It is now possible to generate 16 Megapixel (4096x4096) raw images with SANA 4K model using under 8GB VRAM, 4 Megapixel (2048x2048) images using under 6GB VRAM, and 1 Megapixel (1024x1024) images using under 4GB VRAM thanks to new optimizations
Added 2025-01-12 23:30:15 +0000 UTC13 January 2024 Update
Installers : https://www.patreon.com/posts/from-nvidia-labs-116474081
New 4K Tutorial Video : https://youtu.be/GjENQfHF4W8
Now the APP will use Diffusers Pipeline and it has huge VRAM optimizations
You need to reinstall
The models will be downloaded into your Hugging Face cache folder when you first time generate something
Please make a fresh install
When you enable all 4 optimizations the VRAM usages are like below
Make sure shared VRAM is enabled because initial loading of the model need more VRAM
Enable VAE Tiling + Enable VAE Slicing + Enable Model CPU Offload +
Enable Sequential CPU Offload
1K (1024x1024) : 4 GB GPUs
2K (2048x2048) : 6 GB GPUs
4K (4096x4096) : 8 GB GPUs
Still in any case may work on your GPU test it
Just Enable VAE Tiling + Enable Model CPU Offload works great in many cases
Check above images
Comments
i may make with OneTrainer if it gets fully supported
Furkan Gözükara
2025-01-13 18:49:46 +0000 UTCPlease make a tutorial on how to train this
Samuel
2025-01-13 17:24:45 +0000 UTCthey have training scripts but i havent tried yet
Furkan Gözükara
2025-01-13 09:02:55 +0000 UTCIs there a way to finetune the model already?
Bjarne Desmet
2025-01-13 08:58:42 +0000 UTCyes it is way better right now. thanks for comment
Furkan Gözükara
2025-01-13 08:42:20 +0000 UTCV14 is a big improvement over the V9 release. So much faster and doesn't run out of memory on my 3070TI with 8 GB VRAM.
Nibmeister
2025-01-13 01:39:13 +0000 UTC
