The main benefit of Flux was never "realism" or whatever because you already could achieve the same level of that in SDXL (albeit with more work, like inpainting fucked up areas etc). What sets Flux apart is the integrated LLM and the associated prompt understanding and adherence. With Flux, you can input one sentence in completely natural language, and even if the concept is very abstract, Flux will give you precisely what you asked for. Whereas in SDXL / SD1.5, if you want something specific you need to combine 3 paragraphs of keywords with like 50 schizo negatives until it gets even close to what you want, and even then you need like a dozen re-gens before it gives you something usable. I hate how slow Flux is but I'm using it more and more just because the prompting style is so refreshing. Some ML nerd should investigate whether it might be possible to put an LLM into Stable Diffusion.
While I still like the sdxl lightning models, flux is playing in a different league imo. From what I've tested so far I prefer realvisxl v4.0 over the new v5.0 mainly because v5.0 gives me plastic skin in portraits. Maybe there are other areas where 5.0 shines, but I don't know them yet.
This could occur if training dataset for V5.0 contains a large number of overprocessed images. For instance, images that have been processed using frequency separation techniques, which are well known in Photoshop, tend to produce a "plastic-like" appearance in skin textures...
@@fooboomoo While I prefer flux overall, I agree with you that it does do way too plasticy skin. I've read that people do img2img with sdxl or even sd1.5 based models to fix skin, but haven't tried it yet. Honestly I doubt I will in the near future. Hopefully there are many community trained models to come, guess I'll wait and see what they bring.
The "UI" Options in the upper left are just additional presets instead of having only one. This makes it easy to quickly switch the overall settings while switching model types.
Of which I am not a fan because they delete all the settings makes it a pain to switch different model types with the same prompts. And the defaults are not even always fitting any ways, just for the base models.
@@madrooky1398 That's why I've kept two installs - my original one for SD1.5 and SDXL and a second just for Flux, although they share the same model folders.
@@Elwaves2925 It actually used to be possible to open the same UI in two tabs and use different settings in each, it still is, but since Flux there is a bit of an issue properly unloading a model when starting generating in the other tab with a different model type.
thanks for the tutorial sir, but there are too many models/checkpoints to try, would you like to create best top 10 models with its sampler setting in this December?
@Elwaves, I think SDXL has this limitation. With my Flux experience, I didn't find this problem till now. I use N4 and FP8 models on my 4070 12 GB with resolution under 1024. worked for me, though the speed for first run is slow, take about a min, but then it do fast generation. Another part is prompting. Flux understands prompts better. specially with two or more people.
Great vids as always! I want to say though, when you record your vids, instead of green screens you might want to use white screens because it's still segment you the same but the bleed from white screens are much more universally compatible with the light from whatever background you swap to. Just my little 2 cents for ya!
Please, can u do a video training a model with latest dreambooth version?, I have the version of stable diffusion v1.10.1 and is a bit different of the version that you use in your last training dreambooth video. Continue doing that great job!! Thanxs in advance
Background and Foreground seem to be from 2 different images... the edges are like scissor cuts..... Face Loras made with this Checkpoint and used on are almost always elongated, stretched ...even the 1024x1024....
Its only a little bit more precise, but lot of fine details still wonky. To me it seems there is too much upscaled training images that have broken details in those data sets.
@@danielc121 clip skip 1 is default, and if not stated otherwise using other settings most probably does not do anything good. Only very few models can profit using clip skip 2, but i have only a little bit knowledge about it. As far as I have tested it, the benefit especially in quality is very very limited. You basically just skip one layer of the text encoder which can lead to a more specific image generation, basically skipping a high level class. And how it affects the outcome might be very different depending on a lot of different parameters, so I avoid making general statements about this options effect.
Olivio, doesn't need to change anything in his title because he wasn't talking to YOU in the first place, he is actually very much engaged into making the difficult less difficult to digest (understand, grasp) by people not that smart as yourself. There's like nuances and nuisances in this world which paint wrong ideas around intelligence, or savoir faire. The reality, like the deeper one, is the fact that one cant go without the other, for example, some persons can not stand the reality of blood bursting out of someones else's body, This being said, not everyone can manage to clean up really bad road accidents...but hey some do, and we love it, they are like heroes we only meet in times of despair, its absolutely fine you don't work as a firemen or medic, just make sure to keep your ego in check, this wasn't about hands and feet, this is about understanding not everyone carries the understanding you do, in other words, you should/could ? the intelligence you got for free and share it the same way Olivio does, it makes the world including your own a better place for all, ❤
Thanks for a great tutorial! I'm a bit of a newbie on the subject, so forgive me a newbie question. What's "first roles or second roles image"? I was trying to find the explanation but couldn't find anything 🤔
@@lunarismaarfilms no question is a newbie question :) it means that the images shown were generated with only 1 or 2 tries, and not chosen from a large amount of processed images (with the same prompt and settings, but different seeds).
The scifi prompt looks like she jumped right out of a Mass Effect game... Also women having just a touch of peach fuzz isn't unrealistic, even if the makeup and fashion industry would like to convince you otherwise.
Hi. a5 3:53 in your video when you show the A1111 interface, in the top left you have the option to select the Checkpoint type (SD, XL, FLUX, all) how do I turn this on? I have latest version but I'm not seeing this option. Also does that mean that Flux works in A1111 now?
Hi, that's not A1111 you see, that's Forge UI version 2 (very similar to A1111 but supports Flux). Olivio Sarikas has a video where he talks about Forge UI 2.0: ua-cam.com/video/o6wM4M445zM/v-deo.html ("STOP Wasting Time with A1111, Use Forge UI Instead")
@@xeverian5591 A lot of people say Forge is so much faster than A1111 but when i tested it I saw literaly no difference in time when generating same image. Is the difference only apparent with low end cards? If yes, then are there any other advantages that Forge has over A1111? I've recently tested ComfyUI but apart from it generating completely different images with the same prompts ive seen no appeal over A1111, like with Forge.
But it’s still a SDXL model and uses the weak text encoder right? Meaning, it can’t handle difficult and unusual prompts well, properties leaks across subjects, etc.
@@KlimovArtem1 Because Flux is very rigid and does not do styles well or any type of non casual stuff. And not even talking about nudity, Flux can't do certain clothing styles or make up or fictional themes at all.
@@KlimovArtem1 Because as fantastic as Flux is, SDXL can still do a few things better (better being somewhat subjective of course). For one, I like to mix and match celeb names to get a new combo of the two but Flux doesn't handle celebs names well, not without loras. Secondly, there's character consistency. If you have SDXL characters you aren't going to get that same character from Flux without a lora. Thirdly, humans in Flux, especially females seem to be trained on supermodels IMO, so it's harder to get away from that while SDXL can do it easily. Obviously it all depends on what you want but SDXL still has it's place.
I can see the quality for solo portraits, without hands. The fine skin details is something I miss in many other models, even in Flux. For anything else the models seems to be very fragile. And the control over details is not in my favour too. The speed might be an argument though. Its easy to spit out many images and scrap the bad ones. And for what its worth, the often broken details like teeth and eyes can easily be fixed with Flux in img2img (maybe along upscale with Flux) and very low denoise (0.1-0.2) but i am not sure if its worth the effort. Flux is also removing the nice skin details and it might be necessary to use Photoshop and merge the images so you can keep the best of both. But overall, I prefer realistic Pony variants or Flux depending on the topic. I am actually very impressed with the Pony models and how well they do with prompt adherence, much better than anything I have seen in XL, and in my case most often only very specific negatives.
@@sca04245 Nope, i have generally lost interest in XL models except Pony realism fine tunes. XL is in my book too inconsistent and difficult to prompt. Just tested this a bit because of the video, but I am not going to hold on to it. There will be better options with Flux in the near future.
@@OlivioSarikas sorry to disturb you, but i followed your instructions for the forge ui and wether I use flux or realvisxl i get "connection errored out" message whenever i try to generat an image. Have you heard of this error before?
Kommt ein wenig auf deine Hardware an was man empfehlen sollte. Auf meiner alten 2060 mit 12gb und in ForgeUI läuft das Hyperflux1-dev_bnb_nf4_fp8 wirklich gut und schnell (8steps). Und vor allem out of the box. Dazu das Lora "anatomy_fineart_nudity_by_caith". Gibt aber viele Loras die mit der Art von merge nicht funktionieren, vor allem Character Loras. Gibt wahrscheinlich mittlerweile Varianten die etwas besser sind aber hatte noch keine Lust neue zu testen. Jedenfalls brauchste für sowas kein Tutorial, einfach nachlesen, normalerweise musste die 2 clip and das eine VAE noch manuell hinzufügen was kein Hexenwerk ist. Falls du noch das normale a1111 benutzt, weiß nicht ob da Flux überhaupt läuft.
Thank you for saying it - Flux is the most overrated, over-hyped model so far - I'm glad you're re-focusing the conversation back to one of the many other good models available.
Tested the fp16 version of this model, not the lightning, for a couple of days - but had problems with smaller faces, when heads are not as large as in portraits. Even Adetailer struggled with fixing these. It gets worse if you try do depict celebrities. It looks like they deliberately did this. Can anybody confirm, or give a hint how to improve faces here?
thanks for the video, can it do proper text and does it follow prompts such as "an orange xxx on the left, a blue yyy on the right, in the foreground a black zzz in the background a white aaa is sitting over a eeee"? If not...then it's not better than Flux and that makes it clickbait title :D But that's the way things work nowadays unfortunately...
Most of the models have portriates of 25yo pretty models down. I find that when I ask for a 34yo or a not perfect model person, they struggle. They will do 25yo or 70yo if pushed. Also I don't want portraits. I want wider shots!
i heard gguf q5 is on par with q8 with smaller size, or nf4 maybe, i've been using gguf q4 and fp8 turbo on my rtx 2060 super without any troubles, and your machine sure could perform better.
No offense but this model ain't that great, I'd rather stick with Juggernaut if I don't wanna use flux. I just gave this model a go n it ain't all that promising
Sorry Olivia But literally every picture you showed has bigger or small defects ... Such quality was ok a few months ago but now it looks just silly ...
@@mirek190 Well, yes and no. It has flaws, as I wrote in my own comment about it. But it can create very nice images with details even Flux can't do. And its for free for you to use, calling it "unacceptable" is silly of you. Very silly. The entitlement stinks, to be frank. Where is the amazing free model you have fine tuned?
@@madrooky1398 have you seen details? Have no sense ... look at them. Materials are not coherent and weird. Look at the women how and what is dressed... No sense , no coherent dress or armour , look on the room and shelves ... just lol Even a parrot is broken and literally takes whole picture. Gorrila fur is also strange. Juggernaut I think is making better job than this one. Only skin looks..ok on that model but that can be fixed under Flux easily by realistic lora. ..actually skin is better under flux with lora ...
@@mirek190 Perhaps your settings are wrong, I got some pretty fine results after experimenting. But Only portrait without hands or background do work. Hands suck with almost all regular XL models. Pony realistic variants are much more reliable. In regards to Flux, of course if you have a Lora that works well. I can't wait to see proper fine tuned models, they are going to be amazing. Generally I don't like Lora, only few that really work well without impacting base model quality. There is so much spam of crappy stuff. I have trained Loras aswell but never uploaded because I don't want to produce more trash...^^
I feel like the thumbnail is dishonest and does a disservice to your usual efforts to provide accurate decent information because it is misleading. It is not better than Flux, but it is worth a try for some users.
I'll check it out, Pony and SDXL models are still quite viable even after FLUX, Pony models specially have a huge library of highly varied style and character LORAS that Flux doesn't have.
Very viable. I'm still using an earlier RealVis model because it has the right looks for the celebs and made up names I've used, giving me consistency. Flux is fantastic but my two minor gripes are that the females all seem to be trained on supermodels, requiring work to get away from that and celeb names don't work. Pony has a big head start on loras, so that's to be expected but Flux loras are turning up pretty quick IMO and they're much easier to train accurately.
@@DaniDani-zb4wd It's slower but it depends on your system. I have 12Gb VRAM and while it's slow to load the model first time, after that it's fine for me. I wouldn't want to run it on anything less than that.
@@Elwaves2925 ofc it depends on your system but I don’t see any pint in spending that much money just to have a bit of fun with ai. I am running it on a laptop with 6gb vram. It takes 2 min 40 seconds to generate an 896x1152 image. I think I’ll just wait one more year before upgrading since I see no point in having both a laptop and pc ( what s shame laptop has lowvram in general)
@@DaniDani-zb4wd Absolutely, nobody is saying you should spend money. You came in saying it was slow, which it would be for you on 6Gb as it isn't designed for your specs. I was just pointing out that just because it's slow for you doesn't mean it's slow for everyone else. 🙂
Flux is not that good to begin with, it produces the same face and as soon as you push it, it produces CG images and it literally sh!t itself when your prompt is too long.
From a raw quality perspective, Yes. From a style perspective, No. Flux is very rigid and stuck to a casual realistic style, which is not what many people are looking for.
The main benefit of Flux was never "realism" or whatever because you already could achieve the same level of that in SDXL (albeit with more work, like inpainting fucked up areas etc). What sets Flux apart is the integrated LLM and the associated prompt understanding and adherence. With Flux, you can input one sentence in completely natural language, and even if the concept is very abstract, Flux will give you precisely what you asked for. Whereas in SDXL / SD1.5, if you want something specific you need to combine 3 paragraphs of keywords with like 50 schizo negatives until it gets even close to what you want, and even then you need like a dozen re-gens before it gives you something usable. I hate how slow Flux is but I'm using it more and more just because the prompting style is so refreshing.
Some ML nerd should investigate whether it might be possible to put an LLM into Stable Diffusion.
Wow!! thanks for this!
I'm getting really cinematic generations without the use of any style loras. I'm liking the prompt adherence so far.
yes, i love the style of this model too :)
Love the giant censorship bar!
The censorship monolith!
@@carmelodistasio1454 that reminds me of something else... 😅
is it censored or what you mean?
Nah, it's just a Roblox dress.
The bar that brings the boys to the yard ;)
While I still like the sdxl lightning models, flux is playing in a different league imo.
From what I've tested so far I prefer realvisxl v4.0 over the new v5.0 mainly because v5.0 gives me plastic skin in portraits. Maybe there are other areas where 5.0 shines, but I don't know them yet.
This could occur if training dataset for V5.0 contains a large number of overprocessed images. For instance, images that have been processed using frequency separation techniques, which are well known in Photoshop, tend to produce a "plastic-like" appearance in skin textures...
flux especially gives me very plasticy skin. What do you do to avoid that?
@@fooboomoo While I prefer flux overall, I agree with you that it does do way too plasticy skin. I've read that people do img2img with sdxl or even sd1.5 based models to fix skin, but haven't tried it yet. Honestly I doubt I will in the near future. Hopefully there are many community trained models to come, guess I'll wait and see what they bring.
@@fooboomoo try lowering the guidance
Thanks again for another recommendation. Which Vae are you using?
Thaaaanks!! 🤩 Btw.. is that forge UI?
Any setup for comfy?
Its a normal SDXL model. You use it like you would use a SDXL model in Comfy UI.
The "UI" Options in the upper left are just additional presets instead of having only one. This makes it easy to quickly switch the overall settings while switching model types.
they also change the UI though. at least for the Flux model
@@OlivioSarikas indeed, good point!
Of which I am not a fan because they delete all the settings makes it a pain to switch different model types with the same prompts. And the defaults are not even always fitting any ways, just for the base models.
@@madrooky1398 That's why I've kept two installs - my original one for SD1.5 and SDXL and a second just for Flux, although they share the same model folders.
@@Elwaves2925 It actually used to be possible to open the same UI in two tabs and use different settings in each, it still is, but since Flux there is a bit of an issue properly unloading a model when starting generating in the other tab with a different model type.
This is one of many SDXL models for my choices for a while.😉
Can it work with FOOOCUS ?
thanks for the tutorial sir, but there are too many models/checkpoints to try, would you like to create best top 10 models with its sampler setting in this December?
@@duniasainsasrofi this one is one of the main ones tho
show me the hands!!!!
hands will always s*ck because most devs don't pay attention to it
@Elwaves, I think SDXL has this limitation. With my Flux experience, I didn't find this problem till now. I use N4 and FP8 models on my 4070 12 GB with resolution under 1024. worked for me, though the speed for first run is slow, take about a min, but then it do fast generation. Another part is prompting. Flux understands prompts better. specially with two or more people.
Great vids as always! I want to say though, when you record your vids, instead of green screens you might want to use white screens because it's still segment you the same but the bleed from white screens are much more universally compatible with the light from whatever background you swap to. Just my little 2 cents for ya!
Will you do a review of the new juggernaut model as well?
how is the result of hands and feet
It's SD, it's going to be sh!t as usual.
THAT is a very good question
Similar to other SDXL models.
Thanks! Do you get better results on forge than comfyui?
Please, can u do a video training a model with latest dreambooth version?, I have the version of stable diffusion v1.10.1 and is a bit different of the version that you use in your last training dreambooth video. Continue doing that great job!! Thanxs in advance
Background and Foreground seem to be from 2 different images... the edges are like scissor cuts..... Face Loras made with this Checkpoint and used on are almost always elongated, stretched ...even the 1024x1024....
Is there something like this with support to add external model? Like clothes, designs and so on? Would be brilliant for products.
can you do image to image with it ?
Right now, I'm using FluentlyXL and Juggernaut. I guess it's cool to ditch this model and switch it out for this one.
never heard of FluentlyXL. i will check it out
Olivio, have you tried the Fp32 version of this model? thanks a lot!
Its only a little bit more precise, but lot of fine details still wonky. To me it seems there is too much upscaled training images that have broken details in those data sets.
@@madrooky1398have you notice that the autor suggest to use clip skip 1, does that affect the prompt?
@@danielc121 clip skip 1 is default, and if not stated otherwise using other settings most probably does not do anything good. Only very few models can profit using clip skip 2, but i have only a little bit knowledge about it. As far as I have tested it, the benefit especially in quality is very very limited. You basically just skip one layer of the text encoder which can lead to a more specific image generation, basically skipping a high level class. And how it affects the outcome might be very different depending on a lot of different parameters, so I avoid making general statements about this options effect.
i never noticed any effect from clip skip with sdxl based models
@@fooboomoo once for accident I use clip skip 3 in realism engine and I could not believe the result I got lol
You should change the title of this video to "You NEED to try this to see the ultimate compilation of bad hands and feet!"
Olivio, doesn't need to change anything in his title because he wasn't talking to YOU in the first place, he is actually very much engaged into making the difficult less difficult to digest (understand, grasp) by people not that smart as yourself. There's like nuances and nuisances in this world which paint wrong ideas around intelligence, or savoir faire. The reality, like the deeper one, is the fact that one cant go without the other, for example, some persons can not stand the reality of blood bursting out of someones else's body, This being said, not everyone can manage to clean up really bad road accidents...but hey some do, and we love it, they are like heroes we only meet in times of despair, its absolutely fine you don't work as a firemen or medic, just make sure to keep your ego in check, this wasn't about hands and feet, this is about understanding not everyone carries the understanding you do, in other words, you should/could ? the intelligence you got for free and share it the same way Olivio does, it makes the world including your own a better place for all, ❤
@@simbarules777 Are you stuck in a car crash ?
how are we inpainting with it?
Thanks Olivio, how'd it do with hands?
and how do you get the Ui bar?
I try lora to this, but it does not work. THX for share!
Thanks for a great tutorial!
I'm a bit of a newbie on the subject, so forgive me a newbie question. What's "first roles or second roles image"? I was trying to find the explanation but couldn't find anything 🤔
@@lunarismaarfilms no question is a newbie question :) it means that the images shown were generated with only 1 or 2 tries, and not chosen from a large amount of processed images (with the same prompt and settings, but different seeds).
@@celiaa6872 Thanks so much! :) That makes sense!
The scifi prompt looks like she jumped right out of a Mass Effect game...
Also women having just a touch of peach fuzz isn't unrealistic, even if the makeup and fashion industry would like to convince you otherwise.
Hi. a5 3:53 in your video when you show the A1111 interface, in the top left you have the option to select the Checkpoint type (SD, XL, FLUX, all) how do I turn this on? I have latest version but I'm not seeing this option. Also does that mean that Flux works in A1111 now?
Hi, that's not A1111 you see, that's Forge UI version 2 (very similar to A1111 but supports Flux). Olivio Sarikas has a video where he talks about Forge UI 2.0: ua-cam.com/video/o6wM4M445zM/v-deo.html ("STOP Wasting Time with A1111, Use Forge UI Instead")
@@SwampySi this is Forge - basically a mod for A1111 which is much better and faster, try it
@@xeverian5591 A lot of people say Forge is so much faster than A1111 but when i tested it I saw literaly no difference in time when generating same image. Is the difference only apparent with low end cards? If yes, then are there any other advantages that Forge has over A1111? I've recently tested ComfyUI but apart from it generating completely different images with the same prompts ive seen no appeal over A1111, like with Forge.
But it’s still a SDXL model and uses the weak text encoder right? Meaning, it can’t handle difficult and unusual prompts well, properties leaks across subjects, etc.
For text I've been using flux and then img2img controlnet with sdxl models.
@@gotenks157 but if using Flux, why do you need sdxl?
Exactly
@@KlimovArtem1 Because Flux is very rigid and does not do styles well or any type of non casual stuff. And not even talking about nudity, Flux can't do certain clothing styles or make up or fictional themes at all.
@@KlimovArtem1 Because as fantastic as Flux is, SDXL can still do a few things better (better being somewhat subjective of course). For one, I like to mix and match celeb names to get a new combo of the two but Flux doesn't handle celebs names well, not without loras. Secondly, there's character consistency. If you have SDXL characters you aren't going to get that same character from Flux without a lora. Thirdly, humans in Flux, especially females seem to be trained on supermodels IMO, so it's harder to get away from that while SDXL can do it easily. Obviously it all depends on what you want but SDXL still has it's place.
how run it on cpu or igpu intel?
I can see the quality for solo portraits, without hands. The fine skin details is something I miss in many other models, even in Flux. For anything else the models seems to be very fragile. And the control over details is not in my favour too. The speed might be an argument though. Its easy to spit out many images and scrap the bad ones.
And for what its worth, the often broken details like teeth and eyes can easily be fixed with Flux in img2img (maybe along upscale with Flux) and very low denoise (0.1-0.2) but i am not sure if its worth the effort. Flux is also removing the nice skin details and it might be necessary to use Photoshop and merge the images so you can keep the best of both.
But overall, I prefer realistic Pony variants or Flux depending on the topic. I am actually very impressed with the Pony models and how well they do with prompt adherence, much better than anything I have seen in XL, and in my case most often only very specific negatives.
Have you tested portraits against v4.0?
@@sca04245 Nope, i have generally lost interest in XL models except Pony realism fine tunes. XL is in my book too inconsistent and difficult to prompt. Just tested this a bit because of the video, but I am not going to hold on to it. There will be better options with Flux in the near future.
Hey, this deems awesome. How do i install this? I have a very powerful pc
i have videos for both comfyui and forgeui install on my channel. it's best if you have a nvidia card
@@OlivioSarikas thanks, I'll look them up. I've got an rtx 4090.
@@OlivioSarikas sorry to disturb you, but i followed your instructions for the forge ui and wether I use flux or realvisxl i get "connection errored out" message whenever i try to generat an image. Have you heard of this error before?
how to turn on gpu weight ui or ui for select xl sdxl etc
Buddy u got sub for you work ❤
Great Video, but my question is, how did you bring Flux in SD to run? Can you make a Tutorial please? Regards from Germany
you mean flux running in webui forge?
Kommt ein wenig auf deine Hardware an was man empfehlen sollte. Auf meiner alten 2060 mit 12gb und in ForgeUI läuft das Hyperflux1-dev_bnb_nf4_fp8 wirklich gut und schnell (8steps). Und vor allem out of the box. Dazu das Lora "anatomy_fineart_nudity_by_caith". Gibt aber viele Loras die mit der Art von merge nicht funktionieren, vor allem Character Loras.
Gibt wahrscheinlich mittlerweile Varianten die etwas besser sind aber hatte noch keine Lust neue zu testen. Jedenfalls brauchste für sowas kein Tutorial, einfach nachlesen, normalerweise musste die 2 clip and das eine VAE noch manuell hinzufügen was kein Hexenwerk ist.
Falls du noch das normale a1111 benutzt, weiß nicht ob da Flux überhaupt läuft.
Very nice
Thank you for saying it - Flux is the most overrated, over-hyped model so far - I'm glad you're re-focusing the conversation back to one of the many other good models available.
Will it work in Comfyui?
Yep. I'm just getting it from Civitai
Of course
Anything works in comfy, anything.
@@deadlyrobot5179 Do you think comfy gives you better results than a1111?
With comfy UI is working literally everything
will it work with Draw Things?
The model is a nightmare. Penguin hands.
I have done quite a bit of testing with it and it is not as good looking as my latest Jib Mix Realistic 15 model.
Tested the fp16 version of this model, not the lightning, for a couple of days - but had problems with smaller faces, when heads are not as large as in portraits. Even Adetailer struggled with fixing these.
It gets worse if you try do depict celebrities. It looks like they deliberately did this. Can anybody confirm, or give a hint how to improve faces here?
I've been using XL for a while, and nothing I've seen compares to Flux. Flux can now run on 6gb VRAM, perhaps even less
Is there an gguf versions of the fp32 ik there'd an 6.4gb version but is there one in gguf for like 2gb or 4gb size
thanks for the video, can it do proper text and does it follow prompts such as "an orange xxx on the left, a blue yyy on the right, in the foreground a black zzz in the background a white aaa is sitting over a eeee"?
If not...then it's not better than Flux and that makes it clickbait title :D But that's the way things work nowadays unfortunately...
Most of the models have portriates of 25yo pretty models down. I find that when I ask for a 34yo or a not perfect model person, they struggle. They will do 25yo or 70yo if pushed. Also I don't want portraits. I want wider shots!
which flux model would you guys recommend for me. I'm using RTX 4070 Ti Super 8gb rvam with 16gb memory. Shall I go for GGUF?
i heard gguf q5 is on par with q8 with smaller size, or nf4 maybe, i've been using gguf q4 and fp8 turbo on my rtx 2060 super without any troubles, and your machine sure could perform better.
@@XTINGAX thx dude!!!
I would try... But I changed my AI faith to Flux recently 😅
Finally you created your own extra "talented woman" wink wink xd
No offense but this model ain't that great, I'd rather stick with Juggernaut if I don't wanna use flux. I just gave this model a go n it ain't all that promising
i like the analog style. jugger is good, but it looks like any other ai generated image
Bro have you tested this model fully look at the hands and toes 😂😂😒. Come on....
you do know that there are hand and feet fixing tools
downloaded 📂
Sorry Olivia
But literally every picture you showed has bigger or small defects ...
Such quality was ok a few months ago but now it looks just silly ...
latent upscale... for shame.
You don't need to censor them just make the black bar 30% and call it an educational try-on Haul video.
Flux has always been over hyped tbh and I don’t know how people see the world when it comes to realism
hmmm… perhaps it’s time to release my image generator….
Really good but not sure it's better than Flux.
Please, if you censor the NSFW women censor the lip hair too. 😄
yeaaaa... no. this video smells of paid to promote.
No, yeah?
@@madrooky1398
For nowadays use that model looks silly.
Look how many big or small artifacts are there ... unacceptable.
@@mirek190 Well, yes and no. It has flaws, as I wrote in my own comment about it. But it can create very nice images with details even Flux can't do.
And its for free for you to use, calling it "unacceptable" is silly of you. Very silly. The entitlement stinks, to be frank. Where is the amazing free model you have fine tuned?
@@madrooky1398 have you seen details?
Have no sense ... look at them.
Materials are not coherent and weird.
Look at the women how and what is dressed... No sense , no coherent dress or armour , look on the room and shelves ... just lol
Even a parrot is broken and literally takes whole picture.
Gorrila fur is also strange.
Juggernaut I think is making better job than this one.
Only skin looks..ok on that model but that can be fixed under Flux easily by realistic lora. ..actually skin is better under flux with lora ...
@@mirek190 Perhaps your settings are wrong, I got some pretty fine results after experimenting.
But Only portrait without hands or background do work. Hands suck with almost all regular XL models. Pony realistic variants are much more reliable.
In regards to Flux, of course if you have a Lora that works well. I can't wait to see proper fine tuned models, they are going to be amazing. Generally I don't like Lora, only few that really work well without impacting base model quality. There is so much spam of crappy stuff. I have trained Loras aswell but never uploaded because I don't want to produce more trash...^^
I feel like the thumbnail is dishonest and does a disservice to your usual efforts to provide accurate decent information because it is misleading. It is not better than Flux, but it is worth a try for some users.
LMAO... stop this Olivio.. at most a refiner for Shnel
I'll check it out, Pony and SDXL models are still quite viable even after FLUX, Pony models specially have a huge library of highly varied style and character LORAS that Flux doesn't have.
Very viable. I'm still using an earlier RealVis model because it has the right looks for the celebs and made up names I've used, giving me consistency. Flux is fantastic but my two minor gripes are that the females all seem to be trained on supermodels, requiring work to get away from that and celeb names don't work.
Pony has a big head start on loras, so that's to be expected but Flux loras are turning up pretty quick IMO and they're much easier to train accurately.
@@Elwaves2925flux is just so slow….
@@DaniDani-zb4wd It's slower but it depends on your system. I have 12Gb VRAM and while it's slow to load the model first time, after that it's fine for me. I wouldn't want to run it on anything less than that.
@@Elwaves2925 ofc it depends on your system but I don’t see any pint in spending that much money just to have a bit of fun with ai. I am running it on a laptop with 6gb vram. It takes 2 min 40 seconds to generate an 896x1152 image. I think I’ll just wait one more year before upgrading since I see no point in having both a laptop and pc ( what s shame laptop has lowvram in general)
@@DaniDani-zb4wd Absolutely, nobody is saying you should spend money. You came in saying it was slow, which it would be for you on 6Gb as it isn't designed for your specs. I was just pointing out that just because it's slow for you doesn't mean it's slow for everyone else. 🙂
Flux is not that good to begin with, it produces the same face and as soon as you push it, it produces CG images and it literally sh!t itself when your prompt is too long.
Commercials license is all I need otherwise the model is useless
It is still very cartoonish, it is even worse than Dalle-3 or MJ v3... Not for me!
Lol😂 not even close to Flux
Dont feal réal at all
flux намного лучше
From a raw quality perspective, Yes. From a style perspective, No. Flux is very rigid and stuck to a casual realistic style, which is not what many people are looking for.
Women have "lip hair"? So, the model is politically-correct.
the only question... is it 100% free speech?
5:30 Excuse me?! Who doesn't like a beautiful woman with a bit of "lip hair"?! Nice mustache is what makes a woman perfect 😂😂😂
2:28 I don't trust the guy to the right of her.
He looks like a 2-faced liar. (it's a joke you will get when you see him).