Finetuning Flux Dev on a 3090! (Local LoRA Training)

In this video I show how to set up the AI Toolkit training script by Ostris to finetune LoRAs on Flux Dev by Black Forest Labs. Please excuse the rough quality, I wanted to release it quickly.

I am planning a follow-up video with more comparisons and final results.

Links

[seto]

Possibly important to note:
If you've been pulling your hair out trying to figure out why the training image folder is causing issues when trying to run the script, you might need to put "c:\\" instead of "c:\". Don't know why this is the case for some and not for others, but I'm glad I finally figured it out.

[krakenunbound]

I literally just downloaded your LoRA randomly from CivitAI and then went browsing on youtube for something to watch. Curious about training my own Lora I came in to check out your video and see that your name is the same as what I downloaded. Cool!

[cuinheaven]

One key element missing from this video is how to populate the caption files that accompany the training images. Should it be comma separated words and phrases? New line separated? Any words or phrases to avoid? That sort of thing. Other than that it was a short and helpful video, thanks!

[_lm_]

Me: "yeah let's train some Lora for Flux"
My 8GB VRAM card: "calm down dude"

[flin]

Thanks for the video, easy to follow and helpful.

[doctordiffusion]

Thank you! They added a "examples" folder inside of the "config" so you need to add that to the final command.

[AgustinCaniglia]

I can't even run a flux lora on comfy but I am watching a lora training video.

[Gausch]

Awesome guide! Had problems with the yaml, but Claude helped me fixing it. Tried it with 2000 steps for finetuning on my face (22 photos), but thats not nearly enough. Trying 8000 next time.

[nomnomAI]

Thanks. I just got this working. had a pain with the ipadapter. i had to redo all of the steps to get it working. ty!

[televizor]

Just one question - I noticed that your dataset isn't exclusively 512x512, 1024x1024 (or 1:1 AR in general). I had an understanding that Flux wants square 512 images. Am I wrong? All aspect ratios are fine?

[niccolon]

followed the steps and im only getting 1 .txt file even though I have many images. In the CMD prompt it describes all my images but only saves 1 .txt file that describes 1 random image .. any idea?

[AiHello]

You missed out on explaining how the training imagers and captions are laid out. Overall helpful video though

[denismatejcik]

you can use "TAB" to autocomplete names in cmd so u dont have to type them and also avoid typos

[wonder]

Thanks for the tutorial, and looking forward to an update. After I manged to run the script, the models do not load. Loading 1 of 3 tensorfiles, and the first never completes. Using a 3090 with an Eizo monitior that might take more Vram. I think a fast connection speed is manditory. I will be upgrading to that mythical 100mb/s next week so I will try then. I am attempting to make a lora from my own work, which has a unique aesthetic, so I would rather do this locally. I can see potential in offering this possibility to non-AI artists who have a particular painting style. As way to generate ideas in their own visual language.

[taezonday]

Amazing! And here I was getting ready to instal Linux to run Simpletuner lol. Speaking of, do you know how Ostris compares with Simpletuner, are they using much different methods?

[CharlieLee-pq]

At the 12:57 mark of your video, I encountered a bunch of errors and can't progress any further than that. I followed your video from the beginning and installed it accordingly.

[TheKoingWolf]

I think including at least one sample caption would have been really good, so we can see how you captioned your dataset. Was it manually done too or run through an LLM for natural language captioning?

[henrywang]

Is there a way to cache the downloaded flux models so that they don't have to be download again later?

[MisterWealth]

Do you need to change steps or epochs depending on the amount of images you have?

[PetitShya]

what is (moduleNotFoundError: No module named"dotenv") ? I setup env file but……look like doesn't work? 🤔