hair, are dressed in matching, high-gloss white latex outfits that resemble futuristic or sci-fi attire. The outfits have high collars and exaggerated, angular shoulders, adding to the futuristic aesthetic. The teen girl on the left has a neutral expression, while the teen girl on the right has a more intense, contemplative look. The car's interior is sleek and modern, with black leather seats and a minimalist design. The ceiling is lined with neon lights that emit the vibrant pink and purple hues, contributing to the overall otherworldly ambiance. The windows are tinted, and outside, the night sky is barely visible through the tint, adding to the mysterious, isolated feel of the scene. The photograph captures a moment of introspection and modernity, blending fashion and technology in a visually striking manner.
>>101935586Civit adding on site lora training was a mistake, 90%+ of all loras being posted now across all models are objectively dogshit. Zero quality control on that site, just an overwhelming flood of slop.
>>101935545yeah it's overcooked a little bit. still honing in the numbers on my cfg/adaptivethreshold
>>101935615do you use tonemap? if yes you could decrease the multiplier value to uncook some of it aswell
>>101935563this one feels very tron
>>101935509yeah, ponyXL does characters really well but sometimes the backgrounds are not detailed, so these models can work in tandem to create good results because there are TONS of 1.5/xl/pony loras.
>>101935614they're trying real hard to figure out how to drive revenue. they never cared about quality of content but they especially don't care now.
>>101935614Is there any filter or tag that indicates it was trained on site? Really want to filter all that crap.
>>101935644Just speedrunning enshittification >>101935648Not that I'm aware of
>>101935614sort by downloads, the best stuff generally gets lots of downloads. and you only need one good character lora to make a billion gens.
>>101935667flux makes a good inpainting tool too, even without controlnets it manages things well: i'd usually use openpose or canny/depth when making clothes edits.
is 20 generally the sweet spot for steps in flux? (dev model)
>>101935553>need my celebrity slop! saaar!! redeem the emma watson lora!!Hopefully when Kamala makes deepfakes illegal they scrub all celeb slop off of civitai
>>101935709It seems to be 30 actually for the most consistent results.
>>10193570920-25Anyone used this?--bf16-vae --fp16-text-enc --fp8_e4m3fn-unet
>>101935709I use 30+ these days personally. The lower, the less things the model will be able to get right from your prompt. In other words it will follow your prompt less. 20 feels a bit too low for me sometimes so I bumped my default.
>>101935732"realism" in image models was always a mistake. I'm also on my knees praying mommy Kamala and the EU fuck all of the jeets in the ass.
>>101935309Which one is the best to run locally? I was recommended flux but I don't see it in the list.. is it comfyUI?
pixel art is a fun prompt to start things with:
JoyCaption is really good at describing the angle of a shot, not so much esoteric outfits or cyborg stuff>>101935631thanks, im going for a gritty neon cyberpunk aesthetic. ive noticed flux tends to add too much light to images
>>101935732you can just use something like reactor, in flux, to face swap a txt2img gen into her if you wanted to. no need for a lora.ie: john wick on a beach, but ryan gosling (swap)im having way more fun making mikus though
>>101935806*in forge, flux is a model. or with comfy. loras are good for characters but flux generates perfectly fine human females, all you need to do is a faceswap.
>>101935741that looks like command line for setting the checkpoint weight and text encoder weights, you can set that in comfy on the nodes
thread quality always takes a dive during the weekends
Would you guys rather:>load CLIP to CPU, wait 20-25 secs for every prompt changeor>use a quant to reduce the model size on your VRAM, but potentially get worse gens
>>101935850>wait 20-25 secs for every prompt changehell no
>>101935334is this what japanese looks like to normies?Is that why you keep posting NTR as perfect examples of cute and innocent girls?why cant normies and AI just learn jap?
this one turned out cute>part of prompt is holding a mic>ends up being one of the old school radio mics
>>101935553Here's your Flux lora bro
>>101935850i would not go below q4
>>101935860Why did he use Fifa '18 for the dataset, is he stupid?
>>101935751its nice to see the aesthetic is nice enough to be worth replicating>>101935850quants only matter if you care about reproducibility or if you're doing text stuff imo
>>101935852This happens regardless right now, because of how LoRAs are busted.It's actually quicker to just load the clip to cpu than to wait for it to load up again on vram, oom out then load again.
>>101935863Don't worry I'm sitting on 24GB of VRAM I just start hitting the limit when I upscale and I'm sick of my machine freezing up.Guess I could ditch windows and plug my monitors into on-board graphics
>>101935842It was in the comments here to speed up gen timehttps://www.youtube.com/watch?v=chfUGCE0AVY
>>101935452There is an anal LoRAhttps://civitai.com/models/649472/anal-art-for-flux?modelVersionId=726623
>>101935309>Beginner UIi stopped visiting this place for a few months and now A1111 is not for beginners what happened?
>>101935885Interesting, I haven't tried using loras yet - Just exploring flux after taking a break since SD3 dropped. I remember having this being a problem when I was playing with sigma (before comfy was updated which stopped having to reload the checkpoint after anything in the workflow changed).
>>101935896well for me what made things way faster was fp8 weight for the main model, left the text encoder at fp16. but if you have 24GB vram or 64gb ram then fp16 is fine. it will slow down if you max either.
>>101935854are you japanese?
>>101935334>>101935854Manga translator here. some of the phrases almost have meaning. it's like how SD used to be with english>うし麦し半<>>ちわいってーんい!>土@んてっん上こまそいushimugihan<>>cow wheat/master half<>chiwaitteeeeeeni!>CHWAI NEH I SAY!tsuchi @ nteeeenjoukomasoi>earth @ net roaf thing n deet (TL, could be a muffled 'roof thing thin & detailed')>乃のーも毛のゐってに>てース班いえ4っ母らnonooo moke noyi tte ni>attempt to say no but spelling it two ways then mo also spelt 2 ways followed by soundsteeesu sake ie 4 hahara
Goodnight
So I swapped out .sft to .safetensors on ComfyUI and that seems to have helped performance after I rebooted my PC, not sure if placebo or just raw benefits of a new insane Adaptive Guider workflow.
>>101935869I think it's literally a lora of his custom fifa character judging by the description. Why he thinks anyone else would want this is anyone's guess
>>101935920.sft is just .safetensors, the file hash is the same
>24gb vram>Training at 512 resolution doesn't even really impact outputs>can train at rank 32 LoRAs now24gb chads, we're eating good.
wew, there's a few controlnet models out alreadywhat's the feedback?
>>101935904>i stopped visiting this place for a few monthsthey were added just after christmas so you were gone longer anon
>>101935969Meh
>>101935904Technical debt and bloat. The gens are waayy slower on A1111 for some reason and people eventually move on to other front ends.
>>101935981I should add to this. It's not that they don't work. It's more that the controlnets we're getting feel more like poorly trained publicity stunts than anything practical. Sure, they confirmed that they worked, but the dataset and time it took to train them reflects the poor quality. See xinxir for an example of how a good controlnet should operate.
>>101935987and A1111 doesn't support Flux right? Damn that guy is slow, I mean, I have a lot of respect for him when I learned SD in 2022, but the reality is here, forge and Comfy are better devs than him
>>101935994can they be used with 12Gb?
>>101936011I don't know, sorry.I know a good way to find out!
>>101936014yeah I'm on it
>>101935940>Training at 512 resolution doesn't even really impact outputsi have only heard people say this is a bad idea to do
forge defaults to 896x1152 even when ui-config.json has height/width set to 1024x1024?why
>>101935320>photo of an Indian being forced to teach his replacement as a condition for receiving unemployment
GGUF in ComfyUI now works with the native LoRA loader. No extra fuckery required.https://github.com/city96/ComfyUI-GGUF/pull/29
>>101936029nm other people are having a similar issue, it's a bug
>>101936045>it's a bugit's a pain in the ass is what it is, but you can easily load a png into the png info tab and send the parameters to txt2img
>>101936044thank you for your service fren
>>101936023I've unironically heard the exact opposite.
>>101936044Based City.
>>101936044KINGHow did you do it?
>>101936044this may be the way to use loras without 24gb, neat
>>101936044wait what is this, you can load conditioning?
>>101936044dumb question, but how do you get the little denoising preview on the k-sampler?
>>101936061the place i heard what i heard is here, so you probably shouldn't listen to me at all
>>101936044loadcond replaces the clip text encode nodes? how are you prompting? im a comfy noob
>>101936120Ill be back in a few hours and see how my megumin LoRA looks. (It's being trained at 512).I've done 1024, 768 etc and they all turned out great. I wanted to see how those LoRAs looked at a higher rank, so I turn down the resolution.
>>101936044Speed is a bit shit since the calculations are done in FP32. Comfy added an option to do them in any dtype but that just got merged and I'm not gonna deal with 20 more reports of "it don't work :(((" by instantly using it kek>>101936072Instead of applying the LoRA weights to the quantized tensors I keep them on the CPU, start loading them asynchronously, do the dequant (by which time they're already on the GPU), do the math on the actual weights, then delete the copy on the GPU>>101936095>>101936121Ignore that, that's just for testing so I don't have to load T5 each time lol. I just saved the outputs of the text encoder node to a .cond file.
>>101936098Launch with "--preview-method auto" or "--preview-method taesd" (if you have downloaded taesd)
Good morning sirs. Does anyone else get this error trying to apply the lora to gguf?
>>101936044THATS MY GOAT
>>101936166It is telling you that your failed the math exam, unfortunately you will have to repeat the grade.
>>101936130>I wanted to see how those LoRAs looked at a higher rankah yes, the HoRA
>>101936166Which LoRA? I only tested on the ones I trained on ostris's trainer.
>>101936166did you update?
>>101936183SDXL LoRAs were often 64 to 128. 32 isn't THAT high rank.
>>101936166connect it at the start of the model loader rather than after anything else than modifies it, that's how i fixed that error
>2 weeksLook at where we are now. This shit is moving faster than any other local model since 1.5.I want the niggers who spent 2 days arguing that it was impossible to train to speak up and accept defeat.
>why is my speed so dogshit>cfg 1.1
>>101936157This is also just an option in comfy manager.
>>101936188It's the official flux realism lora converted for comfyui>>101936236it don't work :(((
>>101936189Nice. How did you get the swastika?For now the best gens I got with schnell were with the [*background] [(*face)style] prompt style
>>101936257>loras not possible>finetunes not possibleapparently a lot can change in a week
>>101936202Yeah I updated the gguf custom node, didn't update comfyui, kinda afraid to after seeing the UI change
>ingame screenshot of world of warcraft starring miku hatsune as a wizard
>>101936323>kinda afraid to after seeing the UI changeThere is an easy fix for this. Get your autism medicated
>>101936294From here?https://huggingface.co/XLabs-AI/flux-lora-collection/tree/main
>>101936325
>>101935542>>101935563>>101935595These are sickEverything else in this thread was gay and a waste of time
>1000 steps into my 512x512 megumin testIt's starting to look like her.
>>101936329>>101936308Why is it generating the same leather jacket for every gen? I noticed it does the same for some background, like it knows only one type of leather jacket or city
>>101935470Daaaw is the ameriturd scared of a banny wanny?Pathetic
>>101936308Just ask for it desu, flux knows swastikas unlike past SD slop that would always just give me messed up checker patterns instead>>101936329Yeah, the one that was converted for comfyui
>>101936344First sample for reference
>>101936294Odd, I have the anime one and it seems to load.>>101936323I also hate it. Just make a copy of the "web" folder and update it, delete the new one and copy the old one back.Then before each update you'll delete the web one again, do a "git reset --hard" then a git pull before copying the old one back kek.
>>101936354Thanks for the hack to keep the UI, I might resort to that as well o7
>>101936354Comfy really needs to add an autism option in the settings that changes it back to old. Does not sure what motivated him to do a UI change when the primary people comfy appeals to are actual autists.
>>101936189i wonder if flux devs being german will get any controversy for being able to generate swastikas so well>>101936338thanks anon ill definitely be using the first image in my future cool music video where i rap about fucking robots>>101936257>This shit is moving faster than any other local model since 1.5.this is also because people have more compute on average, and techniques like lora/controlnet were already discovered/invented
>>101936351This works for mehttps://huggingface.co/XLabs-AI/flux-lora-collection/blob/main/realism_lora.safetensors
>>101936380We had all those tools with XL too, but it took ages for it to be worthwhile. Then we've had pixart and hunyuan and nobody gave a shit about those either.the primary motivator is how good the base model is.
https://github.com/bmaltais/kohya_ss/issues/2701Lurked this thread for flux kohya ss training configs. Running one from there right now with some old danbooru style dataset I have, at least it worked and started training.!Watch your temp/ power limit!
>>101936386>https://huggingface.co/XLabs-AI/flux-lora-collection/blob/main/realism_lora.safetensorsIn comfyui? The one that wasn't converted? That one gives me different errors and ends up not being applied at all
>>101936397Use this one>https://github.com/bmaltais/kohya_ss/issues/2701#issuecomment-2294611159It's what I'm training on right now, it's the only one I've had success with samples are promising at 3600 steps
>>101936402>That one gives me different errors and ends up not being applied at allThat was fixed last week I think https://github.com/comfyanonymous/ComfyUI/pull/4302
world of miku:
>>101936402Yes
>>101936428Thinking about making WoW classic LoRA to really force that aesthetic. Interested?
>>101936420I got invalid Additional parameters with that one.
>>101936351
>>101936444Can you try it with leather jacket?
>>101936440Could be --highvram? Did you git pull? Every other config I tried from that thread trained to completion but the finished loras had little to no effect at all
>>101936393>the primary motivator is how good the base model is.you're right. SDXL would be half of what it was today without pony
>>101936379The default UI is still the old style one. Right now you have to go in the settings and manually set an option to get the new one.People have been complaining about the UI since the beginning, the new UI is the beginning of one that's designed by actual UX people.
>>101936439im just testing a range of prompts, save your training time for a character/aesthetic you like. but ty
>>101936490I-its not like I'd be making it for YOU. I just like the aesthetic is all, BAKA.
so if Q4 works with loras that means people with under 24GB can generate loras, right? Even Q8 would work. I tried the regular fp8 model on a 4080 and it took 3 minutes (16GB vram).so this is very promising stuff
>>101936460
>>101936498it'd be a neat style in general, would be popular on civitai
>>101936474That's weird because I didn't change anything, I just did a pull and it changed. I definitly didn't pick an option to get it. I assume you made a change between then and now.I personally like the new UI but please understand that your audience is mostly people with varying levels of actual autism and change makes them angry.
>>101936474the specific design isn't what people have been complaining about and your userbase is predominantly autistics who hate change
>>101936499*generate images with lorastraining is another thing entirely
When will the Q2 quants drop?
>>101936505Did you already have the beta UI enabled? (the bar instead of the floating side menu).
now we're getting closer, now that I specified long twintails.
>>101936511dont do that to yourself anon. buy a used 3090 or even a 3060
>>101936509>>101936509?? I've been using loras on q8 just fine on a 4070ti Super 16gb. I even trained one locally
>>101936518>the bar instead of the floating side menuYeah I did. I guess that explains it.
>>101936534probably cause I used the default fp8, I was running out of vram, ill try the Q8/Q4
>>101936523Looks pretty cool Anon. I wonder if specifying her username, the map name in the minimap and the chat contents would replace the gibberish texts.
>>101936534how do you connect the gguf or lora loader nodes to a clip/prompt node? it just has model out, any screenshot of your workflow?
>>101936380Can we get nudes of the Russian girls? How/where?
Flux sure likes necklaces
>>101936501based
>>101936534>I've been using loras on q8 just fineSo we can just use any Flux lora? I wonder if controlnets will be the same.
So is nf4 old and busted with the GGUF being the new hotness?
>>101936590Is that what you think Russian women look like?
>>101936625yeah
>>101936590definitely not on a blue board, also base flux sucks with nudity and nipples
>>101936590russian-nude-girls.com.exe
>>101936474is the old one going to get removed
>>101935309I like how even if the base model can do it, a LoRA can help keep the prompts shorter. Not fully accustomed to boomer prompts so that's another great use case for LoRAs besides adding style.
>>101935614>just an overwhelming flood of slopJust like real life.
>>101936633he really choose the most East Asian looking gen in the whole thread to ask that question too kek>>101936695>Just like real life."90% of everything is shit"
>>101935850>wait 20-25 secs for every prompt changeEvery gen is several minutes, so whatever.I do both and I'm still in pain.
>>101935850I have a 7700 and even a prompt with 3k characters is processed in under 10 seconds
>>101936674Maybe in a long time idk.
>>101936723what prompt is that anon
>>101936702To be fair, women from Vladivostok tend to have that look, sort of, due to the proximity to Mongolia, China, and North Korea. Russia is more ethnically diverse than people realise. Flux seems to agree.>Prompt: A woman from Vladivostok Russia wearing traditional clothes.
>>101936729a made up prompt to test the processing speed
>>101936726I think you should leave it in forever as an easter egg if someone uses the arg --autism
>>101936741I doubt the model has any specific information on what part of Russia they models come from.
>>101936741Wait this is the wrong pic, it was this one
>>101936702I was referring to the ones in white, not that stinking gook looking girl
https://civitai.com/models/644109/pepe-flux?modelVersionId=720509a new age of pepe
>>101936756Yeah the vlm wouldn't know, but I am curious now, gonna run some tests.
>>101936741>that watermarked chin
>Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.what it be
loras do in fact work, neat
>>101936804it checks if okay on GPU so don't have to go to CPU
>>101936808aight, what do
>>101936791It's all just vaguely Russian looking regardless of the region. None of them are correct.
>>101936817it be faster
>>101936821They are all sisters from the same chin.
>>101936829They are all gens from the same seed.
>>101936826What I'm 'posed to do?
>>101936839don't compile torch yourself
>>101936726yeah i can understand if it starts to cause trouble
>>101936858
>>101936858Training LoRAs, cant gen right now.
Anyone get joycaption running locally yet?
>>101936878yes
>>101936889Cool, how?
>>101936804that's usually proceeded by indexing errors or something like that. it's an issue with the code, nothing you can do yourself, report the issue with the full traceback
has science gone too far? this is what happened with "elf" by itself.
>>101936891I cloned and ran it.
>>101936806>loras do in fact work, neatit's real
>>101936846>>101936900I just updated my nvidia driver and now I get all kinds of shit, I don't what these mean
>>101936858I found a well priced Titan. Is it a good choice for those 24GB?
>>101936937I did that with 16gb, you dont need 24 but it's nice for training purposes.try >>101936912 or try fp8/q8 in forge with the update
>Goofgufs now working with LoRAsThis is so good.
>>101936474>UX "people"It's fucking owari
>>101936973That's almost a cameltoe. Is it learning?
>>101936991>Is it learningNo, unless he has a camel toe lora that is purely coincidental. It can't learn, it's a pre-trained model. We have to feed it additional training in the form of loras and fine tunes.
>>101936998B-but AI!
<lora:FLUX-Pepe-1:1> Pepe riding a skateboard in summernew age of memes begins now
>>101936991pure luck just by prompting for tight high-wasit shorts
so how is gguf lora formed?
>gguf+lora works but not with dynamic thresholdingowari da
>>101936912What all this stupid celeb loras show, that is posible delete the chin butt with a a proper data finetune
>>101937023how is goof form?
>>101936044any plans for Q8 T5?
>>101936998You don't say? I thought it could communicate through the internet by interfacing with anon's GPU, altering its weights and syncing them via bittorrent protocol in real time.Thanks for clearing that up. It's nice to have intelligent people like yourself here.
>>101937029It also shows that Trump and Kamala had an affair
>>101937023looks like no conversion required
>>101937040>he doesn't know
Flux werks but i dunno what to gen
>>101937040You joke, but it wouldn't be the most retarded question asked seriously in here, so I just assume everyone is an idiot.
>>101937026I moved away from dynamic thresh because it was tanking my performance and the visual gains were subjective at best. Unless you very specifically want a certain thing I'm usually fine without it.
>>101936878>git pull it from hugging face >copy+paste venv from comfyui or whatever into the folder>edit app.py and change the model path to a publicly accessible model, like the quantized llama one (check archives for link) >If on windows, create batch file to run app.py (ask chatgpt if you don't know how) >open UI in browser >done if you encounter any issues or errors ask chatgpt how to fix them
I'm gonna post it
>>101937051sex gifs
prompt cartoon to get cartoonish pepe instead of RL pepe:
>>101937055dont
>>101937040you'll find anons on the DALL-E 3 threads firmly believing that the more they prompt a certain celebrity the better the model gets at reproducing themstupidity is real
>>101937055DO IT FAGGOT
>>101937066DALL-E jeets aren't human
>>101937060I don't think the flexibility and successfulness of flux LoRAs is being discussed enough. Even at low rank they nail aesthetics. Go ahead and find an artist you like, take 50-100 of their images and just shove them into a trainer, don't even bother captioning it and it will ape their style almost perfectly.
>>101937055you won't
>>101937055Don't do it anon...
>>101937066Ask them if they can gen this.
>>101936786(((The creator of this asset requires you to be logged in to download it)))How does no one seemingly have a problem with this?
>>101937082yeah, even for testing a SDXL lora I used 10 images from one piece and was able to make a nami that looked just like it. and that's with BASIC settings, not even tons of training.
>>101937066just remember the average IQ in India is literally the IQ standard for qualifying as mentally retarded in the first world, and everything makes horrifying sense
So, I found a way to avoid the chins might be to prompt ethnicity and avoid writing "woman" or "girl". You can already see that with Asian. It's not 100% foolproof, but it might be worth looking into.This is a "Jewish beauty".
>>101937088There's literally paid loras on civitai now
>if only you knew how GOOD things are
>>101937086only the censorship around DALL-E would stop this from generatingif you played with it in the first week there should be no doubt it can gen that plus tongues and saliva (Flux can't into tongues and saliva)
>>101937096Mental retardation determination by IQ only makes sense after normalized to 100 in the test population.
>>101937101You could be exploiting people's need to cum and make your own LoRAs.>>101937104I like how trump's eyes are also vaguely pepeish.
>>101937116You know what DALL-E can't do that FLUX can? Run on my GPU.
>Been like 4 days since the 24gb fine tuning guy said he could get it working on a single 3090.>Even Kohya was impressed>Nothing has come of it yet.Where the fuck is he?
>>101937101Signs the site has hit rock bottom, we are in dire need of a new website
>>101937133DALL-E 3 is 8B+5B, all you have to do is get it out of OpenAI's hands
>>101937146Imagine breaking into open AI and the only thing that was stolen were the Dall-e weights.
cartoon, Pepe is holding two pistols like John Wick and has an sad look on his face, he is wearing a business suit and diving in the air while firing. feels bad man.
>>101937118I see you qualify as average in India, good job champ
>>101937101>buy lora early access>it's shit
That's a long barrel
>>101937188for you
>>101937101I don't really see why anons are so bothered by this. Just ignore it until it's free, who gives a shit?
>>101937207Annoyance to paywalls is usually directly proportional to how horny someone is in that moment.
>>101937207the precedent
>>101936044Damn, even with this I OOM, the fuck did Comfy do, a few days ago I could run loras just fine
>>101937217Coomers were a mistake
>>101937221I'm gonna be honest dude, I don't get the outrage. If someone wants to put in the work of gathering the data and the compute to train it then wants to charge for it, that's their right.
reminder, open source always wins
>>101937235>put in the workyou say this like LoRA training is hard. It's not hard. It's just expensive, in terms of resources required.
>>101937191kek
>>101937235that was already perfectly summed up in "who gives a shit"
>>101936166Yeah same, updated everything including comfyui, tried various lora loaders including the standard ones - none of the loras work, even ones made by Black Forest themselves. I have a slight suspicion it has something to do with not being able to fit and allocate everything in memoryt. 8gb vramlet using Q4_0
>>101937221What precedent? The alternative was they paywalled it behind their Patreon or whatever gay service they are using, you weren't getting it 8 days earlier either way. People who are in things for profit are in things for profit, shitjeetai wasn't some holy ground prior to this. Big nothing burger
>>101937235No one's forcing them to release the model. Or even come to an open source site to do so. They could host an API and fuck off to the platform they belong, oh, that's right, now it's civit.
>>101937224I've had this with highvram
>>101937240True. It's up to you, the consumer to decide if that effort is worth your money.
>>101937253>Or even come to an open source site to do solmao what
>>101936344Anon, there is already a megumeme lora on civitai
>>101937249Are you using dynamic thresholding? Turned out that was that caused the error for me
she said it
>>101937253ohhh noooooo, I have to wait an additional week to use this jeetslop lora instead of getting immediate access! what ever will I do, what if it's infinitely more slop than the other 50 slops of the day being posted?!
>>101937274I made my own megumin LoRA because I wanted to see how it behaved at rank 32. I spent less than 5 minutes right click saving images of megumin for my highly curated "dataset"
>>101937271If civit had branded itself a shop for models there wouldn't be an issue. It is supposed to be a place for users to upload open source models, a hub of sorts. Now it is violent that original intent. Of course users are gonna be pissed, unless you're okay with them close sourcing anything they want at will.
>>101937253Ask me how I know you have autism
>>101937295>violentviolating*
>>101936044I have errors everywhere when trying to load the MJ lora lol https://civitai.com/models/640247/mjanimefluxlora
>>101937295>Now it is violent that original intent.No such thing as free lunch, it needs to be paid for somehow and there are worse ways than monetizing access to the product.
>>101937295ignore it for 8 days you insufferable tard, no one cares
>>101937315you're not alone anonhttps://github.com/city96/ComfyUI-GGUF/issues/31
heh
>>101937288looks good anon, megumin will make magic great again
>>101937284>>101937284Speaking of jeets.Do you know what else that allows? Scammers. Scammers are now incentived to upload slop to their site to lure users to pay for their slop. I hope you are proud of yourself.
>>101937261I remover the force/Device on the model loader and it seems to be "loading" I guess, now I have this error kek >>101937315
>>101937347whats your problem man, this kinda shit is the most benign kind of monetization/buzz drain they could havewhat would you propose then?
we are at the point where you can make a fake e-girl and make money off retarded simps. they would never know she is fake.
>>101937275Oh yeah, right, been using modelsaplingflux + tonemap, without them it works. But man, it sucks being stuck with just cfg 1 and no ability to tune the flux results. Hopefully it will get resolved eventually.>>101937329>>101937315Turn off everything that messes with the model
>>101937347Non issue, Civit will end up refunding the jeetbuxx if scammers slip through. You were more likely to be scammed by having to join someone's paid discord or whatever dumb thing they used
Lux gen on Civit. Retards can't help themselves can they. We will have full AGI and you will still be talking to it in 1.5 prompts
>>101937370>they would never know she is fake.This was never an issue. Even in 1.5 days there were "AI influences" big uncanny 1.5 anime faces that had Indians and boomers simping in their DMs
>>101937369He's a broke autist whose adverse to change and has no patience to wait a week for his fomo slop. No actual thought towards the subject at hand has been made by him, it's all kneejerk emotional reactions
>>101937099Does the model have some prompts it bans or what?
>>101937370People say this kind of shit like it hasn't been happening for literal decades already.Oh no AI help me photoshop!
>>101937384>(REALISTIC PERFECT AMAZING OMG 69K HD ULTRA DEFINITION ORGASMIC MASTERPIECE:999)I am glad that is being phased out
>>101937388it's neat how fast things have progressed, imagine in 5 years with more advances in nvidia GPUs.
>>101937372>Turn off everything that messes with the modelsuch as?
>>101937399I am praying that in 5 years I will never have to use NVIDIA again.
>>101937392nah
>>101937372>Oh yeah, right, been using modelsaplingflux + tonemap, without them it works. But man, it sucks being stuck with just cfg 1 and no ability to tune the flux results. Hopefully it will get resolved eventually.oh ok, then I'm not gonna use it without tonemap, I want to do some CFGmaxxing aswell, without it it's boring and slopped as fuck
>>101937410that depends on Intel and AMD finding a way to beat CUDA.
>>101937404Can you please for the love of fuck tidy your workflow. It triggers me every time
>>101937404Re-read the whole post, tonemap node fucks with loras
What do you think Gadot's new movie is about?
>>101937399the sad part is there are plenty of advances in jewvidia GPUs, they're just gatekept for profits on a level that should be illegal
>>101937420
ドッ ドナルド様
>>101937373>>101937369It's not a problem until the site inevitably becomes unusable with models that are complete replicas of each other (that is, they are based off synthetic slop taken from other models) or based off of cheaply stolen ideas.
>>101937419Not holding out for AMD either. They are NVIDIA's controlled opposition.
>>101937425is that this 327mb for no reason LoRA?https://civitai.com/models/647203/gal-gadot-flux
>>101937441until?
>>101937416DamnAnyone knows how to change the font size of the prompt text? I edited the user.css with element.style { --comfy-textarea-font-size: 20px;}But it gets overwritten by html when opening comfyui
>>101937399>imagine in 5 years with more advances in nvidia GPUs.it's been 6 years we're stuck with 24gb cards (since the RTXTitan), we won't go further, 10000 dollars 48gb entreprise cards are now the bread of butter of Nvdia
>>101937429the 7800x3D is such an amazing CPU in terms of watts/performance, idk why they can't engineer something for Radeon that is as good.
baker?
>>101937441>becomes Have you ever once actually looked at Civitai? How fucking new are you?
>>101937448unfortunately yes, razzz must have a whole pipeline to shit out these celeb loras faster than anyone else>>101937454>are now the bread of butterGOOD MORNING SAR
>>101937432我々は再び偉大な国にする、めぐみんちゃん…
>>101937479but it's noon
>>101937448It's that size because the LoRA was trained at rank 32
>>101937488It's 6 am retard.
>>101937372lmao, even stacking two loras on top of each other breaks everything. Basically can't attach anything to your model besides a single lora for now.
>>101937465here's your order of bread>>101937495>>101937495>>101937495
someone made a new thread but I'm mobile posting so too lazy to link it. please check the catalogue at your earliest convenience
>>101937466SD 1.5/XL slop is very different from Flux because Flux is a decent model and actually knows concepts.
Some employee needs to leak the grok2 model so we can all run that shit locally...it has very few guardrails.
>>101937532Good news! It was already leaked to the public.
>>101937532do you mean the text or the image model?
>>101935309Can you spoonfeed me a Flux colab, site or something? Can flux make porn?does Flux worx with an AMD card?
Everyone left, time to post my ugly gens
>>101936729extended tnd pasta x 3