I got it working and added my images and had it add captions. I hit "start training" and after about 10 seconds it says "Training Complete. Check the outputs folder for the LoRA files." Which is obviously way too fast. Any idea why it thinks it's done when it didn't do anything? I kept all the settings at default and I have a beefy NVidia card. Thanks!
Try to reinstall fluxgym again that means the installation was not done correctly, or there is some files missing, update your NVidia drivers and restart it again, test with as little images as possible let’s say 5, to see if it works. Once everything is installed correctly it should start working.
@@AiMotionStudio Ok.. I actually had this happen yesterday and reinstalled it today. Pretty certain all the models are in the right folders. I did it just like the tree displayed it. So I'll try updating my Nvidia driver and see if that does anything. Thanks for the advice!
@@AiMotionStudio There was one error in the train log... on line 39 it said " RuntimeError: use_libuv was requested but PyTorch was build without libuv support" Any idea what that could be? Thanks again!
Images can’t be too big because the fluxgym will automatically crop the images into 512 as required before the training, I will suggest that you check if the model is downloaded correctly and it’s 23GB in size. I am pretty sure a crucial file is missing causing this to happen.
Hi, I got it setup and running, put in 20 images running a 3080, but it gets stuck at downloading the 23gig checkpoint flux model (terminal does not show other next steps). Would this just be a console error and should I wait out or does this mean it is not working at all?
I will suggest that you manually download the flux dev 1 model that’s is 23GB and place it inside the model sub-folder under the fluxgym folder and then run the training again
RTX 4070 with 12 or 16GB VRAM will be more faster, but expect at least 5 hours or more, just set it up during the night when you don't need your computer and run it, once you wake up the next morning it should be done! Training 50 images is more than enough to get very good results.
Hi @minamalak5115! If the LoRA isn’t giving consistent results, here are a few things you can check: Ensure the LoRA is properly loaded: Double-check that you’ve applied the LoRA to the correct model and that it’s being activated during inference. Adjust weight scaling: Try tweaking the LoRA’s weight values (commonly between 0.5 and 1.0) to see if it improves consistency. Verify the base model compatibility: Some LoRAs are trained on specific models. Using an incompatible base model can result in inconsistent outputs. Seed consistency: Set a fixed random seed to ensure reproducible results. Let me know if any of this helps or if you need more details!
Set your chrome to dark mode to en-force the changes, at the bottom right of a new tab page, select customize chrome. Under “Appearance” select Dark and chrome will change to dark theme throughout the browser.
You can train it for 2-3 hours if you're only working with a few images, like 5. I was training with almost 40 images, which is why it took much longer. For nearly perfect training on a person's face, you'll need around 10 different variations of images to get the best results.
No it won’t work due to the fact that the program needs NVidia’s CUDA AI module for training, however you can train via third party website using there hardware check out replicate ai or Runpod. I am working on tutorial for this soon. Stay tuned
try to use comfyUI workflow forge didn’t accept the Lora for me as well I think this is because you trained the Lora with a Full Flux Dev model and in forge you are using either the NF4 flux or the gguf flux versions, try to use the same flux.1 dev you trained it with as the checkpoint diffuser main model to generate. That will surely work also try it in comfyUI not forge.
@@AiMotionStudio Hello. I don't understand here. Isn't this Fluxgym supposed to be for those with low VRAM? That's why it has the 12gb option right? But if in the end you need a high VRAM to run flux.1 dev to work with the lora then it make no sense. If you could clear up my confusion that would be appreciated. I have a 3060 and I'm not sure if I should give it a try or just give up. Thank you very much!
Do you mean the workflow? which you can find on civitai.com to get the dev1 with Lora workflow Or the comfyUI manager with the CPU/GPU/VRAM monitor? You need to just install the comfyUI-Crystools custom node which you can find and download via GitHub, just copy the GitHub code URL and install directly inside your comfyUI manager, and restart the comfyUI to see the monitor.
Did you check to see that the training started ? Check if the GPU is active during training via task manager, also make sure it’s completed installation with all files
@AiMotionStudio ops sory its rtx 3070Ti ,misspelled above And i dont think all in all is only vram bcz u have just 3060 and i have 3070 but Ti which is advance version so maybe its 8gb more powerful then simple 3060 hmmm???? B.d.w r u girl
Set your chrome to dark mode to en-force the changes, at the bottom right of a new tab page, select customize chrome. Under “Appearance” select Dark and chrome will change to dark theme throughout the browser, also paste this to your chrome: chrome://flags/#enable-force-dark | then enable the Auto Dark Mode for Web Contents.
Its worth noting that to not make several lora models in advanced settings the ,, Save every N epochs '' should be set to max epochs we have set in basic settings. Everything else on point
Hi @H2_S0_4! Great tip! 😊 Yes, setting the 'Save every N epochs' to the max epochs in the advanced settings can definitely help prevent the creation of multiple LoRA models. It ensures that you’re saving only the final version of the model. Thanks for sharing, and I'm glad the rest of the setup was helpful!
You can offer this service on Fiverr for clients looking to have professional photographs of themselves without the need for an actual photoshoot. This can also be used for creating pro-level ads where people are holding or interacting with a product. The application varies depending on the specific needs of the project, whether it's for personal use or commercial advertising purposes. check out our other video we trained an electric car for an ads, we also made ads for clients using this method.
Hi @Larimuss! You’re absolutely right-the default learning rate can be too high, especially when traing a large dataset. For 50 images, you might want to lower the training timeframe to prevent over training or unstable training. Try reducing it incrementally (e.g., 1e-4 or 5e-5) and monitor the results. A lower rate generally leads to smoother convergence with larger datasets. Let me know if you want tips on fine-tuning!"
Hi @CreativeGluvia-b2g! The ENOENT: no such file or directory error usually happens when the system can’t find the file or directory you’re trying to access. In this case, it seems like there’s a problem with how the fluxgym.git directory or file is being referenced. Here’s what you can do: Check file path: Ensure the file path is correct and there are no typos or missing folders in the directory. Re-clone the repository: Try deleting the fluxgym.git folder and re-cloning it from the correct repository to ensure everything is in place. Check for special characters: The use of {{input.event[0]}} in the file path seems unusual-double-check if this is meant to be a variable or needs to be replaced with an actual file name. Permissions: Ensure you have the proper permissions to access that directory, especially if you're on Windows. Let me know if that works or if you need further help!
I got it working and added my images and had it add captions. I hit "start training" and after about 10 seconds it says "Training Complete. Check the outputs folder for the LoRA files." Which is obviously way too fast. Any idea why it thinks it's done when it didn't do anything? I kept all the settings at default and I have a beefy NVidia card. Thanks!
Try to reinstall fluxgym again that means the installation was not done correctly, or there is some files missing, update your NVidia drivers and restart it again, test with as little images as possible let’s say 5, to see if it works. Once everything is installed correctly it should start working.
@@AiMotionStudio Ok.. I actually had this happen yesterday and reinstalled it today. Pretty certain all the models are in the right folders. I did it just like the tree displayed it. So I'll try updating my Nvidia driver and see if that does anything. Thanks for the advice!
@@AiMotionStudio Ugh, Nvidia says I have the latest drivers... Can my images be too big? Some are 3000x2000px.... They are all png files. Thanks!
@@AiMotionStudio There was one error in the train log... on line 39 it said " RuntimeError: use_libuv was requested but PyTorch was build without libuv support" Any idea what that could be? Thanks again!
Images can’t be too big because the fluxgym will automatically crop the images into 512 as required before the training, I will suggest that you check if the model is downloaded correctly and it’s 23GB in size. I am pretty sure a crucial file is missing causing this to happen.
thank you for the lesson and the trippy music :)
Glad you liked it!
click the link below to stay update with the latest AI tutorial 👇🏻 :
www.youtube.com/@AiMotionStudio?sub_confirmation=1
Hi, I got it setup and running, put in 20 images running a 3080, but it gets stuck at downloading the 23gig checkpoint flux model (terminal does not show other next steps). Would this just be a console error and should I wait out or does this mean it is not working at all?
I will suggest that you manually download the flux dev 1 model that’s is 23GB and place it inside the model sub-folder under the fluxgym folder and then run the training again
@@AiMotionStudio Oh okay I will try that, thanks for the quick response and tip!
Did it work bro?
@@brechtxbrecht
everything is free in this tutorial?
hello and thanks for your awesome guidance, is it possible to share your comfyui workflow ?
hi! good video, thanks. Do you kwnow how can i setup the output of files, for to be safetensors files? Now it make me other files formates.
the Loras should be authomatically be saved as .safetensors files check to see that you setup the fluxgym installation correctly.
@@AiMotionStudio stf теперь нету safetensors
I have rtx 4070....how much time it will take to train with around 50 images..how many hd images required on an avg to get best output ?
RTX 4070 with 12 or 16GB VRAM will be more faster, but expect at least 5 hours or more, just set it up during the night when you don't need your computer and run it, once you wake up the next morning it should be done! Training 50 images is more than enough to get very good results.
@@AiMotionStudio thanks 👍
why are they always 17mb tenor files? I want them to be bigger and take in more detail.
you will have to train the models for more hour to make the size of the Loras bigger but i don't think that will affect the quality of the Loras.
i did the same thing but when i try to put the lora it never gets me the same person it's like the lora doesn't exist any idea what did i do wrong ?
Hi @minamalak5115! If the LoRA isn’t giving consistent results, here are a few things you can check:
Ensure the LoRA is properly loaded:
Double-check that you’ve applied the LoRA to the correct model and that it’s being activated during inference.
Adjust weight scaling:
Try tweaking the LoRA’s weight values (commonly between 0.5 and 1.0) to see if it improves consistency.
Verify the base model compatibility:
Some LoRAs are trained on specific models. Using an incompatible base model can result in inconsistent outputs.
Seed consistency:
Set a fixed random seed to ensure reproducible results.
Let me know if any of this helps or if you need more details!
after click Train, i get all kinds of errors, sometimes error can't find "sample" folder
sometimes error "list index out of range"
check to make sure that all the flux models are downlaoded, also check if the folder the error is coming from has all the required files in it!
@@AiMotionStudio Hi, after pull the version that the author just fixed, it worked.
@@AiMotionStudio i tried train again with images+parameters same as previous success trained task, but error again 😅
I don't know what it wants 🤣
Hi,please how to change trainer page to dark mode??,i have selected dark mode but trainer is white.
Set your chrome to dark mode to en-force the changes, at the bottom right of a new tab page, select customize chrome. Under “Appearance” select Dark and chrome will change to dark theme throughout the browser.
@kasperskygroup also paste this to your chrome: chrome://flags/#enable-force-dark | then enable the Auto Dark Mode for Web Contents
If we don't have a strong computer, can we use this system on the cloud?
Please explain if there is a way. Thanks
Yes you can run this on Runpod or any cloud service that you can rent a GPU.
Yes you can use Runpod or any cloud service that you can rent a GPU.
@@AiMotionStudio Thanks, do you have any tutorial about it?
Not yet I may do that in future!
@@AiMotionStudio Hope so
looks like its useless to use it if u have not a 4090 rtx ... 10 hours training... its a pain, is there a different way to use Flux Lora?
You can train it for 2-3 hours if you're only working with a few images, like 5. I was training with almost 40 images, which is why it took much longer. For nearly perfect training on a person's face, you'll need around 10 different variations of images to get the best results.
It does not work in Mac Os Silicon M2...... a lot of errors during FLUX installation
No it won’t work due to the fact that the program needs NVidia’s CUDA AI module for training, however you can train via third party website using there hardware check out replicate ai or Runpod. I am working on tutorial for this soon. Stay tuned
I have 4gb vram will it work?
Yes how ever it will take a very long time I will suggest that you use flux Pulid instead. Check out our Flux Pulid video tutorials
I try to make this and FluxGym give me a Lora but is in anime style,i don't undestand why,help please,i use forge for generate
try to use comfyUI workflow forge didn’t accept the Lora for me as well I think this is because you trained the Lora with a Full Flux Dev model and in forge you are using either the NF4 flux or the gguf flux versions, try to use the same flux.1 dev you trained it with as the checkpoint diffuser main model to generate. That will surely work also try it in comfyUI not forge.
@@AiMotionStudio Hello. I don't understand here. Isn't this Fluxgym supposed to be for those with low VRAM? That's why it has the 12gb option right? But if in the end you need a high VRAM to run flux.1 dev to work with the lora then it make no sense. If you could clear up my confusion that would be appreciated. I have a 3060 and I'm not sure if I should give it a try or just give up. Thank you very much!
@@AiMotionStudio thanks i try this
How do I load the same setup in ComfyUI that you have?
Do you mean the workflow? which you can find on civitai.com to get the dev1 with Lora workflow Or the comfyUI manager with the CPU/GPU/VRAM monitor? You need to just install the comfyUI-Crystools custom node which you can find and download via GitHub, just copy the GitHub code URL and install directly inside your comfyUI manager, and restart the comfyUI to see the monitor.
install the ComfyUI-Crystools Custom Nodes to get the CPU/GPU/VRAM Monitors into your comfyUI.
I cant install flux gym with pinokio 😢
Yes you can
i got black images always while training why?
Did you check to see that the training started ? Check if the GPU is active during training via task manager, also make sure it’s completed installation with all files
is my rtx 307Ti is way more powerful then urs 3060 for this ???
Yes I believe so although it also depends on the VRAM you have mine is 12GB
@AiMotionStudio ops sory its rtx 3070Ti ,misspelled above
And i dont think all in all is only vram bcz u have just 3060 and i have 3070 but Ti which is advance version so maybe its 8gb more powerful then simple 3060 hmmm???? B.d.w r u girl
how to activate dark theme?
Set your chrome to dark mode to en-force the changes, at the bottom right of a new tab page, select customize chrome. Under “Appearance” select Dark and chrome will change to dark theme throughout the browser, also paste this to your chrome: chrome://flags/#enable-force-dark | then enable the Auto Dark Mode for Web Contents.
Its worth noting that to not make several lora models in advanced settings the ,, Save every N epochs '' should be set to max epochs we have set in basic settings. Everything else on point
Hi @H2_S0_4! Great tip! 😊 Yes, setting the 'Save every N epochs' to the max epochs in the advanced settings can definitely help prevent the creation of multiple LoRA models. It ensures that you’re saving only the final version of the model.
Thanks for sharing, and I'm glad the rest of the setup was helpful!
le music c'est too bloody loud :P
not working for me
Please state the issue specifically!?
@@AiMotionStudio UserWarning: cuDNN SDPA backward got grad_output.strides() != output.strides()
how can i get money working with this? can I sell my trained models? this is amazing!
You can offer this service on Fiverr for clients looking to have professional photographs of themselves without the need for an actual photoshoot. This can also be used for creating pro-level ads where people are holding or interacting with a product. The application varies depending on the specific needs of the project, whether it's for personal use or commercial advertising purposes. check out our other video we trained an electric car for an ads, we also made ads for clients using this method.
The "model" is an adult film star named "alexa grace" 😂😂😂😂
"wink-wink"😂😂
The default learning rate seems way to high. Good for 20 images but probably horrible for 50
Hi @Larimuss! You’re absolutely right-the default learning rate can be too high, especially when traing a large dataset. For 50 images, you might want to lower the training timeframe to prevent over training or unstable training.
Try reducing it incrementally (e.g., 1e-4 or 5e-5) and monitor the results. A lower rate generally leads to smoother convergence with larger datasets. Let me know if you want tips on fine-tuning!"
annoing music.. grr
just some ai music !
8vram crash don’t even run
yes i most say that for 8vram you might want to try to train few images so you can use less vram
ENOENT: no such file or directory, stat 'C:\pinokio\api\fluxgym.git\{{input.event[0]}}'
the issue when downloading fluxgym
Hi @CreativeGluvia-b2g! The ENOENT: no such file or directory error usually happens when the system can’t find the file or directory you’re trying to access. In this case, it seems like there’s a problem with how the fluxgym.git directory or file is being referenced.
Here’s what you can do:
Check file path:
Ensure the file path is correct and there are no typos or missing folders in the directory.
Re-clone the repository:
Try deleting the fluxgym.git folder and re-cloning it from the correct repository to ensure everything is in place.
Check for special characters:
The use of {{input.event[0]}} in the file path seems unusual-double-check if this is meant to be a variable or needs to be replaced with an actual file name.
Permissions:
Ensure you have the proper permissions to access that directory, especially if you're on Windows.
Let me know if that works or if you need further help!
Ten hours 😂
Yes it’s better you set it up before you go to sleep and when you wake up the Loras are ready.
Yes it’s better you set it up before you go to sleep and when you wake up the Loras are ready.