00:06 FLUX LoRA allows for easy AI image generation and Lura training. 01:58 Training a LoRA model for creating AI influencers and replicating styles. 03:39 Utilize different image sets for effective LoRA training 05:29 Enabling sampling for monitoring Laura's evolution during training 07:23 Setting up models and checkpoints in FLUX LoRA 09:17 Using LoRA for image generation with high resolution 11:04 Adjusting LoRA weight for style variation 12:49 Combining luras is fun, but weights need to be adjusted for complex prompting. This is awesome. I will be doing this a bit later. Thanks for what you do.
@@AiVOICETUTORso Im also doing similar things. Out of respect are you okay if I do a similar voice dub tutorial. My live streams always fail. Seriously don't want to take your style but I do alot with ElevenLabs. Just say don't and ill move on.
Very very good tutorial - you have chosen your words very well without and with great guidance. Finally a video where the selection of the pictures is REALLY explained. Thank you! :-) There is a template of FluxGym on RundPod too - in case you want to use a RTX4090 :-) Could you please do another FluxGym tutorial wfor the advanced settings?
Thanks, I'm glad you like the tutorial. I do plan to make a follow up video that deals with the advanced settings but I'd also like to be able to tell what the ideal settings are for certain kind of image sets. And I'm still gathering info on all that from my own experiments and from what others share online.
Tried it with my 1080Ti 11gb 64GB RAM, trained 10 Images and it would take 3 Days for the training to be finished lol. I retested the same settings on my Predator Laptop with RTX 4060 and it took just about 4hrs. The difference is night and day, although good ol 1080TI is still kicking if you're really patient enough.
after I got the message that training was complete, I went to the lora folder but I didn't find the actual lora file inside, all I found was dataset, readme, train, and sample_prompts files Edit: and yes the path is a local hard drive with no spaces in the folder names, a large free storage as well (300+ gb)
My Lora doesn't generate close to what I put in for the images, I am making a Lora of myself so maybe a Lora of the Rock is easier as it's a well known celebrity. Is it better to add more photos, I have 20 at the moment from various backgrounds and locations from over a few years of selfies, or is it better to raise the trainings per image to 20 or 30 or higher? It takes a long time to run it, going to try different ways, takes nearly a day for me to get a LoRA with my hardware.
I have trained a LoRA of myself with about 20 images and the default settings and it works (I trained another one with about 50 images which works even better). Maybe try to improve your image set and you could also try lowering the trainings per image to like 5 or 10 and set the epochs to about 10. This way you can experiment without waiting too long. Fingers crossed!
I can't replicate the samples I get in Fluxgym. The samples I get are awesome. When I try to replicate them in Forge with the same seed and prompt they are nowhere near. Tried multiple Sampling methods and Schedule types. It would be interesting to know with what settings the Fluxgym samples are generated.
I'm mostly using an older version of Cogstudio (the one that has frame interpolation and upscaling). Think it only uses 6GB of VRAM but it's also very slow. I plan to do make a video about some good ai video tools soon.
I'm doing it right now. Google local lab fluxgym lora training pod. He has a YT video with link to the template. Works great. Setup a network volume though for lora so if you shut it down and haven't DL lora yet you can save it.
pleasepleaseplease~~~Hello, I followed the steps to train, and after the training was completed, there were other training files in the OUTPUT folder, but why was there no file model of Lora?
so, you train wiht dev and the models will also work on schnell. But maybe the training will be faster on schnell, and it will also work on dev. A win win.
Hello, I followed the steps to train, and after the training was completed, there were other training files in the OUTPUT folder, but why was there no file model of Lora?
I cannot use Loras trained in Fluxgym. Terrible image generations! Has anything changed? It was working great last month. I even installed it on another ssd.
Thanks for letting me know but for some reason, RUclips isn't showing the link in the video even though I set it correctly (had this issue before and no idea how to fix it). But yeah it's the video that is over a year old now (ruclips.net/video/5i_Pyw0gH-M/видео.html). I've tried all open source Ai voice cloning tools that came out since then and while the all were much easier to use (easier training to no training at all etc.), none of them came even close to the quality of RVC V2.
@AiVOICETUTOR hey mate, please i need help, really desperate here. i'm trying with my rtx 2050, 4gb vram, 30gb ram. i followed every steps in your video, but i'm facing many problems here. the first one is that my generations are slow, it can take around 30 mins for just 1 image. and the second is that, when i'm using my lora, the generations are really dark, like full black and just showing the face. is there a way to train and use my lora other than using flux ? because when i'm trying other model than flux it just doesnt recognize the lora. how can i use the lora i trained in another ( sd, or else ) please ? or do i need to train another lora with a different model than flux ? if yess how ? ty again
Hello, sir. Can i get your training dataset for the rock’s lora? I’m currently dealing with the bad results of my own trained lora. I’ve done everything i can. So i would like to figure out what went wrong by training with proper dataset
Hi ! Thanks for your video, it's very instructive. However, I have a problem with the text. Do you have a tip for the generated text that respects what I asked him to write in the prompt ? I also have a problem when I want to put two lora of characters I trained myself. It mixes them up, or gives me twins. I haven't found a solution yet.
Hi, for the text you could try repeating the text in the prompt once or twice. But sometimes it just takes a few tries. Great question about the two different people but sadly I also don’t have a solution for that. When I tried, I was mostly getting twins that looked like they are the children of the characters of the LoRAs
@@AiVOICETUTOR Hi. Thank you for your reply. For the text, I'll try what you suggested. One more question. As I don't have a big graphics card, I can't go too high in terms of Expected training steps. Would you advise me to use a bigger number for Repeat trains per image or for Max Train Epochs?
Sorry for the late reply. You could try setting higher steps rather than epochs. Though some people are having good results with lower trains per image (like 5 or 10) and with 10 epochs.
Not a dumb question at all. It seems like only Flux Schnell can be used commercially. You can find the licenses here: github.com/black-forest-labs/flux/tree/main/model_licenses
@@СлесарьЛеха My computer has nvidia geforce rtx 3070 ti with 8GB VRam (and I think a total of 16GB RAM). I'm not that tech-savvy, will this method will work for me?
1 image is probably not enough but it‘s worth a try if you get find a few variations of that logo (or a few similar logos). Theres loras on civitai trained for logos so that might help too.
@@AiVOICETUTOR If i wanted to train my own personal logo, how would i go about doing this? if 1 image of my logo isnt enough, what types of variations should i add please?
you have the most standard settings, this is a video for dummies, on flux for training it is better to choose fewer repeats and more epochs 1-2-4/50. these are the very first recommendations, not to mention extra settings.
From what I gathered since FluxGym launched 2 weeks ago is that there are no perfect settings to train a LoRA and that it depends on countless things. I have seen great LoRAs with all kind of different settings (including many with the default setting). Some people suggest keeping the repeats at 10 but slightly lowering the epochs while others lower the repeats and increase the epochs as you said. Some say you should keep the total steps between 1000 and 2000 while others do 15000 steps. I guess it also depends on what each persons standard for a good LoRA is. Like is it enough when you can generate an image similar to what you want 4 out of 10 tries? Or do you want every mole and tattoo of the person to be accurate 8 out of 10 times? I‘m sure at one point, we will have enough Information so that we will be able to look at an image set and then set the “correct“ training settings according to the characteristics of that set.
The one without a number in the end, is the full Lora model. The other ones are just the steps while training. You can sort them by date, then you will see which was created first and last.
@@AiVOICETUTOR Thanks but.. thats not the case. I installed flux gym using pinokio, did everything the same and now it worked. There are multiple people reporting the same issue with examples on github. Maybe just something with github version
It seems like only Flux Schnell can be used commercially if that's what you meant. You can find the licenses here: github.com/black-forest-labs/flux/tree/main/model_licenses
@@AiVOICETUTOR thanks for the answer, you are the first one on the English-language channel who answered me 😃 and so I figured it out myself, I already made a lore with my face, now I'm trying to do it with a resolution of 1024x1024. video card 4060ti 16gb
Do you know which folder pinokio downloads the flux1.dev.sft into? I already have the file so just want to move it into the correct folder rather than downloading it again.
![[Bad] - YouTube ВСЁ? (победа VK и прощание с подписчиками)](/img/1.gif)
00:06 FLUX LoRA allows for easy AI image generation and Lura training.
01:58 Training a LoRA model for creating AI influencers and replicating styles.
03:39 Utilize different image sets for effective LoRA training
05:29 Enabling sampling for monitoring Laura's evolution during training
07:23 Setting up models and checkpoints in FLUX LoRA
09:17 Using LoRA for image generation with high resolution
11:04 Adjusting LoRA weight for style variation
12:49 Combining luras is fun, but weights need to be adjusted for complex prompting.
This is awesome. I will be doing this a bit later. Thanks for what you do.
Thanks I'm glad that it's helpful.
@@AiVOICETUTORso Im also doing similar things. Out of respect are you okay if I do a similar voice dub tutorial. My live streams always fail. Seriously don't want to take your style but I do alot with ElevenLabs. Just say don't and ill move on.
Hey man, absolutely! Go for it! Feel free to use whatever concepts from my videos you think work well.The more decent tutorials, the better!
Probably one of the best guides I have come across so far. Well, done.
Thank you!
Grisu ! ❤ not have seen in decades. Thanks for the time travel
great ai voice and good info. I can't stand so many ai voices. well done!
Cool, thanks for letting me know!
tnx alot, was veryusefull for me.
Thanks for commenting, Glad It was useful!
Very very good tutorial - you have chosen your words very well without and with great guidance. Finally a video where the selection of the pictures is REALLY explained. Thank you! :-) There is a template of FluxGym on RundPod too - in case you want to use a RTX4090 :-) Could you please do another FluxGym tutorial wfor the advanced settings?
Thanks, I'm glad you like the tutorial. I do plan to make a follow up video that deals with the advanced settings but I'd also like to be able to tell what the ideal settings are for certain kind of image sets. And I'm still gathering info on all that from my own experiments and from what others share online.
Good and old 1080ti 11gb, do you think it gonna work? thankssss
I have the same GC, so I'm just commenting to get the reply notification xD
Yeah I think you guys will be fine and won't have to retire that great GPU just yet.
@@Sebastian78120 I decided to start in Fooocus first. 1080ti is absolutely rocking everything here. Flying high...
@@AiVOICETUTOR 1080ti for Fooocus is awsome. Later i will try FluxGym
Tried it with my 1080Ti 11gb 64GB RAM, trained 10 Images and it would take 3 Days for the training to be finished lol.
I retested the same settings on my Predator Laptop with RTX 4060 and it took just about 4hrs. The difference is night and day, although good ol 1080TI is still kicking if you're really patient enough.
nice job
Thanks!
after I got the message that training was complete, I went to the lora folder but I didn't find the actual lora file inside, all I found was dataset, readme, train, and sample_prompts files
Edit: and yes the path is a local hard drive with no spaces in the folder names, a large free storage as well (300+ gb)
thanks mate, can you share your pc's hardware ??
and any recommendation for budget pc hardware to run AI training model ?
How can you make ultra realistic 4k videos with it? And does it work with wearing a watch? Like what if u have multiple watches how does that work?
My Lora doesn't generate close to what I put in for the images, I am making a Lora of myself so maybe a Lora of the Rock is easier as it's a well known celebrity. Is it better to add more photos, I have 20 at the moment from various backgrounds and locations from over a few years of selfies, or is it better to raise the trainings per image to 20 or 30 or higher? It takes a long time to run it, going to try different ways, takes nearly a day for me to get a LoRA with my hardware.
I have trained a LoRA of myself with about 20 images and the default settings and it works (I trained another one with about 50 images which works even better). Maybe try to improve your image set and you could also try lowering the trainings per image to like 5 or 10 and set the epochs to about 10. This way you can experiment without waiting too long. Fingers crossed!
Best guide ever! Just one query, can we use two loras while generating image? For example: Rock and Undertaker in a single frame. Can you guide please
You can use as many loras as you want at the same time
I can't replicate the samples I get in Fluxgym. The samples I get are awesome. When I try to replicate them in Forge with the same seed and prompt they are nowhere near. Tried multiple Sampling methods and Schedule types. It would be interesting to know with what settings the Fluxgym samples are generated.
Hi! Great video. I just wanted to ask can we do this without training the LoRA? Like what you showed in Fooocus tutorial but with Flux in Forge?
❤❤❤❤🎉😊 awesome
Thank you 🙏
Great explanation, I subbed. For training a "Style", not person, do we just need to have correct caption?
Thanks! Nope the captions are not needed for styles or people.
Question: what do you use when going for image to video and what's the recommended VRAM for the PC.
I'm mostly using an older version of Cogstudio (the one that has frame interpolation and upscaling). Think it only uses 6GB of VRAM but it's also very slow. I plan to do make a video about some good ai video tools soon.
@ i see
btw, is 16GB VRAM for GPU good enough or still too slow?
Is there a way to install fluxgym, pinokio etc in Runpod or rented GPU's ?
I just noticed a fluxgym template on runpod. Trying it out now.
Yeah it‘s possible to do it on Runpod. Hope it worked for you @DrSnakeGamez
@@AiVOICETUTOR yes, the template worked great and was a very easy setup 💪
I'm doing it right now. Google local lab fluxgym lora training pod. He has a YT video with link to the template. Works great. Setup a network volume though for lora so if you shut it down and haven't DL lora yet you can save it.
pleasepleaseplease~~~Hello, I followed the steps to train, and after the training was completed, there were other training files in the OUTPUT folder, but why was there no file model of Lora?
Try to make sure that the folder paths don't contain any spaces or special characters and aren't on a network drive.
Hey! Nice video! Could you use ComfyUI with the lora model that you trained? Or must it go through forge?
Hey and sorry for the late reply. Yes you can use the lora in Comfy or any tool that supports Flux loras.
so, you train wiht dev and the models will also work on schnell. But maybe the training will be faster on schnell, and it will also work on dev. A win win.
Hello, I followed the steps to train, and after the training was completed, there were other training files in the OUTPUT folder, but why was there no file model of Lora?
I have 4060 with 8gb vram - can you help me to install fluxgym for training purposes?
forge works great
my lora dont show up in output folder
I cannot use Loras trained in Fluxgym. Terrible image generations! Has anything changed? It was working great last month. I even installed it on another ssd.
You forgot to link the video where you show how to make this kind of voice. Is the one you did one year ago or there is a new one?
Thanks for letting me know but for some reason, RUclips isn't showing the link in the video even though I set it correctly (had this issue before and no idea how to fix it). But yeah it's the video that is over a year old now (ruclips.net/video/5i_Pyw0gH-M/видео.html). I've tried all open source Ai voice cloning tools that came out since then and while the all were much easier to use (easier training to no training at all etc.), none of them came even close to the quality of RVC V2.
What is your experience with "resize dataset images"? Why not 1024?
@AiVOICETUTOR hey mate, please i need help, really desperate here.
i'm trying with my rtx 2050, 4gb vram, 30gb ram. i followed every steps in your video, but i'm facing many problems here.
the first one is that my generations are slow, it can take around 30 mins for just 1 image.
and the second is that, when i'm using my lora, the generations are really dark, like full black and just showing the face.
is there a way to train and use my lora other than using flux ? because when i'm trying other model than flux it just doesnt recognize the lora. how can i use the lora i trained in another ( sd, or else ) please ? or do i need to train another lora with a different model than flux ? if yess how ? ty again
I dont have the base model option in fluxgym, I have updated fluxgym in pinokio but it still doensnt show up
Odd that its not showing up for you but Fluxgym should be using Flux Dev even if theres no option for it. So you should be good to go
is it possible to pause a train, turn off laptop, and retrain from the point when we paused?
8:35 - How do i know what model "suits" my gpu? thats the only thing thats not well explained, thank you for the tutorial!
Thanks, that's a good point. What I meant was "pick a model that fits into the VRAM of your GPU".
Hello, sir. Can i get your training dataset for the rock’s lora? I’m currently dealing with the bad results of my own trained lora. I’ve done everything i can. So i would like to figure out what went wrong by training with proper dataset
Hi ! Thanks for your video, it's very instructive. However, I have a problem with the text. Do you have a tip for the generated text that respects what I asked him to write in the prompt ? I also have a problem when I want to put two lora of characters I trained myself. It mixes them up, or gives me twins. I haven't found a solution yet.
Hi, for the text you could try repeating the text in the prompt once or twice. But sometimes it just takes a few tries. Great question about the two different people but sadly I also don’t have a solution for that. When I tried, I was mostly getting twins that looked like they are the children of the characters of the LoRAs
@@AiVOICETUTOR Hi. Thank you for your reply. For the text, I'll try what you suggested. One more question. As I don't have a big graphics card, I can't go too high in terms of Expected training steps. Would you advise me to use a bigger number for Repeat trains per image or for Max Train Epochs?
Sorry for the late reply. You could try setting higher steps rather than epochs. Though some people are having good results with lower trains per image (like 5 or 10) and with 10 epochs.
i cant get any LoRas to work with Forge
My trained Loras are all 17mb big, how can I make them bigger to include more detail?
This might be a dumb question but can the images we generate be used commercially?
Not a dumb question at all. It seems like only Flux Schnell can be used commercially. You can find the licenses here: github.com/black-forest-labs/flux/tree/main/model_licenses
How much time it will take on 3060 12GB?
approximately infinite
If you do 512, and max 10 pictures, and lower the repeats to 7 you might get it done overnight
@@СлесарьЛеха My computer has nvidia geforce rtx 3070 ti with 8GB VRam (and I think a total of 16GB RAM). I'm not that tech-savvy, will this method will work for me?
I HAVE RXT3060TI 8GB WITH 32GB CAN I TRAIN USE IT?
After I finished the video, the requirement of needing 12GB VRAM was lowered to 10GB but 8GB won't work for now. Maybe in the future.
I have one 8Gb and one 12Gb GPU, but I always end up using runpod...
Love to know how it fairs
i did try it but it takes soo much time so i cancel it
Only works on Nvidia? No Mac?
i downloaded flux dev, but flux gym downloads it again how can i solve it?
hey so when i train my model, it says training is finshed, but i dont get the safetensor file! please help!
Try to make sure that the folder paths don't contain any spaces or special characters and aren't on a network drive.
NVIDIA GeForce GTX 1070 with 8GB works? If not, do you know a good vps for run this or how to do it in runpod? I need cheaper. Thanks!
I've read that people got it to work with 8GB cards but I can't test it myself.
Can any one tell me the requirements for a computer to work with this method?
Pinokio does not run on a Macbook M1 😤
Is there a limit on how many images you can use on Training LoRA? TIA
There's a limit that can be adjusted: github.com/cocktailpeanut/fluxgym/issues/22
can loras be trained on specific logos? would i use just one image of the logo?
1 image is probably not enough but it‘s worth a try if you get find a few variations of that logo (or a few similar logos). Theres loras on civitai trained for logos so that might help too.
@@AiVOICETUTOR If i wanted to train my own personal logo, how would i go about doing this? if 1 image of my logo isnt enough, what types of variations should i add please?
Just look for a couple of similar logos that look like yours (and that look like the style you want to be able to generate)
It does not work in Mac Os Silicon M2...... a lot of errors during FLUX installation
Sorry it‘s not working for you. It might be worth to try it again once in a while in case it gets fixed.
you have the most standard settings, this is a video for dummies, on flux for training it is better to choose fewer repeats and more epochs 1-2-4/50. these are the very first recommendations, not to mention extra settings.
From what I gathered since FluxGym launched 2 weeks ago is that there are no perfect settings to train a LoRA and that it depends on countless things. I have seen great LoRAs with all kind of different settings (including many with the default setting). Some people suggest keeping the repeats at 10 but slightly lowering the epochs while others lower the repeats and increase the epochs as you said. Some say you should keep the total steps between 1000 and 2000 while others do 15000 steps.
I guess it also depends on what each persons standard for a good LoRA is. Like is it enough when you can generate an image similar to what you want 4 out of 10 tries? Or do you want every mole and tattoo of the person to be accurate 8 out of 10 times?
I‘m sure at one point, we will have enough Information so that we will be able to look at an image set and then set the “correct“ training settings according to the characteristics of that set.
Всё. Получилось! Создал своё лицо с разрешением 1024x1024 на карте 4060ti 16 gb. Использовал примерно 17 фото, заняло 20 часов…
Cool, glad it worked for you!
Только нвидиа? Амд?
Is there any platform that has Loras that I can download to combine with mine?
Yeah for example here: civitai.com/search/models?modelType=LORA&sortBy=models_v9&query=flux
I noticed fluxgym generated four tensor files but you only copied over one of them. Was there a reason for this?
The one without a number in the end, is the full Lora model. The other ones are just the steps while training. You can sort them by date, then you will see which was created first and last.
@alinguist9793 thank you 💪
Can I run it on a M1 Max Macbook pro?
No, sadly it needs a NVIDIA GPU.
@@AiVOICETUTOR what if I use Comfy UI instead?
Does anyone here get a message when training the lora "no samples" ? Someone posted this on github as well but.. no answers..
Maybe check if your folder with the samples (or the names of the images) have any special characters or spaces in them.
@@AiVOICETUTOR Thanks but.. thats not the case. I installed flux gym using pinokio, did everything the same and now it worked. There are multiple people reporting the same issue with examples on github. Maybe just something with github version
can we used this for product?
It seems like only Flux Schnell can be used commercially if that's what you meant. You can find the licenses here: github.com/black-forest-labs/flux/tree/main/model_licenses
How to do this all in ComfyUi???
You can download workflows for Comfy to train loras. Civitai has a few
Too bad i don't meet the VRAM requirement.
I've read that people got it to work with 8GB cards but I can't test it myself.
Can this be used in Macbook M1Pro?
nope, Fluxgym does not yet support Mac
А если у меня есть уже Forge, нужно ли его опять качать?
No you can just copy the finished lora into your forge folder and use it from there
@@AiVOICETUTOR thanks for the answer, you are the first one on the English-language channel who answered me 😃 and so I figured it out myself, I already made a lore with my face, now I'm trying to do it with a resolution of 1024x1024. video card 4060ti 16gb
Can I do with Rtx 4060ti 8gb vram ?
Please someone replay ?
@@narutolover626 I do with my RTX 3070 8GB.
@gunther6638 Thanks for confirming!
I’m gonna try on my 3080 ti 16gb. hope it will train fast 😂
Fingers crossed!
@@AiVOICETUTOR thanks. so I got 30 min training 13 images with 5 train steps and 8 epoch. is that normal?
Yeah you got a pretty good GPU so that speed should be normal
@@AiVOICETUTOR thanks
Is it free ?
Yes
If you have the hardware and time. Yes.
guys , Im starting out as an AI enthusiast making similar content
would appreciate your feedback!
Do you know which folder pinokio downloads the flux1.dev.sft into? I already have the file so just want to move it into the correct folder rather than downloading it again.
The file needs to be in pinokio/api/fluxgym.git/models/unet. It needs to be named flux1-dev.sft
Is there some Lora trainer that does not need to install 32402390492034 different software? That would be an instant hit.
Are you on discord or Telegram or Instagram?
Yeah same name on Instagram as here
Can i do this on my mobile 🥲
Sadly not. You can run it online but I haven‘t looked into that and it‘s not free either
This or rendernet?