Flux 1.1 Pro & Ultra Fine-Tuning API is HERE! + Custom Gradio UI Walkthrough

[@HealthwiseNursingInsights]

Thanks so much for this video! I've been experimenting with various parameters for training LoRAs and trying full fine-tuning as well. Interestingly, I found that full fine-tuning didn’t give me the results I was expecting. Even with LoRAs, increasing the training steps didn’t always improve the outputs. In fact, some of my best results came from models trained on fewer steps. There's definitely a lot to figure out when it comes to optimizing parameters and finding what works best with this model. Your insights here have been really helpful!

[@ababiya_worku]

well ill tell you something! its not about your training parameters, fine-tuning or Lora creation, its about your prompting! without a proper prompting, nothing is gonna work

[@deonix95]

why is there no raw in FLUX Pro 1.1 Ultra?

[@solporcima]

Hi Paresh, thank you, amazing job, can you fix the link of your repo on git, its getting 404

[@paresh-ranaut]

Thank you for pointing that out! I’ve made some updates and tested the new system-it’s now working correctly. Please follow the ReadME.md steps to set it up, especially these steps: -Rename config.example.json to config.json -Add your Flux Pro API key to the config.json file. -Ensure you’ve added credits to your Black Forest Labs account.

[@AusturlandUpplysingasida]

Great video, I'm running into issues though with finetuning, I upload the zip, set the parameters and start fine tuning. It will generate a fine tune ID but then nothing happens, I've done multiple runs and my BFL balance just stays the same as well.

[@paresh-ranaut]

Hi, thank you for sharing your concern! I ran the training on another system, and it worked perfectly. However, here are a few things you might want to check to ensure everything runs smoothly: Image Format: The uploaded file should be a Base64-encoded .zip. Inside the .zip, images must be in .PNG or .JPEG format. Ensure no other formats are included. Checking the Status: If you see an error like "Model not found" or "Invalid fine-tune ID", don’t worry! This is a minor bug and doesn’t affect the process. If credits have been deducted from your account, it means the training has started and is running in the background. Training Time: The training process may take anywhere between 5 to 10 minutes (or sometimes a little longer), depending on the parameters used. Once the training is complete, the status button will update to "Ready". Troubleshooting Steps: After starting the training, wait for 10 minutes, then check the status. If it’s still not ready, wait another 10 minutes and check again. Make sure to use the correct file formats as mentioned above, and keep an eye on your credits. I’m actively working on fixing the bug to ensure the status updates in real-time as the training progresses. In the meantime, please follow these steps, and everything should work as expected. Please check the Terminal in your IDE as well, that might help you figure out the issue. Let me know if you have any other questions or run into further issue.

[@deonix95]

I also encountered a problem that makes the same sid although it is set to 0

[@paresh-ranaut]

Yeah, for random seed generation, You should leave the seed box empty.

[@wkrisdiyanto]

hi, can you add flux fill pro with lora also ?

[@paresh-ranaut]

Sure, Will be working on those soon!

[@deonix95]

If I have a dataset with poor quality images but where the product is clearly visible, I get poor quality images on the output, but the old way of training lora did not distort the image quality even if the dataset in poor quality

[@paresh-ranaut]

I found that as well with these models, they really a on the provided dataset quality but also recreate the Fine-tuned object to have much more closer resemblance to the trained images compared to the dev model. I guess it goes both ways, could work great for you if you have high quality data ,not so much if you don't.

[@deonix95]

@ Can you please tell me in which resolutions to save photos, because I recently did a large 4k and I got a very bad training and can you tell me if I can use a product png file without background for training? I didn't find it in the documentation.

[@paresh-ranaut]

​@@deonix95 I’ve also struggled to find clear guidelines on the optimal resolution and formats for training. I’ve already sent a request to Black Forest Labs asking for more clarity on this, and I’ll definitely keep you updated if I hear back. In the meantime, I’ve had decent results using 1024x1024 resolution images in PNG or JPEG format, similar to what I used with the Flux Dev model before. Hopefully, this helps you while we wait for further updates. I’ll also share any tweaks or updates on the UI to ensure it works seamlessly with this setup.

[@HistoryAi76]

(In mobile)Hy sir, can you tell some methods where I use fulx 1 or flux 1.1 or pro(with loras or not) not flux shenell For free and unlimited I need it for my ai based content

[@paresh-ranaut]

@@HistoryAi76 Hello, The Flux 1.1 Pro models are not open sourced so they don’t let you download the model weights or the LoRA weights. But You can use the Flux 0.1 Dev model instead of Shnell as you mentioned. If you want to use it with LoRa, you’ll need to train the LoRa first-platforms like Replicate are great for this. Once trained, you can download the weights and use them locally with tools like Forge, ComfyUI, or Automatic1111. Forge is the easiest to set up in my opinion. Training LoRa is a one-time cost (around $10-$12, depending on training time), and after that, you can generate for free!(as libg as you have a good enought system with atleast 4-6gb Nvidia GPU), The speed will depend on your system’s GPU and RAM. I also made a video showing how to train LoRa on Replicate for your own product-But for using Local tools, there are tons of videos on youtube about each platform.