Disclaimer: I used AI to format this to be as concise as possible.
I’ve been building LoRA Character Kit around a simple idea that training a character LoRA should be more approachable for people using the computer they already own. You shouldn’t have to start by buying a flagship GPU, renting an expensive server, or becoming an expert in Python dependencies just to find out whether character training is for you.
My development machine has an RTX 4080 with 16 GB of VRAM and 32 GB of system RAM. That’s certainly capable hardware, it’s my gaming PC repurposed for training a character LoRA, but obviously not a workstation with multiple GPUs or a huge pool of video memory. Making training practical within those limits has been a central part of this project.
What is LoRA Character Kit?
The kit provides a browser-based workspace for the character-training process:
Import and organize your training photos.
Generate caption drafts, review them, and keep your trigger phrase consistent.
Remove unwanted images and restore them if you change your mind.
Prepare a dataset revision for training.
Create a training job and follow its progress and logs.
Compare saved checkpoints.
Download an adapter or copy it directly into ComfyUI’s LoRA folder on the same server.
The goal is to make the workflow easier to follow—from choosing the right photos to testing the resulting character—while keeping the important choices visible. It builds on existing tools, including Ostris AI Toolkit and ComfyUI. Those projects do the underlying training and generation work; the kit brings the steps together into a more manageable character-focused workflow.
Working within 16 GB of VRAM
The kit currently includes training profiles for FLUX.1 Schnell, FLUX.1 Dev, and Krea 2 RAW.
For Krea 2, the approach is to train an adapter on RAW and use it with Turbo for generation. My tested configuration uses a 4-bit frozen base, 512px training resolution, batch size 1, and rank 16, along with memory-saving techniques such as gradient checkpointing and cached text embeddings.
A download people can actually use
I intend to make LoRA Character Kit available as a downloadable early preview that people can install and use, with setup instructions and clearly identified requirements. The current development target is Linux with an NVIDIA GPU. I’m not claiming support for every operating system or graphics card yet.
The downloadable kit will also need to distinguish its own code from the models and tools it depends on. Model weights have their own download requirements and licenses; they won’t simply become freely redistributable because they work with the kit. My aim is a practical starting point: understandable instructions, useful defaults, clear errors, and an honest account of what has been tested.
I want your feedback before calling it finished
This is the part I’m most interested in: getting the kit into other people’s hands before deciding what the final product should look like.
When the preview is available, I’d love for people to download it, try a small dataset, and tell me what happens.
Useful feedback would include:
Your hardware: GPU, VRAM, system RAM, and operating system.
Installation: What worked, what failed, and what was unclear.
The workflow: Where you felt confident and where you got stuck.
Training: Whether a short test completed, how long it took, and how much memory it used.
Character results: Whether likeness carries across poses, clothes, lighting, and backgrounds.
Missing features: What would make the biggest practical difference to you.
You don’t need to be an experienced trainer to contribute. If a button is confusing or a setup instruction assumes knowledge you don’t have, that is valuable feedback too. Please remove tokens, account details, and other private information before sharing logs. There’s no need to publish your training photos to report an installation or workflow problem.
Help shape the kit
I’m building this because I want more people to be able to experiment with character LoRAs without assuming that expensive hardware is the price of admission. There are still real hardware limits, and the kit won’t make every model fit every GPU. But the results so far suggest that a carefully configured workflow can make useful training possible on a consumer machine.
If that sounds useful, I’d love for you to try the preview when it’s available and help me improve it. Tell me what works, what frustrates you, and what would make you want to keep using it.
The goal is to learn from real users and real machines before calling anything a final product.
Here is the GitHub url: https://github.com/TigerFan69/lora-studio
Here is my Ko-Fi page should you wish to help support further development and character LoRA creation: https://ko-fi.com/tigerfan69
Comments