Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101678250>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
bless thread of frenship
>>101681371I only use 12GB when genning in comfyui.
bros, is it over for the faggots over at SAI???Unironic question
at 512x512 speed is comparable to SD and is still pretty good
So much fresh bred in a simgle day
>>101681400I feel mildly sorry for the people that didn't want to fuck up but got "overruled" by the leadership which also wasted tons of money (not on their wages).But yes, the alternatives are better, and I don't see much of anything SAI has going for it - the competition is currently better.
>>101681428>don't see much of anything SAI has going for itan image to 3d model no one asked for kek
>>101681383>blessblessed
>>101681535
Looks like they will cooke a finetune of flux dev, nice
>>101681583Pixartfags on suicide watch
the speed of schnell is quite impressiveI shat on it before but I think I was giving it too many steps and getting deep fried images as a resultI'm used to SDXL turbo models lying about only needing 4 steps when they actually need more like 6-8This one actually needs only 4
>>101681440Maybe a future better variant, but I don't particularly expect it to be coming from SAI.Even then I wonder if it's not still an ultimately 2D-ish imagegen component (to re-texture the same generated object) that is going to be just as important as generating the 3D object itself.
>>101681583It can't be finetuned retard, it's a dead end for a lot of reason
>>101681598>This one actually needs only 4but the quality is worse than flux dev so...
>>101681598Two (2) steps
https://reddit.com/r/StableDiffusion/comments/1ehotwi/it_seems_fluxdev_is_another_model_i_have_been/I wonder what is his prompts, he nailed the different styles
Im going to kill myself
me too
>>101681594I don't see the issue for Pixart at all if this can't be finetuned on consumer hardware.
Guys, I think I found a sampler that can replicate styles well, Euler Karras
>>101681684 >>101681702But why?
>>101681711huh, karras does seem to bring out a more painterly stylethis makes no sense, the noise scheduler shouldn't have any effect on that
friendly reminder to attach the sound effect node so it dings when it's donethat way you can switch tabs until you hear the ding
>>101681765https://github.com/kijai/ComfyUI-KJNodesadd status icon to tab, as an alternative
>>101681583will donate next month
All the people that was training before this scale, were prepare only for 8B, so this suddenly movement of flux, will define the scenario. There are many companies, that will pay for improve a open source model that is in the tier level of Dalle3, and there is more room to enhance. Regard the license, you can monetize trough KOFI, or similar just like those are doing right now, and therefore, a company could donate to your project without many announcement. Is just a question of time, to many loras and finetunes will work with this.
>>101681583why do furries have so much money
>>101681601SEETHE COPE DIAL AN 8
Finally a model that can oneshot this prompt (besides Dalle)>A color photograph of a young Japanese woman in a cropped top typing on a retro computer in a dimly lit room. She is holding a gun while typing. The image looks like a VHS still, slightly faded with film grain. The room has a nostalgic feel with vintage decor and low lighting, casting shadows that create a moody atmosphere. The overall scene is retro and slightly mysterious, capturing a unique blend of past and present elements.
>>101681789It took great focus to muster through this post but after doing so, I have come to agree with you.
where's the 5000 series nvidia?...
>>101681583what have they done previously ive never heard of these guys before
>>101681806Elon just bought all. We wont get any
Flux has officially revived the image generation scene after being stale for such a long time (half a year)Can't wait for some finetunes to come out.
>>101681798Slightly better but lost VHS feel kek
>>101681817Totally this. Any idea what hardware requirements are for finetunes tho?
Yo
>>101681826Seems like it's pretty high, so it's going to take a while to see anything worthwhile probably.
>>101681733vramlet :(
>>101681840yea I thought so, well at least the autists already have their finetune datasets, all they need is to rent some server time to get it done, 2-3 months for the first nice ones to appear?
>A 1990s anime still featuring a woman with an 1980s hairstyle and look. She is pouring water from a bottle into a glass in a hotel room. Behind her, a blue light shines at the room during the day, creating an interesting contrast. The scene captures the essence of 1990s anime with its distinctive art style, vibrant colors, and detailed backgrounds.Nice>>101681817We have been waiting for someone more sensible than SAI to step in for years now, first the Chinks (and we know we have their dedication as they will release models on par with Flux Pro soon) and now Flux, it truly is a relief.
>>101681866How much vram and ram do you have?
>>101681817i think it's a very nice model, but there was constant progress either way>>101681866download more vram? maybe there'll be other solutions later
>>101681888>1990s>2000s stylemodel is cooked stylistically
>>101681866Give it some time.Remember when SDXL took 14gb vram minimum to run? You can now run it on a 4gb potato just a bit slower.
>>101681900i hope sooner rather than later>>101681910>Remember when SDXLmy thoughts exactly. the only question is when
>>101681907nta, but whatever ya do, anime styles are not well understood by flux
>>101681910>Remember when SDXL took 14gb vram minimum to runthat was for the full FP32 model. to run this model that way would require 48gb+ vram
It's kinda disappointing that it ends up being based around 1024x1024 again (with quality drop offs at higher res). Are we ever going to move to a higher base res?
>>101681922strangely I experience better results on schnell than on dev in res above 1024x1024 (schouldn't they have the same data set?)
>>101681922Once consumer GPUs catch up to datacenter GPUs in terms of vram we'll move on from 1024x1024.For now it's simply more practical to gen at low resolution and upscale.
>>101681922I think that would shrink the dataset considerably, the number of 2k or 4k images is much smaller than the number of 1 megapixel imagesYou'd probably end up having to cheat by having a lot of the dataset be 1024 images upscaled with ESRGAN or something, in which case what would be the point
>>101681916I think flux-schnell at least is easily fast enough that people will let you use this for cheap/free or via ai horde "trades" or w/e even if nothing changes.
>>101681907True, is training code released? If someone takes it and finetunes it on thousands of different styles then we might be back.
>>101681888>interestingWhy the fuck do tech inept people add this to their prompts? Do they think he AI can do something meaningful with an adjective like ”interesting“?
>1980s, retro, vintage_anime, anime_(1980s), takahashi_rumiko, adachi_mitsuru, hojo_tsukasa, hagiwara_kazushi, buronson, kurumada_masami, toriyama_akira, takahashi_yoichi, 1girl, a fantastical scene depicting a woman in swimsuit sitting on a hill overlooking an airport runway on a rocky outcropping in the middle of the ocean. A violent and powerful storm is visible in the background, and a giant wave is crashing over the rocks at the runway, 1980s anime tv episode still
>>101681947I also got the feeling that some data subjects are more focused on hires, if you prompt nature scenes, trees and the like, you can get incredible detail, but I tried night skies yesterday and got horrible rasterization effects on hires. Also ofc 2D was not their focus
>>101681976Might as well use "bizarre" instead because it actually does something interesting
why does karras work for schnell but not for devthey must be quite architecturally different
post ghibli inspired miku
>>101681977in most of my trials it just borks on booru tagging, 1girl and artistname_name will not work well.. in the style of Akira Toriyama better than toriyama_akira
>>101681976I mean this is exactly how a VLM talks
>photo realistic spooky castle at night on a mountain from a distance red light emanating from the windows a narrow winding road runs upwards towards the castle moon light shines bright illuminating the sceneI'm glad I learned about this model, I'm having a blast with it.
>>101681998Tero isn't bossed around by some lapland burners
>>101682006I like to see what it does (by adding booru tags you add creativity to the model)
>>101681998karras does work for dev, it just requires way more steps to convergeyou can't get away with 20 like you can when you're using simple or sgm_uniform, it has to be 50 minimum with karras(it took me a while to discover this because that's the opposite of how it works for SDXL, where karras is the 'faster' scheduler)also that other anon seems to have been correct that when you use karras, you get way better art styling, which is fucking WEIRD. the scheduler should not control that
ow nice, finally diorama views with individually individually recognizable features and not blurry mess>>101682046ya I get you, doing that still to, just to see how much of booru tagging sneaked into it, or confure the model abit .. still the lack of recognized anime artists is sad, but I guess finetunes will arrive. God bless its apache2.0, no one can take it away from us again
>>101682046>1980s, retro, vintage_anime, anime_(1980s), takahashi_rumiko, adachi_mitsuru, hojo_tsukasa, hagiwara_kazushi, buronson, kurumada_masami, toriyama_akira, takahashi_yoichi, 1girl, pink_hair, space_suit, floating_in_zero_gravity, surrounded_by_aliens, playing_chess_with_a_robot, holographic_chessboard, neon_lights, space_station_window, asteroids_passing_by, distant_planets, glowing_computer_screens, shiny_metal_surfaces, high-tech_gadgetry, futuristic_headset, sparkling_star
>>101682056>acrylic painting of a catleft: euler sgm_uniform 50 stepsright: euler karras 50 steps
love that it has anonymous with the v for vendetta mask
>Even people in the background are perfectly generated. Finally
>>101682084can you change scheduler mid generation? or alter with every step.
>>101682084I think we unlocked the styles at flux-dev, just use karras bro! kek
>>101682067how many steps anon?
ai pandering
>>101681831moar
>>10168213410 on schnell
>>101682192Does 6 work? I was finding schnell tended to stop improving at that point.
>>101682188
>>101682211same on 6 .. maybe less details? also my PC just crashed.. not getting the same result on 10 either, grr
>>101682261for the science .. this on 10
>>101682261>>101682192what kind of prompt are you using? I'll do it on 50, i'm curious if it will fix the fucked up cars and other oddities
>>101682272>>101682261VERY cool
>>101682287would be:>utopian science fiction city, diorama isometric view, in the style of Akira Toriyama, anime, capsule corp.euler, normal scheduler, cfg 1.0
>>101681353My one grip with flux (aside from the artist issue) is that it also produces SD 1.5 faces unintentionally, just like all SDXL mixtures and like SD3 did on release (almost as if Lykon had his hands on the model). In comparison, Pixart doesn't do that
>>101682299Ok will try that, this gen got the view wrong.
>>101682309ya you need to combine isometric view with diorama, or you will get a random perspective, also pic related is Toriyama replaced with Giger>>101682289thx
>>101682306Supposedly flux was made by same team that made SD3 (that left when SAI went collapsed). Now, this is not necessarily a bad thing, since it can do both SD-slop and SD-nonslop as it should, though right now it can do only do so its own. It just means we need to tune the alignment away, after all this is a DiT model.
I am just amazed that even tho the resolution isnt that great it gets every individual stair step correct
>>101682406Yeah, the model is extremely good with fine detail and stuff in the background
>>101682430very nice
Is there a certain way to prompt with flux, or is it the same as SDXL with just keywords seperated by commas?
>>101682406Can you add some utility poles and power lines? SD usually shits the bed with these
>>101682306That won't be an issue for long if it's easy to finetune.
>>101682439100% natural text, describe how you want the image to look. It's using a fully featured language model though, so it is pretty flexible and can incorporate tags if need be.
>>101682459I see. Thanks!
What artists does flux know.I tried rembrandt and it doesn't seem to work.I'm using "In the style of Rembrandt van Rijn", should I word it differently?
Nice gens this morning. Lotta creativity.
>>101682445does it reasonably well, I guess I have to remove some of the artists to make less of a spider web, the prompt became a mess now>utopian science fiction city, diorama isometric view, in the style of Moebius, Jodorowsky, style HR Giger, style Zdzisław Beksiński, anime, in the desert, dune, Arrakis, brutalist architecture, Sandwurm, style Lynch, style Kubrick, overland power lines,
>>101682449>Finetune in progressI hope. Once the code releases for Kohya a simple LoRA or 2 might do.
>>101682445cleaned up, ya impressive>utopian science fiction city, diorama isometric view, in the style of Moebius, tyle HR Giger, anime, in the desert, dune, Arrakis, brutalist architecture overland power lines, utility polesthe place basically became a dune power plant now
>tfw 12GB VramletIt's over for me isn't it
>A beautiful acrylic painting of batman and spider-man sitting at a bar having a drink. Wolverine walks through the door while holding a watermelon. Wolverine yells "Wubalubadubdub!!"Didn't do the speech bubble unfortunately, but this model is going to revolutionize memes. Especially once some good finetunes come out.
>>101682550yes>>101682551add speech bubble and he might actually do it
>>101682550You can run it without 24gb vram as long as you have a lot of ram.I'm running it with 10gb vram and using about 20gb of my 32gb ram.
>>101682563>add speech bubbleI didn't in my previous prompts and it worked fine, but I'll try.
>>101682550maybe fp8 works? maybe if you load t5 to system ram?https://huggingface.co/camenduru/FLUX.1-dev/tree/main
>>101682574I'm not having any look with the view. Perhaps i just need keep going through seeds, or perhaps its because I'm using 50 steps.
>>101682550just make a big swap partition or swap file. Itr works on my 12GB VRAM and only 16GB system RAM, but I've got a ridiculous swap partition setup on an old SSD, and its working fine. The only time I get issues with RAM on this system is when I start loading up like 3 SDXL models along with everything else with multiple ksamplers. Then I usually only have to close everything down and start fresh with just the one Comfy tab open.
>>101682550>vramletlol, im stuck on my 6gb laptop for months without reasonable funds to upgrade so I can only play with this shit on their free demo thing. it doesn't feel very cucked though to b fair.
>>101682670>he only time I get issues with RAM on this system is when I start loading up like 3 SDXL models along with everything else with multiple ksamplers. Then I usually only have to close everything down and start fresh with just the one Comfy tab open.In fact I think I eliminated that problem entirely when I set my fstab to contain discardUUID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx none swap discard 0 0That enables the SSD to do its trim routine properly on swap partitions
Holy shit, I figured it out. It's T5 that doesn't know styles and artist names. But CLIP knows them! It was always good for that (the whole Greg Rutkowski thing from the 1.5 days was due to CLIP).So, you put your artist name and/or desired art style in the CLIP prompt box.And in the T5 box, you only put your desired CONTENT. Don't say anything about the style or the artist to T5, because it doesn't know what you're talking about and it's just gonna make some shit up instead. Overwhelming CLIP's contribution and giving you the wrong style or the "slopped" look.CLIP box: Style/artist nameT5 box: desired image content onlyThis almost completely eliminates the inability to understand styles or recognize artist names in my testing so far. It was never the model itself that didn't know, it was just T5. With the above method it's now recognizing all kinds of names and styles that it seemed to ignore before.
>>101682551Yep, this is the moment we have all been waiting for. Isn't quite Dalle tier yet in terms of total concept knowledge but 80% of the way there. Crazy how the model just dropped casually kek, looking forward to v2 of this model.
>>101682740it's better in terms of prompt following and stuff
>>101682736Meh
>>101682509cute
>>101682639you using dev? I tried it there .. will work sporadically, on schnell it works 90% of the timepic related 10 steps, euler on schnell
>>101682736Big if true. Share workflow in ComfyUI
>>101682639>>101682785while this is 40 steps euler, same seed, same prompt on dev
>Gameplay screenshot of a 2D pokemon game. It's a battle between two creatures, one on the left is a large blue alligator and the other one is a large red anthropomorphic chicken with one leg in the airKek
>>101682795yeah i'm using dev
>>101682805What settings are you using? can't seem to get that style at all
>>101682521>>101682541huge improvement, ty for testing
>>101682795I'm gonna try reordering the prompt, since I had to do that to get it to understand from behind views. I had to use 'from behind' for image related.
>>101682808+1 for lotus attemptreally cool gens, I am hyped.
>>101682824I think I just fucked up with the settings, because is my 2º time with Comfyhttps://files.catbox.moe/ol9h9k.pngYou capture resembles better the output of pro
>>101682736heck you are onto something there, I am sure knows some artists, but this way you can force it on the whole prompt and txxl5 doesnt just think you are rambling on how Van Gogh is portraited in atmospheric ligh.. 10/10 important postpic related, Giger in CLIP
>>101682736>Using clip for styles >Euler Karras 50 stepswe're not there but we're getting close
>>101682232please share workflow or prompt?
Can someone catbox their comfy workflow with the seperate clip artist prompt?I'm new to comfy and have absolutely no clue how to get another clip box.
>>101682851holy fuck you use t5xxl fp8? this text encoder gets really bad at 8bit, I'm really surprised you got this gen with that lol
>>101682551>Wubalubadubdub
has anyone tested if it can do upscales? is it a SD3 situation where only ultimate SD upscale is viable?
>>101682881Here anon: https://files.catbox.moe/efb6nn.png
with "woman by da vinci" in txxl5 I get vaguely rennaince woman, but if do both: "woman by da vinci" in txxl5 and add "by da vinci" in clip I basically get the mona lisa instantly, I guess flux wants you to dual text encode for best effect
>>101682851yeah I just tried your settings I'm not getting what I want, I guess you got really lucky with that seed kek
>>101682881Not that anone but I messing around with it now, double click anywhere empty and search cliptextencodeflux top box is the style, second box is what you want in the image. Connect clip to clip and conditioning to positive on the Ksampler. I think that is how they are doing it.
>>101682938Thanks!
>>101682913you can render at very high resolution without duplication issues on flan, it's making upscale obsolete I guess
BOFT seems to require 20gigs of vram for training (or it just doesnt work well with prodigy)
For those with 24gb vram cards and are doing a 8bit DiT + 16bit text encoder, I highly suggest you to go for the --highvram command, it has enough room to load the both of them and it makes everythng way faster
tattoos aren't a complete mess. I love this
>>101682957I see. I'll have to mess around with it for a bit to get a grip on how it works. Thanks!
>>101682871What have I done?? kek
>>101683000>--highvram commandHow do I do that? Just as I load comfy?
>>101682574updated this to dual text encoder with styles enforced in CLIP>>101682736thank you so much anon you just made my day thank you very very .. this very important information on handling flux!
>>101683000nice tip thanks, no change to step speed but much faster prompt processing
What step speeds are people at 1024x1024? It's not great on a p40, about 20 seconds per iteration.
>>101683024you do that if you're using the "fast" method
>>101683042about 1.4 -> 1.7 secondes per iteration on my rtx 3090
>>1016830421.15 it/s on my undervolted underclocked 3090
>>101682871You could also maybe try these useful nodes?
Is the quality difference between schnell and dev very obvious or is it relatively small?
>>101682752
>>101683069I let you try it anon, you can download my workflow here >>101682938
So what models do i download if i have a 3090? Just the 23gb dev file and run with it?
>>101683087still debating on that, some stuff seems to work even better on schnell as the isometric stuff posted above, as schnell is the one on apache2.0 I guess we will see the finetunes on that?
>>101682752But can it do girls wearing bikini or underwear without huge boobs? I haven't managed to, so far.
1080p 168s on a 3090
>>101683115follow this guide, you need quite a few files>https://comfyanonymous.github.io/ComfyUI_examples/flux/
I still cannot believe we got dalle3 at home just like that
>>101683135ashton kutchers money was holding us back all this time
do we know if flux comes with a 16ch vae?
>>101683164says it in the release, yes it does.
Guidance down to 2.5 (not cfg, the guidance dial on the text encoder) seems to help reduce the slopped look for art gens too.Only seems to work on dev, changing the guidance does nothing on schnell.
>>101683171ty anon
Flux looking hot
>>101683214absolutely so
I hope finetuning/lora training isn't too bad so we can maybe get some good stuff this month.
1080p 25 steps deis sgm_uniform 86seconds
>>101683246i have a feeling the t5xxl is gonna make shit really painful to train in
>>101683257>i have a feeling the t5xxl is gonna make shit really painful to train in
>A woman doing breakdance at ParisI prompted that in french and it understands that well, nice
>>101683257You can process the prompts and then never touch T5 again for the remainder of the training run.
>>101682803
How do I fix the blurriness in some gens with flux?
>>101683409dont use cfg at all, flux doesnt support it set it to 1.0, use euler and normal scheduler
>>101683383I remember shitposting done on sdg with this prompt when dalle released, retards were coping with muh controlnetflux really is great even if it has styling issues.
How are we gonna cope with the fact that we'll never be able to use negative prompt on flux-dev?
flux team suffering from success
My Hunyuan style selfies still work with Flux, just gotta find the proper realism keywords or perhaps tune it, nice.
>>101682760has the saem issue as Dalle where the skin seems too smooth. Otherwise this is an amazing model, can't wait for people to figure out how to finetune it for more skin details etc
>>101683488Not quite as good but if I find the right keywords I will get there
>>101683509So far it gives me what I want but it's anime, guess it still somewhat works out
>>101683409>>101683425Thanks! That and using the clip_l as the style prompt and t5 as the content prompt made all the difference, with upping the steps to 50.
>>101683455by using the positive prompt more intelligently? I mean to exclude things from the image, words like 'alone' or 'uncrowded'. Be more specific in your prompts. I rarely use negative prompts these days since using pony models, and this thing. It well... it shits all over pony, I don't have to keep spamming queue for imperfections.
>>101683455is this true, or overexaggerated doombait? because i don't think it's good enough to do away with negative prompts entirely. already there's a bunch of ugly slopstylization i'd like to try and neg out.
>>101683586no it's true, this model can't make good pictures at other cfg than 1, and cfg = 1 also means you cdan't use negative prompts
My comfy just crashes when I attempt to run the workflow>12 GB VRAM 32 GB RAM>using fp8 vae and weightsPlease help before I shill out 800 bucks for a 3090
>>101683711>My comfy just crashes when I attempt to run the workflowmake sure you have 24 or so gb free on your C: drive
>>101683683cfg works, it just needs to be figured outsee the web demo
>>101683683>cfg = 1 also means you cdan't use negative promptsyou're pulling this out of your ass?
>>101683717Wait what, why? There's no way I can make that happen, I have like 10 there that I keep having to free every now and then
>>101683726comfyui uses it while inferencing. it will not work without said space
>>101683725you don't know how CFG works clearly
>>101683725no, that been confirmed, just try it for yourself, negative tags have zero impact
>>101683736Huh, I wasn't aware. I'll see what I can do. Or just boot into my linux repo.
>>101683742neither do you. one of the devs claim that cfg is not even actually used for flux.
>>101683751that's what CFG 1.0 means you absolute retard
>>101683099I just tried the conditioning (concat) node and it seems to work decent. The image is not exactly what i want because the dev model does not seem to understand isometric or Birdseye view but it seems to get the art style correct. So the way it works is you connect the image prompt to the conditioning_from connection and then the art style to the conditioning_to connection. This way the image is first considered then the art style is applied second, basically how concat works.
>>101683766give us one of your outputs through a catbox anon, so that we'll be able to get your workflow
>>101683751setting a cfg to zero will force it to use the negative prompt which in this case won't work so that is why cfg is set to 1
>>101683817give me a second, i'm testing it on this example to see if its working as intending >>101682871
>>101683817ok its not working, damn it, its the style but not the right character, maybe i need to tweak some things. but the pic is the basic connections i've done, which I think would be the correct way if I was using an SDXL model, perhaps i need to flip them around or use a different method like average or combine instead of concat.
>>101683671>>101683605Nice, those look professional though, fake skin. Best I've gotten so far is >>101683488 everything else looks a bit unrealistic I guess, some higher failure rate but I'm sure it's due to censorship E.G. a simple LoRA can fix this,
Are there any loras or models that allow finer control over facial features, or is that new flux thing better at it? I want to create characters but SD and XL just do sameface
>>101683953prompt facial features, try a different model, prompt some obscure famours person and use that as your personal sameface..and no for FLUX there is nothing like loras or anything yet .. sameface there? hm remains to be seen once anon wrangled the model enough, but I guess not as much as SDXL finetunes
Is it possible to combine flux with tensorRT for performance gains?It works on 12 GB, but yeah it's pretty fucking slow, uses like 8 GB of system ram at fp8.
dual wielding
>>101683983Maybe it's just the models I've tried but none of them offered much control about facial features and didn't seem to know the terms required. Fiddling around with celebrity faces does a bit more but it's still just a crutch at best
>>101681601like sdxl was?
>>101683098>>101682752how do you get it to write something?
>>101684057nta but you need to explicitly state that there's a speech bubble>a speech bubble above her head says "pixart was here"
>>101683817ok the ConditioningAverage node is definitely working at 0.5 strength. doing it at 0.3 strength now, will post that image on catbox once its done so you can see the difference, the concat node does nothing to the image, no change when removed.
>>101684057this >>101684091or simply>saying "Some text."or>caption "THIS AND THAT"
pic related is base imagehttps://files.catbox.moe/ae4ovc.pnglink is 0.3 strength for averaging
>>101684057I prompt it like>neon sign with text ">speech bubble with text ">title with text ">etc
I thought CFG is supposed to be set to 1, should I keep it at 3.5 instead?
Just cancelled my ideogram sub. Flux is so good I don't need cloud shit anymore.
>>101684402no, cfg is pointless, keep it on 1
>>101681353>local image gen that is literally better than the best closed source corporate modelsLeftist journalists are going to crush Flux once this gets out. This will be over before it even begins.
This thing does not like to denoise until 0.95 when routing to a second ksampler with an entirely different prompt. Why would that be? This is odd.
>>1016818403xA100 server by the minimum just for a fine-tune A Lora requires a single A100
Top is Flux Pro on Replicate, bottom is DALL-E 3 on Microsoft Designer, this is the prompt:>Detailed, realistic oil painting, 1990s science fiction illustration, waist up view of a female science fiction adventurer wearing futuristic clothes, looking at the camera, posing in front of a hovering sleek futuristic concept hovercar, on a street with futuristic pedestrians in a science fiction cityscape with sleek futuristic architectureDALL-E still has more of the styles I want from non-photorealistic images. I think Flux was intentionally trained not to replicate traditional art drawn by humans: it's either photorealistic or it has a very smooth digital art style.
>>101684478Why? Artist names and styles don't work well at all and it doesn't know celebrities.
>>101684679nta but it absolutely does know at least the few big political ones. agree on styles though.
>>101684665>I think Flux was intentionally trained not to replicate traditional art drawn by humans:I think blackforest learnt alot from SAI scandals and problems.. so they are leaving the dirty work for the finetunes. Smart move.
>>101684708meanwhile midjourney still making hundreds of millions per year off artist tears
>>101684723yea spent well on lawyers I guess
>tfw 4080>tfw automatically converted to low vram modewew lad, flux is a bit mental
>>101684767praise NVidia for being greedy on VRAM!if the 5090 really only has 28GB I gonna nerdrage
>>101684679Being open means it will eventually know much more (granted it is easy to train). Instead of pony types what we focus on first is artist finetunes, both Japanese and Western ones, also art/photo movements and photography. Ideally such tunes do not decide or converge on a specific style since their purpose is to teach. Then we do the pony.
>>101681583Poor furries scamming rich furries
>>101684767same on 4090 though
>>101684478There is nothing "journalists" and politicians can do, the weights are out, distributed over tens of thousands of machines all around the world. Even if they can drum up a big enough moral panic for HF and other mainstream providers to ban it, it's gonna continue underground.
>>101684783>>101684812model is fucking sick though. After months of stagnation and SAI cuckery we finally have something that mogs DallE. Just need to make it run on consumer hardware though
>>101684478this >>101684818also.. Apache 2.0 .. its joever
>>101684818Can't steal artist styles out of the box, the only argument they have left is cunny, but it's not a strong one.
>>101684834>consumer hardware thoughwont happen now, this will be reverse, the hardware you need to run it will be consumer in a year or three from now
its good, but not 12b good. i shall remain waiting for bigma.
>>101684854all it needs is a severely autistic person to optimise with wizardry it and we're golden.
>>101684834which shows how shitty SAI is, still, the better is always to come, this model is bad as a foundational model, it barely knows any art, its heavily tuned on "aesthetic" crap because its intended to be a consumer product for their generation services
>>101684880wah wah wah that's what fine tunes and loras are for.
>>101684878yea will probably happen, and for 1024x1024 FLUX.schnell meme production that the masses crave you will probably see it sooner rather than later
>>101684880>it barely knows any artIpdapter exists though. It technically knows all, so as a tool this is the best model, and I'm sure someone will figure out something for artists that is more efficient than IPAdapter or LoRAs, I mean especially with this model it's become a problem for this and every DiT model going forward.
>>101684919there is no implementation because T5 conditioning has no visual understanding of anything it's trained on. clip is the only way to do it proper but clip l is very small
https://huggingface.co/camenduru/FLUX.1-dev/tree/mainfp8 of flux dev
>>101684892>that's what fine tunes and loras are for.Why even wait for companies to release a model then if we can just train one? lol
>>101684996Go ahead and make a 12B model from scratch
>>101684991What does this mean
>>101685007Well, your answer to "this model lacks the knowledge of basic public domain art" is "thats was fine tunes are for" then a model missing everything also just needs to be fine-tuned
>>101685026>nogenSeriously, move on. There's plenty of models that know your esoteric art styles.
>>101685020lower precision and smaller model size, therefore more gpus can run it. we have a fp8 loader node so it doesnt really matter
>>101685042>>nogenI accept you concession, thank you very much
>>101685055You're here in bad faith, you'll always find something to bitch about because you're negatively biased and want models to fail.
>>101685066I can see it in 6 months>The urethras on my cockgens are a couple of mm off what I expect, dogshit model
>>101685094
>>101685042>>101685055>nogenI need more time!
cozy thread today
>>101685053It doesn't work.
Here, a image so the niggers can understand it visually>>101685133cute
>>101685219>retarded posterThe best part of seeing gens is you can tell when someone is autistic and retarded. It's like if I saw you IRL and seeing you're an ungroomed 400 pound man I can disregard everything you have to say.
>>101685219>not foundationalcan't say for certain just yet, we'll just have to wait and see what happens. fingers crossed it's not another sdxl.
>>101685133>>101685243After all that time I can't even post it because nipples
>>101685112
>>101685266I can say. If it's easily tuneable it's the new SD, 100%. Reddit is shitting and cumming over it.
Which anime was this from again?
>>101685243cope lmao>>101685266>can't say for certain just yetIts incapable of making heavily abstract or classical paintings because its turned to hell on aesthetic images, yeah if you train anime or porn on it will work, but thats very far from my point, though this is the only thing retards care about so its not like it makes sense to talk about anything else it seems>>101685243here, a (you) to make your day a little more brighter
>>101685305>wah it doesn't know my fetish art, here's an image with my lack of taste and creativityThanks sweaty 400 pound man for your input
>>101685314>impasto, piet mondrian, or any fucking painting style that isnt hyper ultra realistic slop>my fetish artok then, you win from mental retardation
>>101685314nta, but atleast it knows my fetish art
>>101685305
>>101685288
>>101684880try dropping guidance on -dev to 1.0-1.5 if you're having issues with style/prompt adherence
>>101685351>it doesn't know fine art styles but here's my generic anime girl slop showing I don't try or care I just bitch because really my motive is I want the model to fail
>>101684185Not sure why but the conditioning node doesn't want to play nice but using the "split between clip_l and t5" method works really well. But I am also running it on schnell with 6 steps for quick gen purposes.
>>101685367now try making a heavily abstracted person/portrait like 1.5 and SDXL can do
:eye:
I keep getting signatures that look like real signatures. I wonder if they are.>>101685405try licking my ass
>>101685405
>>101685400Yeah its not working right, completely loses the character. using another Ksampler for second pass different prompt does not work either. I might try IPadapter to transfer style if it even works, after some rest.
Morning bread ready to eat...>>101685374>>101685374>>101685374
>>101685398I already tried, it can't make impastos or classical art, if I hadn't tried I wouldn't have said it, its extremelly easy to achieve a painterly, textured with heavy brushstrokes image on 1.5/XL and a uphill battle on FLUX>>101685421Thats not a person, also doesn't work with animals, or any of those popular subjects, its always heavily shaded and defined
Yea! Beksinki signed this.
>>1016818914 GB VRAM, 16 GB RAM
>>101685266>can't say for certain just yet, we'll just have to wait and see what happens. fingers crossed it's not another sdxl.this model lacks a lot of styles, artists and characters, fortunately that can be salvaged with more training, they did the hard part, good anatomy, good prompt understanding, good textures, this model has so much potential
