🖼️Loradex Trainer Program
Gotta train 'em all!
Last updated
Gotta train 'em all!
Last updated
The team here at Glif is all about supporting AI innovation and creativity. This limited program is a way for us to help AI creators by giving them easy access to the tools they need to create their first Low-Rank Adaptation (LoRA) model.
Assemble up to 10 inspo images for training.
Read the rules below.
Submit the images with an intro to the #loradex-applicationchannel in our Discord server, along with your Glif profile link.
Aspiring LoRA trainers eager to level up their Flux Dev skills (no experience needed!).
Creative minds bursting with ideas for unique styles and models.
Open-source diehards ready to share their creations with the world.
Team players who thrive on collaborative growth and friendly competition.
Our team will send you details on how to get started after your application is approved!
You can also find these rules in the #rules channel in the Glif Loradex Trainer section in our Discord server.
Keep all training community-friendly. No NSFW material allowed.
Ideally, train on your own art or on public domain resources (Artvee.com has a ton of amazing styles that are public domain).
No training on clearly protected IP (sorry, Disney fans!). When sourcing your training materials, be mindful of copyright concerns:
Avoid training on collections of images from a single artist or franchise.
Diverse image sets from various sources are preferred.
Consider whether your model might generate infringing outputs.
Focus on innovation: Aim to create unique styles rather than replicating existing ones.
Do not reproduce real people: These models will be posted publicly, and we're not ok with real faces being attached, consenting or otherwise.
Image recommendations
We recommend images that are:
Diverse in content
Consistent in style
Captioned simply
In our tests, we found that using a known generic token that relates to the style works best, such as “vintage anime style.”
Captions
Keep your captions short and sweet, describing details in the images such as the scene or characters portrayed. For style training, avoid adding extra words to describe aesthetic elements, unless you plan to use them in prompts later.
Use ‘[trigger]’ to insert the trigger word in your caption.
Start with the presets!
We've developed some stellar presets to get you started! Before adjusting numerous settings, begin with hyperparameters that we're familiar with. This approach makes it easier to troubleshoot if you encounter any issues.
Sometimes 1000 Steps is Enough!
Depending on what you are trying to train, you may find that 1000-1500 training steps are perfect for you. It is often easier to tell if a model is undertrained than if it is overtrained, so start with a shorter training session and try again with 500-1000 more steps if your LoRA doesn't seem ready.
Don't Go Crazy with the Dataset
Sometimes less is more when training Flux Dev. Try sticking to 20-40 training images to start out. It's possible to train with as few as 5 or 6. We find that the training can be a little trickier under 10 images or over 50.
Success is Relative
Ease of training depends a lot on what you are trying to achieve. If you are trying to train something really unique that the Flux model doesn't have much context for, it will take a lot longer than if you train something more generic or iconic. You can usually tell after 1000 steps how quickly the style will learn!
Synthetic vs. Human Created Data
In general, training is much faster when your training images were generated with AI. However, if your AI images have hallucinations in them, the model will learn that as well, and sometimes this leads to poorer results. Human-made data often takes longer to learn, but typically has much cleaner and crisper details, and preserves small details better. You can also mix synthetic and human data for cool results!
Example Images
Once your training has been submitted, it will automatically generate example images in increments of 500 in the #live-training Discord channel.
A: If you are accepted into the Loradex Trainer Program, you will receive credentials and a link to our private tool.
A: Yes, your models will be shared open-source on our Huggingface page. Please keep this in mind when you are preparing your models.
A: All stored LoRAs are publicly accessible on our HuggingFace account: https://huggingface.co/glif-loradex-trainer Creators can download their LoRAs and use them on other platforms, including but not limited to Glif!
A: You are recognized as the creator of your LoRA in highlights of the LoRA. For commercial use, Flux Dev LoRAs may inherit NC licensing via the Dev License unless you have a license with BFL. All LoRAs published during early access are under CC0 licensing (no rights reserved).
A: We have a handy guide that covers this in more detail on our company blog!
A: Chat with us in the #discussion channel in the Glif Loradex Trainer section of our Discord server!