Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101704253>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/ComfyUIInvokeAI: 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>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>Kolorshttps://gokaygokay-kolors.hf.spaceNodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper>AuraFlowhttps://fal.ai/models/fal-ai/aura-flowhttps://huggingface.co/fal/AuraFlows>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>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>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
Blessed thread of frenship
always a pleasure seeing your own gens in the bread collage
Hello /long dick general/, is it correct to say that flux schnell is a worse version of flux dev for people with smaller GPUs? How much worse is it?
>>101709609Dev is better for coherence but takes twice as long or more to generate. The wolverine pictures above are Schnell. For text you need to use Dev but for regular text-free images Schnell seems fine.
>>101709662>>101709595it's like warhammer fantasy movie done with 80's aesthetic
>>101709690its a character description of Mohg The Lord Of Blood from Elden Ring. I'm cycling through the character wikis
>>101709625Thanks, I had been using schnell this whole time and was already very pleased with it.
>>101709433Wish flux wasn't so bad with generic styles. Why does it give them all an anime style? Even when prompting for manga. It's uncanny, they went overboard with the DPO.
I like idealised depictions of ancient China.
What scheduler and sampler do you recommend for Flux (Schnell)?Also, this is a big sheep.
Can someone make me a realistic milf with big tits wearing a dress?
Hello
>>101709974ye probably
>>101709924These painting turned realistic 3D cause huge uncanny valley for me.When flux is trying to imitate 2D styles and failing due to being mostly trained on 3D, it accidentally does the most convincing (and thus uncanny) 2D -> 3D conversions I've ever seen in image gen. Compression and pixelization artifacts are also the best.
>>101709974
Replicate forcing the safety filter on and then writing:>The safety checker is intended to protect usersWhat if I don't want any protection? I want to live dangerously.
>>101710056This better?
>>101710118please understand
>>101710118By users they mean shareholders
>>101709924Great image senior, I was truly a frog in a well.
>>101710090Tits look fake>>101710153That's nicer, is that flux?
>>101710282>That's nicer, is that flux?it's bstaber
>>101710282>That's nicer, is that flux?Nah just 1.5 mix
>>101710124actually there's nothing wrong with those uncanny photorealistic paintingsthis one looks uncannily like a Chinese painting made by Shishkin, lmao. (and 3D is still a bit leaking). This model is incredible.
Flux Schnell img2img:Denoise up to 0.81: very few changesDenoise >= 0.82: almost entirely different imageIs this normal?
>>101710318On flux1-dev-fp8 when img2img with euler karras 6 steps, denoise 0.85 works well, but denoise 0.9 turns the picture into blurry mess.
Sorry if this is a stupid LoRA question, but... How do people make LoRAs with specific objects/options constantly appearing?Like, I would like to make a LoRA where each picture generates with a similar border around it (like a polaroid picture frame with some tape on it) and have it generate the image inside of it. Do I just do a regular style lora and tag everything BUT the picture frame?
It's really frustrating how Nvidia is holding consumer AI back by being so fucking cheap with vram. THIS IS NOT OKAY!!!!
>>101710373>Do I just do a regular style lora and tag everything BUT the picture frame?you can still tag the frame, it will just need the tag when generating, and tagging is always the better option because by not tagging you risk the model not really learning that thing
>>101710310I love me some Shishkin. He and Vereshchagin were staples of my gens back during 1.5 times. These gens I'm posting now of course use no artist tags at all.
>>101710373I think you want to tag the frame since SD should know polaroid style photos. This kind of lora could also work without tags, try it out.
>>101710394it's a monopoly issue (AMD doesn't compete) that should hopefully be solved when intel makes more GPUs
>>101709772Now make the loli a frog instead.
>>101710124I like it alot
What is guidance and what are the min and max number for it?
>>101710480>intel makes more GPUsDon't hold your breath.
>>101710480Intel would be lucky to not die and be saved by the American government in the next years, they're in no position to invest in better GPUs.
I like flux, but the fact it can never get better is really hurting my hype.If I can't train a LoRA or find better fine tunes, then what's the point? It's just a toy. Can controlnets even be trained for it?
>>101710978it's been 1 day, relax
>>101710978>then what's the pointIt helps me with some stuff for my illustrations, im a digital artist
Bros, why don't we just make our own models? Imagine a model as good as flux but low on vram and incredibly easy to finetune. WHOA!
>>101711009I'm sure we all have the hundred million USD and knowledge/dataset to do the compute.
>>101711009Just spend 10M bro, just hire a team of 12 phds bro
>>101711009>we (read: someone who isn't me)
>>101711009bigma
>>101711065Bigma who?
>>101709595Very nice
>>101710978Let's just wait and see. At the very least Flux is proving what is possible. This is the first model that makes me think the output is good enough. It's not perfect by any means, but it's good enough. And the fact it can run even on 12gb is great news. In the coming 5 years, VRAM should be more easily available and we may have cards specialized in AI as well, and there's also still a lot of room for quantization. In 5 years it's entirely in the real of possibility that mid sized communities could train a model comparable to Flux, so even in the worst case scenario, I think we will be okay. Fully customizable models will be a thing, although it sucks to wait.
Anyone managed to generate real micro bikini on this piece of trash?
>>101711337its not in the dataset, you can't
>>101711377FUCK
>>101711382Just wait. Porn models will be coming any day now.
>>101711382just like there is no genitalia in the data set, it is what it is.
https://x.com/ostrisai/status/1819865159227338946
>>101711397>>101711401I can and will simply wait a year.
>>101711406tfw no 16ch
The more I see from pony dev the more I realise he's a piece of shit. He and his dataset should be separated and given to someone with the balls to use it.
>>101711497What did he do?
>>101711504He's an anti loli moralist cuck.
>>101711504Every time a promising model comes out he screeches about the license then does nothing with it.
>>101711406This... this it is! HERE WE GO!!!!!
>>101711079bigma deez nutsnot going to be the level of Flux though
>>101711497Pony is the only good sex model>>101711519Make your own model then retardI'm a ossss developer and hate users so much, no matrer how mhch you give them they always demand more and more and never thank you, they just nag
>>101711406>...intentionally breaking down the step compression to hopefully end up with a training base that we can train LoRAs on that will work with schnellmedium hype
>>101711539No. Fuck you. You're going to make models for the community and you're going to make them good. Now stop this whinny bullshit and start working. And no, we won't say thank you.
>>101711519it also doesn't help that he has no competition currently, maybe if someone made a good enough finetune to replace pony he would do something about it.
>>101709433Flux is so good at text bros. How is it so much better than ideogram? Crazy.
>>101711536BIGMA HYPE
How do you get flux to put text on something? I want to generate a character that has a text tattoo.
>>101711746thank you, honey
>>101711009Just wait for Bigma. When it comes out 10B will seem like a waste.
>>101711791>A tattoo that reads "text"
>>101711791double quotes
>>101711746Amazing. What is the prompt?
what a time we life in
Man, I thought the advice to use an LLM to flesh out prompts was for absolute scrubs after I tried it for SDXL, but it really worked well just now when I tried it with flux. I used windows copilot and I suspect they must have used an OAI model to generate their tags because it really seemed to speak flux's language. Flux handles long prompts way better than SDXL, too.
>>101712010This is no secret. It's because of how the model was trained.
>>101711497>>101711519Previous gen datasets don't work with this model. Even if Pony could train the model right away, theoretically he'd still be unable to since he would have to manually caption the model for it to be meaningful (otherwise you're just going to be getting SDslop by teaching it booru tags and no one wants that), this an expensive task. Couple that with how much more exponentially expensive this model is to finetune. It's safe to say that it's likely outside of Pony's budget since a full team has to be hired just to do the captioning and to do everything else.
>>101712099It going to be tagged by llm. No human is going to do it.
>>101712131Look at this list-https://huggingface.co/spaces/opencompass/open_vlm_leaderboardThere is no fully uncensored VLM, he would have to make one himself
>>101712148He can't, he's a retard. Just a furfag with a lot of money.
>>101712310Offensively boring generation.
>>101712310Helen of Troy looking ass bitch.
>>101711753Lol
>>101712322Sorry to offend you
does the huggingface link for FLUX from the OP work with mobile? I want to send to someone to try but I'm on desktop
I'm speechless. Flux is just phenomenal. And it's only going to get better
Something Flux seems to excel at over other image generation is that most things are proportionate. SD 1.5 and SDXL struggled with this and SD3 continued to struggled with this.
>>101712683>forgot pic
>>101712683>And it's only going to get betterThat's where you're wrong, bucko. Flux is dead on arrival. It's an untrainable advertisement for their saas model.
flux is insane. i am a vramlett but good for you chads to be able to run such a beast locally uncensored. bless the makers.
>>101712800
>>101712800>>101712830Prompt for mixed anime and real?
>>101712842Stolen from /lmg/.>A professional real estate photograph selfie in a living room, 24mm, f/16 lens. The background is sharp and in focus. An anime cutout of Hatsune Miku is edited into the photo. There is a fat otaku man with manboobs standing beside her with his hand around her shoulder.
>>101712848lol posted my fat otaku edit.>A professional real estate photograph selfie in a living room, 24mm, f/16 lens. The background is sharp and in focus. An anime cutout of Hatsune Miku is edited into the photo. There is a photogenic man standing beside her with his hand around her shoulder.but its pretty much the same.
So is there any way to run flux now on 3080 ti 12gb that i have 4 of
https://comfyanonymous.github.io/ComfyUI_examples/flux/#simple-to-use-fp8-checkpoint-versionIf anyone is looking for an easy to use single file version of flux dev in fp8. Just make sure you update ComfyUI.
>>101712870>that i have 4 ofLLM?
Is there a way to run flux dev without comfyui? I have tried the example diffusers script and it doesn't even load the model.
>>101712874Ty
>>101712882Managed to fit a 70b model on all of em
How far we have come. I remember the 64*64 horror pics not so long ago.
>>101712756This is proof of what's possible with for image generation. And honestly, it's impossible for it to get worse from here on- it can only get better
https://twitter.com/ostrisai/status/1819865159227338946
>>101712966>forgot the pic...
>>101711825You can 't just force things to be over because you go and gen a fat lady singing. It's just not the right way.
>>101712756>>101712966i dont understand, why can't flux get better?
>>101712856>>101712848>The background is sharp and in focus.>Doubt
>>101713048speculation that finetunes aren't possible
>>101713048it can, just not in a reasonable way right now >>101712977think of it like this, what crysis was for pc gaming, flux is for imggen
>>101713115>it canNot a sure thing yet.https://x.com/ostrisai/status/1819865743929790533
>>101712756This is you>>101713055Flux loves the blurry background. I bet SD fags especially are bothered by this. I think you always needed to explicitly prompt a blurry bacckground. lol
>>101713071From what I've heard it's just a hardware limitation. In that case, consumer level hardware will grow into it in a few years
>>101713244and something about distillation from the pro modelI dunno, I'm not well versed in this stuff
>>101713244>40gbs for a lora>consumer level hardware will grow into it in a few yearsnot with nvidia at helm
finetuning wont be limited by architecture but hardware. people cried about sdxl finetuner thanks to the refiner but we still have sdxl finetuners and barely any need the refiner. the problem is the model is so fucking big that nobody can afford to finetune it. sdxl already took months to get semi-decent finetunes. with this it will be years. nvidia has local by the balls, you need $100k+ to seriously consider finetuning this
>>101711518Ponies are lolis though, just in pony form.
>>101713291>nvidia has local by the ballswhat happened to trust busting?
>>101713253how do you get accurate copyrighted characters like Goku? mine are always way off. do I just need better models?
>>101713279Whatever happened to QLoRA (which is now QDoRA?) Don't tell me you don't plan on running Flux Q4.
>>101713384Unironically write the prompt as if you were praying.
>>101713384I just wrote "with Goku from Dragonball Z"
>Create anything, just make it really cool to impress everyoneNLP might indeed solve my issue with ordinary styles
>>101713384With flux its easy.
>>101713291Nvidia should unironically be forced to license CUDA. It's pretty much a form of vertical integration that is giving them a virtual monopoly.
>>101713384>filename huh
>>101711065
>>101713456>>101713456Most of mine are ending up with 3D characters, how to I make them 2D? Just roulette?
>>101713486>A professional, high-resolution photograph inside an opulent mansion, 24mm wide-angle lens, f/16 aperture for deep focus. President Donald Trump and Russian President Vladimir Putin standing side by side, appearing photogenic and formal. Behind them an anime cutout of goku from dragon ball Z is edited into the photo. A thoughtbubble above their heads that reads "Flux is cool..."
>>101711079
>>101713291So- out of the frying pan and into the fireIf we don't get any extras for flux for a year or two then I think I can cope. Who knows how far we can get with just by improving prompting methodology. And there's always img2img with flux as input if it comes to that.
Is PAG node worth it? It doubles the gen time
>>101712874>17GBWhy is it bigger than this ones? https://huggingface.co/Kijai/flux-fp8/tree/main
>>101713558my thoughts: >>101686791 we wont get anything insane for flux in a year but by that time flux will be outdated. it was only a year ago that sdxl released. right now the local limit is around 8b. 12b is simply too much for any kind of ecosystem to develop around it. flux is a nice model that proves what's possible, but it's unwieldy thanks to gimped local hardware.8b is the sweet spot for the 24gb high-end. a good model baked with 8b as the target, not distilled, will be local's next mass-adoption. i don't think what flux does is only possible on 12b, i think the comprehension and more can be fit on 8b especially if they improve their datasets.
>>101713291All that does is give birth to or create a need for finetuning services. A pretty smart business model if that's something flux devs are considering imo.
The truth is: I don't care if Flux gens are accurate to the prompt when they're all uglyThere, I said it
>>101712874Why do we need this as opposed to just doing this?
>>101713634is this not just renting out runpod or something? i mean for an actual finetune (not lora) you'd easily need money in the 5-digits minimum
>>101713644You can do it that way too if you want. There's just a lot of people confused about the multiple files.>>101713596That's a fp8 version of the diffusion model only, Mine has everything in it so you can use the regular "load checkpoint" node with it.
>>101710630
>>101712618>does the huggingface link for FLUX from the OP work with mobile? I want to send to someone to try but I'm on desktopplease, is anyone here on mobile who can check?
>>101713947
>>101713870>>101713644>>101712874Produces different results. Left is yours, right is with weight_dtype
>>101714093forgot pic
>>101714093>>101714108fp8 T5 vs fp16 T5?
>>101714014how are you prompting pixelart? just "in the style of pixel art?" or "pixelized"?
>>101714137I guess? I don't really know what I'm doing. It's like this. The rest is the same.
>>101714151You are free to implement such a feature.
>>101714151it's pretty jank. stenography would be better but has risks of hidden files
>>101714146Literally just "pixel art" at the front of the prompt.
You are forcing me to download ComfyUI.
>>101714235Foxfucker has monopoly as tight as nvidia.
>>101714235technically voldy is
>>101714260prompt? I know cutout of miku works for mikus with real people
>>101714235>forcingDon't like comfy? or just can't be bothered?
>>101714235you can set up the workspace by loading a file or dragging an image. it's honestly very simple.
>>101714256I think that auto1111 is based for not giving so much of a shit about his webui anymore, considering how little you get in return for open-source software development. As time goes on, the effort to maintain the project increases, and so does the amount of annoying comments, demands, and criticism that you receive (maybe this comment itself fits in the "annoying comment" counterproductive category).Meanwhile, comfy's efforts are commendable. He works nonstop on his project and is very passionate. I'm glad he figures out how to run the model for me so that I don't have to spend weeks experimenting myself. >>101714292kek yes.>>101714334I agree. The UI is very fast as well. But it feels nicer to use auto's UI, at least to me. I'm used to Gradio UIs and find them nice for their amateur look and big buttons.
>>101714355for flux which has no stuff like controlnet/etc yet, it's fine. I like auto for sdxl/ponyxl and making characters or doing controlnet stuff like canny meme edits of photos with characters/loras.
>>101714311"Photograph of Donald Trump smiling and hugging a body pillow. The body pillow has a large anime drawing of Hatsune Miku on it." It can do this kind of stuff surprisingly well.
>>101714413neat, cutout also works really well
>>101714344Neat
>>101714445I love how Donald Trump has become the new Emma Watson for testing shit lmao
>>101714516how did you prompt text per panel or panels in general? 4 panel comic or something?
>>101714531anon, just describe LITERALLY, there's no issue with that. literally describe that it's a comic with N panels, describe each panel, just PROMPT IT NORMALLY
>>101712286That's pretty funny, catbox?
>>101712870Might have found something myselfhttps://twitter.com/cocktailpeanut/status/1819482952071323788
>>101714531It's very simple as >>101714537 says, all I'm using is>A comic scene with 4 panels. In the 4th panel (the last one), the beautiful anime girl is holding a sign that says "..."
>>101712010vlms are used to caption their datasets, it's why all these models respond so well to llm sloppa
>>101714537
>>101714454Thanks friend
>>101714720Ironic
>>101714750that's not safe
>>101714751
Interesting...https://files.catbox.moe/bdr8uh.png
>>101714760
>>101709433What is the best model to generate images for development purposes?I'm trying to get realistic images and they are very good, I got these results at https://fal.ai/specifically whit fal-ai/realistic-visionI would like to know which one is the most recommended in terms of cost and ease, I probably want to create images very similar to an image I already have, that is, have the model generate another one with the same person that is in the input imageI can probably fine-tune it and/or deploy it to a server or maybe use an already existing model and use the api to get the image I wantI would like to know the best way to learn
>>101714789Very christian of you
>>101714901this reeks of shilling and llm
>>101714912christianity is when ned flanders
>>101709433Does anyone have the full size top center image?
Is there an alternate GUI to run flux locally besides comfyui
>>101715019stable swarm UI, which is an auto1111-like UI on top of the comfy backend
>>101715034Thank you
>>101714531That's the magic of good prompt adherence, you can literally just write exactly what you want in what part of the image.
>>101714739Catbox? Love this
Why is SDG lapping us? I thought we were winning?
>
How many steps do you guys do for complicated gens? I just tried one with a really detailed prompt and it followed it better and got the text more accurately with 50 instead of 20 steps. But fuck, it's so slow.
cozy thread
>>101715240I don't need to 'beat' them, it's just nice to have an alternative imagegen thread without the insufferable avatarfagging and lonely people talking about their personal lives using imagegen as a pretext
>>101715290
copey thread
>>101714750They're gonna shut it down before we get the video model. I can already tell. The video model + this + a local text to speech would be too much for them.
>>101715399They don't care about dead babies. Only what you can create, share and expose with it.
>>101715377who is them?
>>101715434>>101715333
>>101714235I mean, if you go to their github instructions you can run a limited webUI
>>101715377What is it like to live your life knowing you are completely retarded? Does it hurt? Are you depressed?
you guys had me all afraid I'd need a billion 4090s to run Flux, it works just fine on me 3060
>>101715506wouldn't hurt tho
>>101715506What's the gen time for one image? I'm contemplating whether to download the 24 GB blob or not
>>101715506If you want to run it at full power you will need a 4090 but fp8 is good enough, and significantly better than both XL, Pony, and SD3. If it could make a vagina we'd be able to delete every other model.
>>101715493>He doesn't get the motive behind (((safety))) and censorship
>>1017155182.5~ minutes for me
>>101715531It is though and by a lot
>>101715518it's about 1 minute for every 10 sampling steps, so 2 mins for the 20 steps comfy's workflow uses>>101715529I'm doing everything fp16 (model and clip) and it works, uses almost all my vram though
>>101715534You dumb idiot. It's already out. We can already animate all this flux shit. There's nothing you or your conspiracy theories can do to stop it.Flux successfully proved you can make a "safe" model not shit.
>>101715574>almostFuck me. On my 4080 it takes 10-20 minutes per image if I run it on full.
almost got it :(
>>101715580Retard, where did I say anything about Flux?
>>101715608Now I'm confused, my vram usage is the same whether I select fp8 or default bfloat16. wtf
>>101715679The image you replied to was flux, it's a safe assumption. What do you think they are going to shut down? The internet?
>>101715636kek
>>101715636HOLY SHIT ANON I WON THE FUCCKING DIFFUSION LATTETY NO WAYYY>Breaking news on NBC showing a reporter on the left side with Trump making a speech on the right side. The text on the bottom reads "BREAKING NEWS", and below it "HIS IMPERIAL HIGHNESS ANNOUNCES TOTAL NIGGER DEATH"I SHIT YOU NOT THIS IS A DEV GEN FROM THE API SCALE 3.5 SEED 51646902 (replicate API flux dev, you can test that it's the same, AR 16:9)This gets the text right like 25% of the time
>>101715518Trying Flux for the first time on my 2060 12GB, looks like about 4-5 minutes for a single 1024x1024 image (also first time using SwarmUI and don't know how to see precisely how much time is estimated)
>>101715705catbox your workflow?
>>101715727KEK
>>101715727incredible
>>101715727what a time to be aliveoverlay prompt?
>>101715733Correction, exactly 6 minutes
>>101715776The whole prompt is there
>>101715776The prompt I shared is the complete one, I did in on replicate api
>>101715740it's just the default flux workflow suggested by comfyanon https://files.catbox.moe/fpiehr.png
>>101715717Retard, the US company behind Flux. If it were Chinese it wouldn't matter (except to Xi).
>>101715793You must be hitting lowvram mode and bleeding into RAM is my guess.What card?
>>101715795The US company behind flux that just got 40 million dollars? That company? They're going to shut it down are they?
>>101715803yeah, it is saying lowvram mode but the speeds aren't bad at all compared to what some people have reported, ~2 mins per genRTX 3060
>>101715832How many steps do you use?
>>101710027We have no donations here Reimu sorry
>>101715832How the fuck are you getting 2 minutes per gen? This is me on my 40801024x1024 gen fp16
>>101715832at around 1MParound 2 it's 8~ minutes
>>101715849What resolution are you prompting?
>>1017158671024x1024 I said already
>>101715727
>>101715727fucking kek
>>101715877how are you getting it so clear? my broadcasts are a bit blurry, this is fp8 on a 4080 btw
>>101715889sorry I'm using replicate api which is probably fp16, dev model.
First Flux proompt, 6 minutes generation time on a 2060 12GB. Going to take a while to figure out how to get the styles I wanted.
>>101715877finally got one that works myself :D>>1017158471 minute per 10 steps, so 2 mins for a 20 step gen>>101715849idk anon... are you on windows? Only thing I can think of is that it might help that I'm on linux, where my vram is almost empty until I start up flux
>>101715910imagine how many boomers on social media can be tricked with this
Ok it's over I got the perfect one>Breaking news on MSNBC showing a reporter on the left side with Trump making a speech on the right side, standing next to the flag of the USA. Behind Trump is the White House logo with the text "THE WHITE HOUSE" and "Washington" below it. The text on the bottom reads "BREAKING NEWS", and below it "HIS IMPERIAL HIGHNESS ANNOUNCES TOTAL NIGGER DEATH". On the bottom right is the text "MSBNC" with the logo.
>>101715933probably easier and more effective to grab any screenshot of a story about trump and edit the headline
>>101715937Forgot pic, and his face is a bit bad after all, but yeah
>>101715939is it tho
>>101715939inpainting with this model like in auto1111 or an inpaint node would be fun too
The difference between fp8 and fp16 is very noticeable.Fuck. Now I need to buy a 5090. :(
>>101715965I've tried both and they're basically indistinguishable. Even when I think the image is screwing up due to being FP8, changing it to FP16 doesn't help.
>>101715979what about fp4 for the ramlets
for scheduler (basicscheduler), some have said "beta" is also good results, like euler
>>101715889Set cfg to 1 if its local
It didn't want to say Pooping news, and three s
>/g/ - Technology
baking
>>101716006I know Flux is nice but >>>/pol/ is that way
>>101715992I have a basicscheduler that outputs sigmas (options are scheduler/steps/denoise, no cfg) but if I use ksampler it outputs latentwhich node do I need for flux?
>>101715965vs pro holy shit fuck me
Le Fresh Bread...>>101715949>>101715949>>101715949
>>101716023nm most of the workflows I see just have the node with 3
>>101714445Can it do kigurumi?
>>101714355One is a slav and a Dota 2 addict and the other is a massive attentionwhore Of course only one will keep going long term because his ego depends on it