SD1.5/SDXL was relatively terrible at it's start versus what we have now. That took time and effort of people, but it was worth it. I don't think this will kill flux, but it could be a good alternative depending on your use case. Personally I hope both continue to exist and develop, just specializing in different things and I hope the community effort gets divided to both instead of one model "winning".
Perfect said!!! 👏 Also, if we have competition, I believe there should be less chance of companies making overly compromising decisions, as Stability AI did when Stable Diffusion 3.0 released. That was really sad.
Exactly. The fact that the base model can already go toe to toe with the shiny new toy (flux) and unlike it, can actually be fine tuned, makes it a huge improvement.
damn ! true i though FLUX had that. that explain why it's not working yet ! good comment ! SD3.5 is looking more interesting by the day (day n°2 :p). training and giving customized style always take me back to SDXL rather than FLUX. i hope for the 2.5B to replace SDXL and i want theses fine anime model from the SDXL era back in this next gen sauce ! we will have to wait for 2.5B
Thanks for making this video, especially the part where you talked about what you have in the pipeline. I am one of your patreon supporters and I had asked a question about Lora training and I am glad to see that there is a video coming up soon. I think that the most important things to me in AI image generation are image generation of anatomically correct characters (D&D themed) in 1080x1920 resolution. For that reason I am going to stay with Flux 0.1 for at least a little while, unless Stability improves its support for the things that matter most to me. Also even if I use Stability AI for something like...landscapes...or car images....I will still use Flux for portraits in the immediate future (at least 8 months or so). Thanks again.
Flux was good, and the simpler prompting style that also had the model follow it better was amazing, but it was very restricted in style. This will most likely become the new standard in open source after a week or so with new trained models and lora. And the fact that it is smaller is also a major point over Flux. edit: Btw, why are there no open source music generation models out yet? This is the only big thing that is missing. Udio and Suno are good, but they are extremely limited in capabilities that will definitely come with open source models.
TRUE ! i can't phrase it beter. i can extend saying that researcher won't put as much effort training a broken distilled weight even if it give beter output. the full extent of flux is proprietary. dev is non commercial, finetunner hates that but user are just enjoying the out of the box experience. not realizing on what can be missed on the long run
Yes I'd definitely want an open source music generation model that can compete with Udio and Suno level. I am very uninterested in how there is a new free video gen every month (most of which can't be run on consumer hardware), but music generation, which seems to be way less computally intensive is completely missing from the open source models.
It would seem the one two blows. That were the punch in the face that was the backlash to SD3 2B. And the kick in the balls of their former engineers releasing Flux. Got Stability back on track. This is good. It means that there is still competition. And that if the boys at Black Forest Labs want to keep the crown that they took from their former bosses. They can't afford to rest on their laurels.
I think you just need to add or something similar and u shouldnt get balls or engineers in it. sounds like a classic 'put it in the negative ! wrong hole haha ' situation :)
Let's be honest, the capability of NFSW is one of these biggest/main reasons to pick SD3.5 over Flux. I want to say a majority of users are doing it for NSFW content
I love their ham so I'm praying for a holiday desert thing but its in a cool tin that you can design on their website (and they throw in some ham cuz they empathetic german angels)
Your point about SD 3.5's fine-tuning flexibility compared to Flux really is key. It’s refreshing to see an open-source model like this with so much potential for community-driven improvement. As impressive as Flux is, the limitations on customization do make SD 3.5 intriguing, especially for users with specific creative needs. Hopefully, Stability AI keeps up this momentum because the prospect of a truly customizable alternative in the text-to-image space is exciting.
like tuning a violin, bow or harpsichord, rest assured your hard work will generate some super cute red head elf queens (different kingdoms) .. once comfy stops going nuts in the dos window or whatever
I don't want a Flux killer, I just want another useful tool and it does fit that. Keep them trying to outdo each other and we all benefit. I still need to run Flux locally. I have a 24 GB graphics card and and ComfyUI installed, so I just need to take the last step.
@@freakguitargodI mean ones a base model the other is iirc a fine tuned distilled model so it pretty much answers ur question base model isn’t meant to be as good
I hope you make a video of installing PuLID for Flux, which allows you to upload an image of a person, then create images of this same person, without the need to train a Lora.. Then add it to the Ultimate Flux Workflow.
Did not agree with your conclusion, what was good with the 1.5 era was the comunity driven works. and this new range of models and the medium range is in my mind even more interesting for that. I hope to see again a lot of people able to use on their computer on cool model and train it. the Support and tools whe had on 1.5 was mind bogling, and if we have a good finetuning with the prompt adherence we will be golden. because good prompt is out of reach for most people due to the size of the models. these ones have a Bright future. it can hit the niche of the 1.5 models occupied back then.
Any people have any experience with Forge and Comfy UI. I use forge, I just wondered if Comfy UI is better, worth changing to, has any advantages, or wether forge is good enough?
I just tested it - is it heavily censored, we'll have to wait for finetuned ones that are uncensored. Also your manual instructions are missing steps where you install ComfyUI Manager, then install missing custom nodes, then reloading the workflow.
fair take ! i'll be waiting myself for an UNION controlnet and an IPadapter (some work from SD3.0 can port over but never really took off for obvious reasons)
I'm still 100% behind SDXL, it's under a true open license, which doesn't require you to notify stability ai, or pay after 1 million. The latter isn't likely an issue I would run into, but having to notify stability ai on what you're doing with the model is annoying and just makes it less open. Also, SDXL is pretty damn good these days with the amazing fine-tunes that have been done by the community.
Sorry for putting it here. But i want to start the adventure with AI. I saw some tutorials or different tools but I want to use something I can use on my hardware. I would really ask for directions what should I start first and should work? I have only 3080 10GB, Ryzen 5900x, 32 GB RAM. What should I try to run? Of course I would like best one but i know I am restricted by my hardware. But not sure where to start. As far I see 3.5 should work (?), but is it possible to use Large or Large Turbo or it will not work? And how it will compare with FLUX versions (I think there are versions with "8" for a versions with lower memory)? Or any other recommendations?
I would suggest installing Stability Matrix first, to use it as an installer for other packages with shared models folder to save on disk space. Try Stable Diffusion Forge first, as it offers maximal simplicity and memory/speed optimisations. Choose the *nf4 Flux models, SD 1.5-3 models should also fit in 10 GB VRAM freely - explore CivitAI for that, sort the models catalog by type and popularity.
Given how this is an actual base model and not a distilled, already fine tuned one like Flux, there’s no comparison. This blows flux out of the water. People who don’t know shit about the technical aspects and just blindly type in spaghetti prompts into whatever the latest model is won’t be able to tell the difference, and will probably prefer flux simply because it’s more refined. But long term, what makes or breaks a model is how well it can be optimized and fine tuned. Flux, at least in its current version, will never get there simply because of how it’s designed as a distilled model. This is the actual SD3 release we should have gotten, and now that it’s finally here, there’s nowhere to go but up. Unless flux releases an actual base model that can be fine tuned, in 6 months no one will care about it anymore and every finetuner will move onto SD3
Also the fact that flux currently doesn’t support negative prompts, and many common configuration tools like CFG are not available either, makes it a very weak model for any actual useful workflow that isn’t just creating the most generic looking images. As mentioned in the video, this is a big disadvantage for flux. They really need to release a proper base version that is up to par with at least SDXL in terms of features. Otherwise, it’s just an interesting tech demo.
Totally agree, and honestly it blows my mind that people like Flux. It's impressive and all but once you start tinkering around with it, its so rigid and inflexible. And honestly I hate the faces it produces, all the women have that same high bone structure over trained face with cleft chins, and its very poor in terms of diversity (everyone is white and blonde unless you tell it otherwise). I went from being excited to going straight back to SDXL, which is still prior to 3.5 imo the best model the open source community ever got.
I want a prompt adherence and proportions model. One that isn't great with quality, but does prompt adherence and proportions perfectly. Then, I want that to go into a quality model, which will add realism to the proportions model output. Then I want that fed into an Upscale model. A model set for each different thing. So Car Proportions model, Anime quality model, and an Anime Upscale model. Landscape Proportions Model, a Painting Quality Model, and a Painting Upscale model. 8b parameters focused purely on human proportions, would blow anything else away. I can easily change models if I want some other specialised model. It's the "Jack of all trades, master of none" philosophy. Time to make specialised models that are lit.
3.5l actually does better than flux..... at 1 mega pixel. It can't do anything BUT 1 megapixel so it has far more training than flux when it comes to 1mp gens by a longshot since flux is spread throughout so many resolutions of training... But as a result in flux you have to change resolutions a lot, sometimes 1.3mp will work while 2mp is blurry or vice versa for instance.... So sd3.5l is a lot better for its specific use cases of 1mp gens... When you get worse quality gens it's likely a prompting issue, i learned 3.5l has so much different of a prompting style than Flux, when you learn to prompt it right and use unique prompts for it the results are often much better than Flux imo.
What is a weak GPU? I got a RX 7800XT, and flux is very hard on it, especially if I run controlnet. It max out the 16gb in no time. Maby there was a pytourch a few days ago? Yesterday I could not run it at all?
I don't know if I am alone but for image quality is no longer what I am looking for but rather prompt adherence and to me it looks like Flux and SD3.5 is almost the same in that regard.
i think we're at a plateau that is going to be overshadowed by the big video wave that hits open source after the corporate folks tell us their spells. I agree and the big thing that I noticed with Flux was handling text so well. Otherwise, I'll forget it's SDXL or FLUX on the workflow and I won't notice either way. It's been little changes but I love all of the control net stuff - keeps me busy even tho we're having to add lots of our own ingredients to make our gumbo presentable lol
2:18 What I would say is: Never test your text to image on a hot women facing the camera in a portrait mode since that it probably the single most common image on the internet and the easiest image for an AI to create based on its training.
If you want hands I'll give you hands DPO ;) Merging DPO into the model makes it harder to train and lowers variety of output (making it too similar to flux), so we decided not to do it on release.
I don’t know why they released this, it would be cool if at one moment the controlnet for a1111 was updated with support for the model, I’m staying on Forge with FLUX, the ssd is not rubber to produce models, as for the content, all this has already been shown many times today
Glad to see SD back in the ring. Shame it's so bad at human details. They need to step up and do a bigger model. The 5090 will have 32gb and that pushes the prices of everything down as well.
3.0 wsa like when ea added loot boxes to that star wars game or a decision sooooo awful that they just kinda pretend it's not there. then black forest curb stomps it and u have to outsource cause scaling gonna scale and the black boxes need nuturing - but 3.5 made it out within that retaliation windows so it's fun but slow down guys. they need to meet up and pop adderall in some mansion and boom 'anything to video' model, 60fps, 40 second clips that are quickly changed to 90 by some wizard in an add on, and it leans heavily towards nudity because they realized all that really matters after 2 months is if it can handle a complex nude with text that says "JASON IS BAD AT APEX LEGENDS HAHA" and maybe it has this guy 'Jason's mom's face on it.. crazy I know but some enjoy imagining another's mother under hte covers. As long as Jason yells into the mic I think the naysayers will be satiated.
we know so many cool words now. i dont want to date a model, i download or train them myself. gguf is fast but comfy needs a node wrapper buffalo yolo BLIP CLIPs with Lora Karros
I'm not interested by SD3.5 for now. Open AI lost completely my trust with the scandalous SD3 licence thing. Maybe later when there will be a lot of models/loras on it. And, to be honest, I have no trouble to make good illustrations with Flux. ^^ I will try a little Verus Vision, a new model based on Flux ( it was long but it is possible)
When loading the workflow i get an error saying Missing Node Types. I watched another one of your videos where there was a manage button but this version does not have one. Anyone able to advise?
It seems to me like this model was only dropped only because FLUX is the talk of the town at the moment. You know what I mean? It seem rushed out the kitchen. Nothing you're missing out on here. And we can already do NSFW with custom 1.5 or XL checkpoints.
Love Flux, but if you have less than 16Gb of vram, it's really hard to run without making compromises that effect quality. Hopefully SD 3.5 will be better with that.
@@Airbender131090 I saw that too. But at only 2B parameters, I'm not sure it will have the image quality to compete with the 8B version or Flux. Might be a small improvement over SDXL though, which would be nice.
Flux isn't going anywhere. SD3.5 certainly has it's plusses, like the license and art styles, but Flux is better in so many other ways. I still find Flux generates slightly faster (with no help) but the model loading is so much slower with SD3.5. Whether SD 3.5 succceeds or not will depend greatly on how much the community takes it on board.
Yep, entirely based on the community, flux has a head start with dedistilled attempts etc, though the factors that might aid SD3.5 large is, it is a smaller model, so lower requirements, it has a more open license than flux1-dev(though flux1-schnell is similar), and it isn't distilled, it is a proper base model. I am going to play around with SD3.5 large for a bit, but at this point I might be going back to flux, at least until the community jumps in.
I REALLY want to be optimistic on this but it seems pretty terrible at anatomy. Forget about fingers, you sometimes get wrong number of limbs. Hand interactions with tools terrible most of the time. I really hope I am wrong on this but at this scale this is going to be very hard to fix with finetunes. Other than this, it's really a great model. More creative too.
Not watched yet, going to in a bit. But having tested flux and sd3.5 with the same prompts, flux has way better coherence and handles more surreal things with a lot better ease. If your after a test try prompting for a blue haired squirrel in a gothic-scifi armour. flux does this... with various degrees of success. sd3.5 fails on gothic-scifi armour (it has a style it thinks is that since it fails consistently on that part but is not even close) and often fails on the squirrel part..... it knows a squirrel is usually grey so thats what it does. (most of the time) Admitedly my comment is going on the title as i cant watch the video for a bit. But sd3.5 is far from a flux killer.
@@lefourbe5596 I planed to redo my sdxl lora's for flux but never quite got the time. I just dont feel sd3.5 is following prompts as well as flux does (atelast not when using natural language, there may be a new prompting format we need to learn to get it to comply) At the minute i prefer the results of flux and hate the fact sd3.5 cant even do gothic-scifi power armour!... but its still early days.
@@DaveTheAIMad from the lykon discord. I made Harris try style. Placed at the end the influence is non existent. At the front it's much more noticable. Will try for myself soon. The yoji shinkawa style was convincing to me
Flux "uncensored" refuses to draw a dinosaur eating people, and so many other things, so I take those claims with a grain of salt ... i'll check that 3.5 though. I am using SDXL much more now
You say 8B parameters is "very very" close to 12B parameters as if it wasn't just 66% of the parameters. Flux is like 10x better as it stands right now.
pretty much yes. A1111 times are over. let it to rest... and it's for the best : i advise Krita AI diffusion nowadays. a very easy to use interface for image generation inside an excellent image editing software. it run on comfyUI backend and as such, there is no limit to what you can do inside it without dealing with a single node !!!
yes FLUX is beter as it is. but you cannot train it efficiently, it is not as open and business friendly. so to me as a Lora maker maniac i will jump on SD3.5 as soon as Ipadapter and controlnet gets decent. coz i know that this model will become AT LEAST beter than FLUX !
@@yaroaram7237 , Ppl were saying the same about SD3 😆 and now 3.5 is out. Kinda doubt it'll be smth revolutionary. I'm still very satisfied with what FLUX Does atm. If someone can't get any desired result, I think it's more of a prompting issue, not the FLUX.
HELLO HUMANS! Thank you for watching & do NOT forget to LIKE and SUBSCRIBE For More Ai Updates. Thx
Hi 👋
use stability matrix for automatic installer
Just FYI, 16:9 would be 1280 x 720, not 768. ;)
Glad to see Stable Diffusion back in the game. Competition is always a good thing.
SD1.5/SDXL was relatively terrible at it's start versus what we have now. That took time and effort of people, but it was worth it. I don't think this will kill flux, but it could be a good alternative depending on your use case. Personally I hope both continue to exist and develop, just specializing in different things and I hope the community effort gets divided to both instead of one model "winning".
Yes, they need to go different ways and start reaching agreements, teaching the same lore in a circle is not a great pleasure
Perfect said!!! 👏
Also, if we have competition, I believe there should be less chance of companies making overly compromising decisions, as Stability AI did when Stable Diffusion 3.0 released. That was really sad.
Exactly. The fact that the base model can already go toe to toe with the shiny new toy (flux) and unlike it, can actually be fine tuned, makes it a huge improvement.
There is another big plus for SD3.5: It uses cross attention like the old ones, which allow for IP-Adapter, InstantID etc.
damn ! true i though FLUX had that. that explain why it's not working yet !
good comment !
SD3.5 is looking more interesting by the day (day n°2 :p). training and giving customized style always take me back to SDXL rather than FLUX.
i hope for the 2.5B to replace SDXL and i want theses fine anime model from the SDXL era back in this next gen sauce ! we will have to wait for 2.5B
@@lefourbe5596 The possibilitys are there, now its up to the community 😁
Yes, we need SD 3.5 fine tuning training video, LETS GO...
if only we could figure this out somehow... either making the video or actually doing the 3.5 thing. I say we get the host to do it!!!
Thanks for making this video, especially the part where you talked about what you have in the pipeline. I am one of your patreon supporters and I had asked a question about Lora training and I am glad to see that there is a video coming up soon. I think that the most important things to me in AI image generation are image generation of anatomically correct characters (D&D themed) in 1080x1920 resolution. For that reason I am going to stay with Flux 0.1 for at least a little while, unless Stability improves its support for the things that matter most to me. Also even if I use Stability AI for something like...landscapes...or car images....I will still use Flux for portraits in the immediate future (at least 8 months or so).
Thanks again.
Flux was good, and the simpler prompting style that also had the model follow it better was amazing, but it was very restricted in style.
This will most likely become the new standard in open source after a week or so with new trained models and lora. And the fact that it is smaller is also a major point over Flux.
edit: Btw, why are there no open source music generation models out yet? This is the only big thing that is missing. Udio and Suno are good, but they are extremely limited in capabilities that will definitely come with open source models.
TRUE ! i can't phrase it beter.
i can extend saying that researcher won't put as much effort training a broken distilled weight even if it give beter output.
the full extent of flux is proprietary. dev is non commercial, finetunner hates that but user are just enjoying the out of the box experience. not realizing on what can be missed on the long run
Yes I'd definitely want an open source music generation model that can compete with Udio and Suno level. I am very uninterested in how there is a new free video gen every month (most of which can't be run on consumer hardware), but music generation, which seems to be way less computally intensive is completely missing from the open source models.
Very interesting and informative.
Thank you!
I'd like to see some easy to follow vids on training 3.5. Using Flux for training would be cool. Combine them to make the best model possible.
It would seem the one two blows.
That were the punch in the face that was the backlash to SD3 2B.
And the kick in the balls of their former engineers releasing Flux.
Got Stability back on track. This is good. It means that there is still competition.
And that if the boys at Black Forest Labs want to keep the crown that they took from their former bosses.
They can't afford to rest on their laurels.
we love to see it ! well said 😄
tried it, and its not so great...needs months of user fine-tunes before its actually even EVEN with flux.
@@omegablast2002 illustriousXL training set over SD3.5 gets me more excited on SD3.5 than FLUX will ever do.
I think you just need to add or something similar and u shouldnt get balls or engineers in it. sounds like a classic 'put it in the negative ! wrong hole haha ' situation :)
Let's be honest, the capability of NFSW is one of these biggest/main reasons to pick SD3.5 over Flux. I want to say a majority of users are doing it for NSFW content
The strong point of Flux is it's amazing adherence/understanding to the prompt. Stablediffusion never excelled at that. Is the version 3.5 different?
He said it was much better at the start of the video yes
I just want whatever black forest labs is working on next tbh.
it's a video AI for now
I want the best of both. Competition is good.
I love their ham so I'm praying for a holiday desert thing but its in a cool tin that you can design on their website (and they throw in some ham cuz they empathetic german angels)
Simple, pump out a whole bunch of stuff from flux to train 3.5, win win
Your point about SD 3.5's fine-tuning flexibility compared to Flux really is key. It’s refreshing to see an open-source model like this with so much potential for community-driven improvement. As impressive as Flux is, the limitations on customization do make SD 3.5 intriguing, especially for users with specific creative needs. Hopefully, Stability AI keeps up this momentum because the prospect of a truly customizable alternative in the text-to-image space is exciting.
hi, you can actually finetune flux but it's really easy to break at the same time. There is also a de-destilled model some people made.
like tuning a violin, bow or harpsichord, rest assured your hard work will generate some super cute red head elf queens (different kingdoms) .. once comfy stops going nuts in the dos window or whatever
It's not even close to flux in my opinion.
I don't want a Flux killer, I just want another useful tool and it does fit that. Keep them trying to outdo each other and we all benefit.
I still need to run Flux locally. I have a 24 GB graphics card and and ComfyUI installed, so I just need to take the last step.
Yeah, flux is still better
defintely it works well even n fp8 and stuff!
@@tails_the_god yeah, Flux fp8
@@freakguitargodI mean ones a base model the other is iirc a fine tuned distilled model so it pretty much answers ur question base model isn’t meant to be as good
Love your channel!
-written by my AI assistant
SD is back. training, if easy enough, will kill it.
I hope you make a video of installing PuLID for Flux, which allows you to upload an image of a person, then create images of this same person, without the need to train a Lora.. Then add it to the Ultimate Flux Workflow.
In SD 3.5 we trust. Thanks for the vid :-)
The low res and anatomy is a bit of a downer though.
I'll trust it when I see a well placed cornhole with upscaling. i've seen them duplicate
already tested it, so cool
Interesting would be a tts with voice cloning in it.
Whats peoples big obsession about the voice ting? I don't get it...
Thanks for another great video. I would like to see lora training for flux.
YES, we need SD 3.5 fine tuning training video. LETS GO...😄
Competition is always good for the consumer
Did not agree with your conclusion, what was good with the 1.5 era was the comunity driven works. and this new range of models and the medium range is in my mind even more interesting for that. I hope to see again a lot of people able to use on their computer on cool model and train it. the Support and tools whe had on 1.5 was mind bogling, and if we have a good finetuning with the prompt adherence we will be golden. because good prompt is out of reach for most people due to the size of the models. these ones have a Bright future. it can hit the niche of the 1.5 models occupied back then.
Any people have any experience with Forge and Comfy UI.
I use forge, I just wondered if Comfy UI is better, worth changing to, has any advantages, or wether forge is good enough?
the 1mp limit is a hard bummer...
and a bit more info about the text encoder would also be nice. didn't have time yet to dive into the topic myself.
best comparision❤
Pony is the best!
I just tested it - is it heavily censored, we'll have to wait for finetuned ones that are uncensored. Also your manual instructions are missing steps where you install ComfyUI Manager, then install missing custom nodes, then reloading the workflow.
Literally. I spent the last 2 hours, trying to get it to work and I still have no idea what the fuck is the manager
Tried it, I'm sticking with Flux until we get some fine tuned sd3.5
fair take ! i'll be waiting myself for an UNION controlnet and an IPadapter (some work from SD3.0 can port over but never really took off for obvious reasons)
I'm still 100% behind SDXL, it's under a true open license, which doesn't require you to notify stability ai, or pay after 1 million. The latter isn't likely an issue I would run into, but having to notify stability ai on what you're doing with the model is annoying and just makes it less open. Also, SDXL is pretty damn good these days with the amazing fine-tunes that have been done by the community.
what was the website used at the 4:09 mark? great video as usual
Sorry for putting it here. But i want to start the adventure with AI. I saw some tutorials or different tools but I want to use something I can use on my hardware. I would really ask for directions what should I start first and should work?
I have only 3080 10GB, Ryzen 5900x, 32 GB RAM.
What should I try to run? Of course I would like best one but i know I am restricted by my hardware. But not sure where to start. As far I see 3.5 should work (?), but is it possible to use Large or Large Turbo or it will not work? And how it will compare with FLUX versions (I think there are versions with "8" for a versions with lower memory)? Or any other recommendations?
I would suggest installing Stability Matrix first, to use it as an installer for other packages with shared models folder to save on disk space. Try Stable Diffusion Forge first, as it offers maximal simplicity and memory/speed optimisations. Choose the *nf4 Flux models, SD 1.5-3 models should also fit in 10 GB VRAM freely - explore CivitAI for that, sort the models catalog by type and popularity.
will this work with Forge?
Damn! Stable diffusion be like: Jhuk ke rehna hoga mere aage 👇😎
Came here to say that!
@@wakegary Are you a ghost? Or just a creepy dark web user?..........your channel 🤨🤨
nice man
also question how did you animate your avatar it was a 2d image you rigged some how what program did you use?
Given how this is an actual base model and not a distilled, already fine tuned one like Flux, there’s no comparison. This blows flux out of the water. People who don’t know shit about the technical aspects and just blindly type in spaghetti prompts into whatever the latest model is won’t be able to tell the difference, and will probably prefer flux simply because it’s more refined. But long term, what makes or breaks a model is how well it can be optimized and fine tuned. Flux, at least in its current version, will never get there simply because of how it’s designed as a distilled model. This is the actual SD3 release we should have gotten, and now that it’s finally here, there’s nowhere to go but up.
Unless flux releases an actual base model that can be fine tuned, in 6 months no one will care about it anymore and every finetuner will move onto SD3
Also the fact that flux currently doesn’t support negative prompts, and many common configuration tools like CFG are not available either, makes it a very weak model for any actual useful workflow that isn’t just creating the most generic looking images. As mentioned in the video, this is a big disadvantage for flux.
They really need to release a proper base version that is up to par with at least SDXL in terms of features. Otherwise, it’s just an interesting tech demo.
Totally agree, and honestly it blows my mind that people like Flux. It's impressive and all but once you start tinkering around with it, its so rigid and inflexible. And honestly I hate the faces it produces, all the women have that same high bone structure over trained face with cleft chins, and its very poor in terms of diversity (everyone is white and blonde unless you tell it otherwise). I went from being excited to going straight back to SDXL, which is still prior to 3.5 imo the best model the open source community ever got.
If I would the channel owner, I would for sure pin this! Really bro, thanks for mention the point!
I would like to to see the flux Q4 trained with fluxgym and which Lora to use with the Q4 optimized flux
Hmm, I'm seeing several fine-tuned Flux models on Civitai.
I want a prompt adherence and proportions model. One that isn't great with quality, but does prompt adherence and proportions perfectly. Then, I want that to go into a quality model, which will add realism to the proportions model output. Then I want that fed into an Upscale model. A model set for each different thing. So Car Proportions model, Anime quality model, and an Anime Upscale model. Landscape Proportions Model, a Painting Quality Model, and a Painting Upscale model. 8b parameters focused purely on human proportions, would blow anything else away. I can easily change models if I want some other specialised model. It's the "Jack of all trades, master of none" philosophy. Time to make specialised models that are lit.
3.5l actually does better than flux..... at 1 mega pixel. It can't do anything BUT 1 megapixel so it has far more training than flux when it comes to 1mp gens by a longshot since flux is spread throughout so many resolutions of training... But as a result in flux you have to change resolutions a lot, sometimes 1.3mp will work while 2mp is blurry or vice versa for instance.... So sd3.5l is a lot better for its specific use cases of 1mp gens... When you get worse quality gens it's likely a prompting issue, i learned 3.5l has so much different of a prompting style than Flux, when you learn to prompt it right and use unique prompts for it the results are often much better than Flux imo.
What is a weak GPU? I got a RX 7800XT, and flux is very hard on it, especially if I run controlnet. It max out the 16gb in no time. Maby there was a pytourch a few days ago? Yesterday I could not run it at all?
SD4 or 5 will be flux1 successor .
unlikely, they didn't planned a next txt2img model after 8B. FLUX 2 would eventually but they focus on video ai right now.
@@lefourbe5596 that require large compute power 'H.W' for PC users !
More styles are possible with Flux when you go for negativ CFG scale...like -3.
it's not a bug it's a feature...
why is this a thing ??? but nice to know ... further renforcing my doubts about FLUX future
@@lefourbe5596 Dont know the technical stuff deeply, but from what I got its just the way to tell it "fuck your conditioning" 🤷♂
Looking forward to you making Lora training guide for Sana from Nvidia
Is there a link to the workflows you used in the video?
I don't know if I am alone but for image quality is no longer what I am looking for but rather prompt adherence and to me it looks like Flux and SD3.5 is almost the same in that regard.
i think we're at a plateau that is going to be overshadowed by the big video wave that hits open source after the corporate folks tell us their spells. I agree and the big thing that I noticed with Flux was handling text so well. Otherwise, I'll forget it's SDXL or FLUX on the workflow and I won't notice either way. It's been little changes but I love all of the control net stuff - keeps me busy even tho we're having to add lots of our own ingredients to make our gumbo presentable lol
2:18 What I would say is: Never test your text to image on a hot women facing the camera in a portrait mode
since that it probably the single most common image on the internet and the easiest image for an AI to create based on its training.
I am very curious if auraflow takes off after pony_v7 release
If you want hands I'll give you hands DPO ;)
Merging DPO into the model makes it harder to train and lowers variety of output (making it too similar to flux), so we decided not to do it on release.
jesus christ brother how can we find missing nodes for god sake manager cannot find anythin
I don’t know why they released this, it would be cool if at one moment the controlnet for a1111 was updated with support for the model, I’m staying on Forge with FLUX, the ssd is not rubber to produce models, as for the content, all this has already been shown many times today
When I’ve used your flux one click installer do I use the sd3.5 one click? Or will that install comfy twice?
Glad to see SD back in the ring. Shame it's so bad at human details. They need to step up and do a bigger model. The 5090 will have 32gb and that pushes the prices of everything down as well.
This is what 3.0 should have been, still, the fine tunes will be great!
3.0 wsa like when ea added loot boxes to that star wars game or a decision sooooo awful that they just kinda pretend it's not there. then black forest curb stomps it and u have to outsource cause scaling gonna scale and the black boxes need nuturing - but 3.5 made it out within that retaliation windows so it's fun but slow down guys. they need to meet up and pop adderall in some mansion and boom 'anything to video' model, 60fps, 40 second clips that are quickly changed to 90 by some wizard in an add on, and it leans heavily towards nudity because they realized all that really matters after 2 months is if it can handle a complex nude with text that says "JASON IS BAD AT APEX LEGENDS HAHA" and maybe it has this guy 'Jason's mom's face on it.. crazy I know but some enjoy imagining another's mother under hte covers. As long as Jason yells into the mic I think the naysayers will be satiated.
What if you generate max 1mp image with 3.5, being say 1280x720, then resample that up to 1920x1080 for instance?
When you think the hands are bad... try to let someone holding a knive 😂
Is there a forge version?
This has so much potential I hope the community makes great finetunnings, control nets and Ip Adapters
there was a gguf version released yesterday
we know so many cool words now. i dont want to date a model, i download or train them myself. gguf is fast but comfy needs a node wrapper buffalo yolo BLIP CLIPs with Lora Karros
The 1 megapixel issue is a deal breaker to me
does that mean you're not going to use it at all? what was the deal?
FLUX 1.1 Pro still have an edge bruh!!!! 🙂🙂🙂
so this is literly taking longer to load than flux for me, am i doing something wrong?
I'm not interested by SD3.5 for now. Open AI lost completely my trust with the scandalous SD3 licence thing.
Maybe later when there will be a lot of models/loras on it.
And, to be honest, I have no trouble to make good illustrations with Flux. ^^
I will try a little Verus Vision, a new model based on Flux ( it was long but it is possible)
When loading the workflow i get an error saying Missing Node Types. I watched another one of your videos where there was a manage button but this version does not have one. Anyone able to advise?
One more UNCENSORED title 😂
Really. Try to create some kinda uncensored thing?
It seems to me like this model was only dropped only because FLUX is the talk of the town at the moment. You know what I mean? It seem rushed out the kitchen.
Nothing you're missing out on here. And we can already do NSFW with custom 1.5 or XL checkpoints.
Discord link dosent work
What ui are you using for the image generation.
At the moment it is only available for Comfyui.
Does it have any clue on how to do human anatomy? Was it censored in its training?
Can it write text though?
yes it can !
Now with fewer legs?
is it uncensored tho?
Love Flux, but if you have less than 16Gb of vram, it's really hard to run without making compromises that effect quality. Hopefully SD 3.5 will be better with that.
2B sd 3.5 will come out october 28
@@Airbender131090 I saw that too. But at only 2B parameters, I'm not sure it will have the image quality to compete with the 8B version or Flux. Might be a small improvement over SDXL though, which would be nice.
@@Airbender131090 it's goes back to 2.5B just like SDXL.
flat out replacement if it deliver on promise
Flux isn't going anywhere. SD3.5 certainly has it's plusses, like the license and art styles, but Flux is better in so many other ways. I still find Flux generates slightly faster (with no help) but the model loading is so much slower with SD3.5.
Whether SD 3.5 succceeds or not will depend greatly on how much the community takes it on board.
Yep, entirely based on the community, flux has a head start with dedistilled attempts etc, though the factors that might aid SD3.5 large is, it is a smaller model, so lower requirements, it has a more open license than flux1-dev(though flux1-schnell is similar), and it isn't distilled, it is a proper base model. I am going to play around with SD3.5 large for a bit, but at this point I might be going back to flux, at least until the community jumps in.
Is it possible to Merge SD 3.5 x Flux x SDXL?
My curious mind wants to know 😂
I REALLY want to be optimistic on this but it seems pretty terrible at anatomy. Forget about fingers, you sometimes get wrong number of limbs. Hand interactions with tools terrible most of the time. I really hope I am wrong on this but at this scale this is going to be very hard to fix with finetunes.
Other than this, it's really a great model. More creative too.
Not watched yet, going to in a bit. But having tested flux and sd3.5 with the same prompts, flux has way better coherence and handles more surreal things with a lot better ease.
If your after a test try prompting for a blue haired squirrel in a gothic-scifi armour. flux does this... with various degrees of success. sd3.5 fails on gothic-scifi armour (it has a style it thinks is that since it fails consistently on that part but is not even close) and often fails on the squirrel part..... it knows a squirrel is usually grey so thats what it does. (most of the time)
Admitedly my comment is going on the title as i cant watch the video for a bit. But sd3.5 is far from a flux killer.
as a lora maker and occasionnal finetunner. i would much prefer SD3.5 as a SDXL replacement.
@@lefourbe5596 I planed to redo my sdxl lora's for flux but never quite got the time. I just dont feel sd3.5 is following prompts as well as flux does (atelast not when using natural language, there may be a new prompting format we need to learn to get it to comply)
At the minute i prefer the results of flux and hate the fact sd3.5 cant even do gothic-scifi power armour!... but its still early days.
@@DaveTheAIMad from the lykon discord. I made Harris try style.
Placed at the end the influence is non existent.
At the front it's much more noticable. Will try for myself soon. The yoji shinkawa style was convincing to me
Flux "uncensored" refuses to draw a dinosaur eating people, and so many other things, so I take those claims with a grain of salt ... i'll check that 3.5 though. I am using SDXL much more now
Is there a1111 version?
That CHIN!
Doubt it, this is just one of the delayed sd3 models from the initial release.
UNION controlnet and ipdapter are much beter on SD arch.
i still have hope
7:11 totaly agreed... there are lack of style inside flux and you play anything with prompt you get same generic look.
Can always spot the Flux chin.
not as good as Flux but I cant use Flux Dev because of the stupid licence
Lets lora this.
You say 8B parameters is "very very" close to 12B parameters as if it wasn't just 66% of the parameters.
Flux is like 10x better as it stands right now.
So, have we abandoned Auto1111?
pretty much yes. A1111 times are over. let it to rest...
and it's for the best :
i advise Krita AI diffusion nowadays. a very easy to use interface for image generation inside an excellent image editing software.
it run on comfyUI backend and as such, there is no limit to what you can do inside it without dealing with a single node !!!
Try reforgeui so much better than auto1111
devs let a1111 rot
Flux Killer ? i dont think so.
Tested both, compared results, FLUX is still better.
Did you compare it with Flux schnell?
yes FLUX is beter as it is.
but you cannot train it efficiently, it is not as open and business friendly.
so to me as a Lora maker maniac i will jump on SD3.5 as soon as Ipadapter and controlnet gets decent.
coz i know that this model will become AT LEAST beter than FLUX !
@@julx97 , Yes. And other versions as well. Even schnell is better than SD3.5
Only for now tough. SD3.5 is going to get a ton of community support and upgrades while flux wont because it is harder to augment and train.
@@yaroaram7237 , Ppl were saying the same about SD3 😆 and now 3.5 is out. Kinda doubt it'll be smth revolutionary. I'm still very satisfied with what FLUX Does atm. If someone can't get any desired result, I think it's more of a prompting issue, not the FLUX.
Flux still the king ...
It's just another SDXL (2.0)
i dunno there still isnt much for me to move past sdxl
i just make silly memes though
Focus on Flux!
Hype 🎉
yes
Flux still rulez! And SD 3.5 is not uncensored!
👀