This is great. I've been doing 3x3 character grids, cropping them out manually in Photoshop and then doing post work on each. This will be a big time saver for creating character data sets, thank you :)
Thanks a lot for sharing your sweet work! Gonna try this out this wknd. Please continue uploading flux workflows for consistent characters. I'm really curious to test it out for my project.
Thank you SO MUCH for this video. I literally spent 2 weeks trying to better understand how to create consistant characters, and training a lora. I finally got something to work, but wanted to try something better. I'll be giving this workflow a shot, and that Lora trainer. Your content is so good, at this rate you are going to end up with a new Patreon member. 😂
It' really great to see your tutorials developing just with the opportunities of AI 😄 Keep up your great work and thank you very much for sharing so much knowledge for free!
Mastered the advanced functions of flux and comfyui... also uses paint to crop photos. 😅 Jokes aside, this is really impressive. I used to use an equivalent setup on webforge, but much more primitive, to generate reference photos for my 3d modeling. Basically I had a simple rendering with my 3 views (face, side and 3/4) as input and I used inpainting to generate a set of reference photos. But that's another level !
Hi there. QQ:Perhaps I missed a step. When you get this deep into a character you most likely have at least one good shot generated. This seems like it is only applicable if you want a new character created. Is there an option to start with a character you have already generated and started using?
I once read that to train Flux Loras you don't need to tag the images, I'm telling you in case you want to try it the next time you train a Lora, greetings!
FYI, the non-numbered loras in the fluxgym output directory are the latest ones. You were using loras from epoch 12 (not necessarily a bad thing but you mentioned latest in video).
are you using the normal controlnet union or promax version for flux ? I don't get it, in the google doc for flux controlnet you put a sdxl controlnet link, is it an error ? I guess you used the union pro from instantx ?
Would you suggest getting the styling perfect before making the lora? I'm working on a project with a very distinct style (sd1.5), so using sdxl or flux now would be difficult.
love your videos , is there a way to input an image of a character and input a pose image and apply a controlnet on using both input images (keeping clothes,face consistent )? or using controlnet with the lora is the only way to achieve this ?
Great tutorial, great guy. I really like your workflows. But am I the only one where the workflow takes super long. A single run takes about 30 min with my 4080 😅
this is amazing. could this also be used to generate every angle of a character and then use gaussian splatting for example to generate a 3d model that can then be retopologized and rigged? that would be huge
Hey Mick ! Thanks so much for all those resources =) Which workflow would you use to animate a creature that is not totally humanoid ? Most of the process out there animate humans, or anime characters Thanks !
Thank you! Interesting video. With the “Expression Editor (PHM)” node, I get the error “invalid load key, ‘\x10’.” Maybe someone has already solved it, I haven't found an answer yet.
It isn't working well from the first step. It generates a bad sheet image of a cartoonish woman but like broken. I didn't change any setting (just checked that the model and controlNet files are correct). Any idea what is going wrong here?
Thank you for this but I've looked everywhere and I'm unable to find the "ControlNetApply SD3 and HunyuanDIT" node, is it part of a particular node-pack, or has it had a name change in the updated ComfyUI ?
very cool stuff! I'm having problems getting the ControlNET to generate the 'profile' heads in the first step. Any suggestions? (tried setting ControlNet weight as high as 1.5, no love). Maybe you have a super-high resolution version of the OpenPose image , and I could extract the profile (look up/flat/look down) from that and run those seperate? thanks! How did you make that image, maybe i can make my own with proper profile angles....
@@GNeuro interesting, can you elaborate? His WF doesn't use anything 'instantX' that I can find... for me, the controlNet is actually working OK (maybe not great for head poses) which I was able to partially solve by making new source image thats JUST the heads... but ideally I would have a way to make my own. Any thoughts on that?
I get Exception during processing !!! Allocation on device when it goes into the sampleCustomAdvanced. Im not really an expert on this. What does that mean?
As far as some other creator said, 12 Gig is the minimum. He tried it on a 8 Gig card but stopped it because it was taking over 8 hours and still hadn't completed. Maybe search to see if anyone has gone through with it. If you really want to do this you may need to rent a GPU online. I saw someone train a small lora like he did in this video for less than $10 Dollars.
Another great video. Question, what are the system requirements to run flux in comfyui? Can I run it on my 8GB mobile 4070? Is there a way to run fluxgym with 8gb vram ?Thanks.
Hey does anyone know how to make this workflow work on a 2020 M1 MacBook Air with only 8GB or RAM? Is there a good 3rd party website which can host 3rd party GPUS for this?
Can't load the workflows missing nodes, go to missing nodes shows no missing nodes to install :( missing like simplestring and expression editor in the SDXL one as an exmaple.
@@lexisgecko3623 I just got it to work. I was still missing a file. Once I got them all in the correct folders, I was able to install the missing nodes using the manager. Still getting a ton of issues, but the nodes work. The installation process is definitely not for beginners... sigh.
Same here I seem to have problems with the instantX Union Controlnet model. It does not seem to be working fine. Also I am not sure if it's the right one because I had to search for it myself, since the one in the document is not the one in the video, is an SDXL model while in the video the controlnet been used is the Union Flux
I have tried everything and could not make it work, but I think the rest of the models are fine and working fine because as soon as I lower the strength of the controlnet I get the character I am looking for, but just one image, not the char sheet. Could someone point to the good controlnet model? Please
@@Kaoru8168it isnt obvious as his previous tutorial started the same way, I thought he was doing another trick today, instead of a lora, that said why are you getting butthurt and all fanboyish while i already edited and mention that is was my bad?…..
Where do you have that InstantX_flux.safetensors from? Also when i use your linked flux\diffusion_pytorch_model.safetensors im getting the error Only InstantX union controlnet supported. Could not find key 'controlnet_mode_embedder.fc.weight' Got any idea how to fix that?
I have said it before, and I will say it again... "BRILLIANT!" Thanks so much for sharing your incredible insights and workflows.
The man keeps delivering. Million thanks.
You are an insipiration. Following you from the beginning & look how far you have come. Congrats & thanks for your guidance. 🔥🔥
You are some kind of wizard! That node sheet is beautiful! Class! 🌟
I’m beginning my AI journey and have no idea about this stuff yet, but I’ll bookmark for later!
This is great. I've been doing 3x3 character grids, cropping them out manually in Photoshop and then doing post work on each. This will be a big time saver for creating character data sets, thank you :)
This is insane. Thanks for sharing your workflow!
Your Channel is totally underrated. Fantastic learning curve... I tried your 2d set extension workflow couple of weeks ago... Awesome, thank you!
You crushed this topic. Thanks for sharing. Gonna subscribe and donate to Patreon.
Thank you so much for this informative video. You fulfilled my request. Thank you.
Thanks a lot for sharing your sweet work! Gonna try this out this wknd. Please continue uploading flux workflows for consistent characters. I'm really curious to test it out for my project.
Thank you SO MUCH for this video.
I literally spent 2 weeks trying to better understand how to create consistant characters, and training a lora. I finally got something to work, but wanted to try something better.
I'll be giving this workflow a shot, and that Lora trainer.
Your content is so good, at this rate you are going to end up with a new Patreon member. 😂
thank you man for sharing and keeping them free. YOU ARE A SAINT
Awesome Video Mick, Great work
It' really great to see your tutorials developing just with the opportunities of AI 😄
Keep up your great work and thank you very much for sharing so much knowledge for free!
Flux is such a game changer,
This is absolutely brilliant! Thank you for figuring this out and sharing with AI image generation enthusiasts!!!
Thank you! Very nice tutorial.
Mastered the advanced functions of flux and comfyui... also uses paint to crop photos. 😅
Jokes aside, this is really impressive. I used to use an equivalent setup on webforge, but much more primitive, to generate reference photos for my 3d modeling. Basically I had a simple rendering with my 3 views (face, side and 3/4) as input and I used inpainting to generate a set of reference photos. But that's another level !
Pure gold, thank you! 💪🏻🔥
Great job! Thank you!
Wow! Great tutorial!
Really impressive.
Excellent video , smooth learning 👌
SUPERB !! - liked subscribed.
Hi there. QQ:Perhaps I missed a step. When you get this deep into a character you most likely have at least one good shot generated. This seems like it is only applicable if you want a new character created. Is there an option to start with a character you have already generated and started using?
GEM!! thanks!!!❤
I once read that to train Flux Loras you don't need to tag the images, I'm telling you in case you want to try it the next time you train a Lora, greetings!
Thnak you so much for amazing videos , can you make video on how to make consistent character pose of custom character or self potrait .
FYI, the non-numbered loras in the fluxgym output directory are the latest ones. You were using loras from epoch 12 (not necessarily a bad thing but you mentioned latest in video).
Liked and subscribed. Great stuff
great ! just a small thing.. the last Lora generation is the one at the top with no numbers..Thank you
Yep, I left the same comment before reading this. (Don't worry, I deleted it. lol)
bro your the greatest
thanks for this!
Pretty cool. Though tinaai seemed to have the lizard's hands in that last image!
Dude nice video! GG
So cool
fluxgym is downloading models even run locally.
how to stop downloading models for fluxgym?
I already have models locally.
omg finally!!!!
Sehr nice, 😍 mal testen.
are you using the normal controlnet union or promax version for flux ? I don't get it, in the google doc for flux controlnet you put a sdxl controlnet link, is it an error ? I guess you used the union pro from instantx ?
+1
+1
감사합니다 👍👍👍
Hey! Great tutorial! What if I want to create consistent character from an image previously generated by another AI Tool?
Would you suggest getting the styling perfect before making the lora? I'm working on a project with a very distinct style (sd1.5), so using sdxl or flux now would be difficult.
can we use referance character image for create a consisten character?
Exactly what I wanna know! Begin with loading character.
I have 500MB of display memory, can I use comfy ui in my PC without any cloud services?
love your videos , is there a way to input an image of a character and input a pose image and apply a controlnet on using both input images (keeping clothes,face consistent )? or using controlnet with the lora is the only way to achieve this ?
Yes please
Great tutorial, great guy. I really like your workflows.
But am I the only one where the workflow takes super long. A single run takes about 30 min with my 4080 😅
this is amazing. could this also be used to generate every angle of a character and then use gaussian splatting for example to generate a 3d model that can then be retopologized and rigged? that would be huge
Hey Mick !
Thanks so much for all those resources =)
Which workflow would you use to animate a creature that is not totally humanoid ?
Most of the process out there animate humans, or anime characters
Thanks !
Thank you! Interesting video. With the “Expression Editor (PHM)” node, I get the error “invalid load key, ‘\x10’.” Maybe someone has already solved it, I haven't found an answer yet.
It isn't working well from the first step. It generates a bad sheet image of a cartoonish woman but like broken. I didn't change any setting (just checked that the model and controlNet files are correct).
Any idea what is going wrong here?
Great video, thanks for sharing, but I have 5700XT :( :D
I wish ComfyUI made more sense to me.
Thank you for this but I've looked everywhere and I'm unable to find the "ControlNetApply SD3 and HunyuanDIT" node, is it part of a particular node-pack, or has it had a name change in the updated ComfyUI ?
Can I use FluxGym for training a SDXL Lora aswell?
I wonder is it possible to use our own lora ?
During the sheet creation is it possible to implement loras there as well?
Did you check Flux license? Commercial use is not allowed for free.
0012 its not the last one. You need to check the time of creation of the files
how do you calculate the x y and width and height of image for cropping?
Not getting the full set of images in my generations, mainly just getting the 3 full body ones, nooobin - what can i do to adjust?
very cool stuff! I'm having problems getting the ControlNET to generate the 'profile' heads in the first step. Any suggestions? (tried setting ControlNet weight as high as 1.5, no love).
Maybe you have a super-high resolution version of the OpenPose image , and I could extract the profile (look up/flat/look down) from that and run those seperate? thanks!
How did you make that image, maybe i can make my own with proper profile angles....
For everyone, who meet trouble wit controlnet - i fixed it by intentional update the instantX node in comfyUi manager
@@GNeuro interesting, can you elaborate? His WF doesn't use anything 'instantX' that I can find... for me, the controlNet is actually working OK (maybe not great for head poses) which I was able to partially solve by making new source image thats JUST the heads... but ideally I would have a way to make my own. Any thoughts on that?
Is there a mistake in the docs? For the flux workflow it links to the SDXL controlnet model.
I get Exception during processing !!! Allocation on device when it goes into the sampleCustomAdvanced. Im not really an expert on this. What does that mean?
does 16gb vram enough to run all those?
How do I use fluxgym on a 8GB GPU? As always cracy good videos!
As far as some other creator said, 12 Gig is the minimum. He tried it on a 8 Gig card but stopped it because it was taking over 8 hours and still hadn't completed. Maybe search to see if anyone has gone through with it. If you really want to do this you may need to rent a GPU online. I saw someone train a small lora like he did in this video for less than $10 Dollars.
What kind of computer you operating?
Another great video. Question, what are the system requirements to run flux in comfyui? Can I run it on my 8GB mobile 4070? Is there a way to run fluxgym with 8gb vram ?Thanks.
You can run GGUF variants of Flux , Try GGUF Q5 or Q6 . Q6 will be slower on 8gb ram but you can make it work.
@@Jaswant7785Thanks!
how long it take to create lora on a lets say RTX 4070 ?
Hey does anyone know how to make this workflow work on a 2020 M1 MacBook Air with only 8GB or RAM? Is there a good 3rd party website which can host 3rd party GPUS for this?
What are the vram-ram needed, please? i have 9 vram and 32 ram.
Can it be done without proper pc ?
does this work on a MacBook
👏🏽👏🏽👏🏽👏🏽👏🏽👏🏽👏🏽👏🏽
With such data there will simply be overfitting, there will be no generalization
Can't load the workflows missing nodes, go to missing nodes shows no missing nodes to install :( missing like simplestring and expression editor in the SDXL one as an exmaple.
+1 😭
@@kehmesis you have this as well ?
@@lexisgecko3623 I just got it to work. I was still missing a file. Once I got them all in the correct folders, I was able to install the missing nodes using the manager. Still getting a ton of issues, but the nodes work. The installation process is definitely not for beginners... sigh.
The initial image is a garbled mess. Any suggestions on fixing that?
Same here I seem to have problems with the instantX Union Controlnet model. It does not seem to be working fine. Also I am not sure if it's the right one because I had to search for it myself, since the one in the document is not the one in the video, is an SDXL model while in the video the controlnet been used is the Union Flux
I have tried everything and could not make it work, but I think the rest of the models are fine and working fine because as soon as I lower the strength of the controlnet I get the character I am looking for, but just one image, not the char sheet.
Could someone point to the good controlnet model? Please
Hi 👋
Can I add my own face to this?
a step to AI movie making
wait.....Free workflow? Its behind a pay wall....
just do a lora?...
EDIT: lol ok you talk about it later, my bad :)
its obvious he talking about a custom character where you would get your dataset before creating original lora? the first steps are crucial my guy
@@Kaoru8168it isnt obvious as his previous tutorial started the same way, I thought he was doing another trick today, instead of a lora, that said why are you getting butthurt and all fanboyish while i already edited and mention that is was my bad?…..
imagine saying free multiple times; but you have to pay to get it
I thought you can't use Flux Dev for commercial use
Lipsync?
too much work, let's see if it get's easier a couple months from now
👍🏿 it's not there yet I say 6 Months to a Year but I respect he tried
Yes ! New ai models can easily be a feature
If it were easy, everyone would be doing it.
Something that would have taken months, done in 8 minutes but still too much work because it’s a little technical lolllllll
I hope everyone feels this way!! More opportunities for us nerds 😂
its not free....lol
Do not touch AI while you only have a crappy PC.
At this point its as much work as you would do it analog yourself
Where do you have that InstantX_flux.safetensors from?
Also when i use your linked flux\diffusion_pytorch_model.safetensors im getting the error Only InstantX union controlnet supported. Could not find key 'controlnet_mode_embedder.fc.weight'
Got any idea how to fix that?
Very helpful video❤❤ 🇮🇳 #nkheartbd #shayana