This tutorial will guide you through the process of training a LoRA (Low-Rank Adaptation) model and integrating it into your Weavy workflows.
[0:00:01.500] This is a really quick tutorial on how to train a LoRA, then bring it in Weavy and use it in your workflows.
[0:00:08.000] To finish this tutorial, I’m going to examine if training on a specific style is way better than just using some reference images with… “Nano-Banana,” for example.
[0:00:19.800] The first thing you need for training a LoRA is a dataset. So, if you’re training for your own image, it can be lots of beautiful pictures of you.
[0:00:27.300] And if you’re training for style, it could be lots of examples for your desired style.
[0:00:32.600] In both cases, you will need a zip file with all your images. The number of images in this file could be somewhere between 5 and 100, maybe even more. I go for 20, 30 examples. Usually, it works for me.
[0:00:47.800] So, first, we head over to Fal.ai or Replicate.com.
[0:00:56.500] Under “Explore” in Fal.ai, I am going into the “Training” tab to watch all the training options. You can see and check all of them for yourself.
[0:01:05.300] Today I’ll be training style. So I’m going for “Flux LoRA fast training.”
[0:01:09.200] And the UI is fairly simple. I’m picking up the zip file under my style.
[0:01:12.700] It has all of my images.
[0:01:16.000] There’s a debate around choosing the right trigger word. I recommend choosing a word that does not have a representation in the normal language. So I’m going for something like “ohwx”, which is totally random and style connected.
[0:01:31.500] Is it a style? Yes, it is. We can go for additional settings, but in most use cases, I leave this on default and clicking “Start.”
[0:01:40.800] We will see our training happening right here on the right side and will be notified when it’s finished.
[0:01:47.300] Once it’s completed, we’ll click on “Show Files” and download the diffusers LoRA file, *.safetensors file. Download it locally.
[0:01:57.000] Once it’s downloaded, let’s head over to our workflow tab to open the quick search menu and import LoRA.
[0:02:05.000] Clicking on the upload, uploading the safetensors file. So we can go back a little bit. While it’s uploading, I’m going to add Flux Dev LoRA, wait for it to finish.
[0:02:16.100] Connect the LoRA output to the LoRA input, add prompt and let’s try it with something pretty similar to what we trained on.
[0:02:24.300] A “Weavy” soap product in “ohwx” style. Please notice you have to call out your LoRA using the trigger word in your prompt. Run.
[0:02:33.900] Okay, we got some results. I think that text is not so good, but the style is pretty consistent.
[0:02:41.000] Okay, guys, so that’s the way to train your own LoRA and bring it inside Weavy and use it. Let’s go over it quickly.
[0:02:48.000] First, you go into Fal.ai, you find the “Flux LoRA fast training.” If that’s what you want to train, then you add your zip file and the trigger word. Run it.
[0:02:58.400] Download the LoRA, upload your LoRA inside Weavy, plug it into Flux Dev LoRA and run it with that trigger word in the prompt.
[0:03:08.500] Now, just as an afterthought, let’s compare this style training to just regular usage of reference for, let’s say, Nano-Banana, right?
[0:03:19.600] “The entire training set was generated by me over here. So we can manually compare how similar it is.”
[0:03:27.200] “We can see it’s pretty similar, even though I think that purple color is way more dominant in my set than in the results.”
[0:03:35.700] Let’s say we are satisfied with that. Let’s compare it to just using a reference for this product set. So I’ll be importing a few of my images. Let’s go for: One, Two, Three, Four. Let’s say these are a good representation of my style. I’m adding Nano-Banana and adding some references.
[0:03:55.200] Okay, I went for a “Weavy” camera. “Use only the style of the reference images provided.”
[0:04:00.700] Let’s run it a few times. Okay. And we got some nice results. I mean, I’m not sure that he got all the reference because he’s repeating this background a few times. But generally speaking, I would always try using references with Gemini before going for LoRA and see if I’m getting satisfactory results, just because LoRA training needs a bit more labor. And also limited for flux and Wan, etc.
[0:04:27.000] Okay, guys, I would love to hear your thoughts about this. If you have different conclusions than me, I would totally love to hear it. Please share your experience in the comments. See you soon.