This media is not supported in your browser
VIEW IN TELEGRAM
Local alternative to Kling AI 3.0 Motion Control (ComfyUI, 16GB VRAM)
https://redd.it/1v6ibi8
@rStableDiffusion
https://redd.it/1v6ibi8
@rStableDiffusion
I spent a year building a free SDXL & Anima trainer that runs on my 12 GB GPU — here's what came out of it
A little over a year ago I got frustrated trying to fine-tune SDXL on my RTX 3060. Every option either forced lower resolution, locked away important settings behind massive config files, or needed a 24 GB GPU to do anything meaningful.
So I started building my own trainer. That was a mistake. A good mistake, but a mistake.
What followed was several months of failed attempts just to get SDXL stable inside 12 GB, then another six months pulling SDXL apart architecturally to understand why things kept breaking. I went back through the original papers and implementations, rewrote the optimizer approach, and eventually built something I actually wanted to use.
The result is Aozora, a free GUI trainer for SDXL and Anima fine-tuning on consumer GPUs.
Current results on my setup: Aozora can train roughly 80–90% of the full SDXL UNet within 11.8 GB of VRAM at around 1.55 seconds per iteration. It can also train 100% of Anima at 1152×1152 resolution while using approximately 11.4 GB of VRAM at around 2.67 seconds per iteration.
The GUI exposes the controls that actually matter — learning rate curve, timestep distribution, loss weighting, optimizer behavior, layer targeting, and training metrics — without burying you in config files or options that rarely change anything.
It is still beta and has mainly been tested on my own hardware, so expect rough edges. If you hit installation issues let me know and I will sort out compatibility.
GitHub: https://github.com/Hysocs/Aozora\_Trainer
Edit — answering a question I received by DM: Aozora is not a wrapper or frontend for another trainer. The training code is standalone and intentionally kept minimal. with an optional attention backend for better performance. It was created and tested on windows only as of now
https://preview.redd.it/oogx8vsevefh1.png?width=1502&format=png&auto=webp&s=d126541e8f556253d6591dd8727a5dd467be8358
https://preview.redd.it/q5uz16tgvefh1.png?width=1502&format=png&auto=webp&s=f5fb882c465a179f4e90177780bd94b60fbf2e48
https://preview.redd.it/ycczaa1ivefh1.png?width=1502&format=png&auto=webp&s=2ec00324d7b8f895f53c375919a9cc58e741d42c
A training guide is coming soon.
https://redd.it/1v6ej0k
@rStableDiffusion
A little over a year ago I got frustrated trying to fine-tune SDXL on my RTX 3060. Every option either forced lower resolution, locked away important settings behind massive config files, or needed a 24 GB GPU to do anything meaningful.
So I started building my own trainer. That was a mistake. A good mistake, but a mistake.
What followed was several months of failed attempts just to get SDXL stable inside 12 GB, then another six months pulling SDXL apart architecturally to understand why things kept breaking. I went back through the original papers and implementations, rewrote the optimizer approach, and eventually built something I actually wanted to use.
The result is Aozora, a free GUI trainer for SDXL and Anima fine-tuning on consumer GPUs.
Current results on my setup: Aozora can train roughly 80–90% of the full SDXL UNet within 11.8 GB of VRAM at around 1.55 seconds per iteration. It can also train 100% of Anima at 1152×1152 resolution while using approximately 11.4 GB of VRAM at around 2.67 seconds per iteration.
The GUI exposes the controls that actually matter — learning rate curve, timestep distribution, loss weighting, optimizer behavior, layer targeting, and training metrics — without burying you in config files or options that rarely change anything.
It is still beta and has mainly been tested on my own hardware, so expect rough edges. If you hit installation issues let me know and I will sort out compatibility.
GitHub: https://github.com/Hysocs/Aozora\_Trainer
Edit — answering a question I received by DM: Aozora is not a wrapper or frontend for another trainer. The training code is standalone and intentionally kept minimal. with an optional attention backend for better performance. It was created and tested on windows only as of now
https://preview.redd.it/oogx8vsevefh1.png?width=1502&format=png&auto=webp&s=d126541e8f556253d6591dd8727a5dd467be8358
https://preview.redd.it/q5uz16tgvefh1.png?width=1502&format=png&auto=webp&s=f5fb882c465a179f4e90177780bd94b60fbf2e48
https://preview.redd.it/ycczaa1ivefh1.png?width=1502&format=png&auto=webp&s=2ec00324d7b8f895f53c375919a9cc58e741d42c
A training guide is coming soon.
https://redd.it/1v6ej0k
@rStableDiffusion
GitHub
GitHub - Hysocs/Aozora_Trainer: A highly memory-efficient fine-tuning toolkit for SDXL and ANIMA, combining modern VRAM-saving…
A highly memory-efficient fine-tuning toolkit for SDXL and ANIMA, combining modern VRAM-saving techniques with custom optimizers to enable full-quality model training on GPUs with as little as 12 G...