RTX 3060 12GB for Local Stable DIffusion (AI Image generation)

Year ago, I bought a PALIT RTX 3060 12GB to replace my old GTX 1650 that took by my brother. I choose RTX 3060 12GB is just because the VRAM is much higher than my preferred GPU (RTX 4060). Then, the times i try local AI image generation is come, yes, i tried Stable diffusion.

Peformance of image generation

Well i didnt do a benchmark for this, but i used some popular checkpoint to create image generation. It worked and i think it is in a OK-ish result when generate a 512x512 pixels with a default sampling steps and CFG scale. It is take my 6GB of VRAM to generate that.

For the larger images started with 720x720 pixels, it will take almost of the GPU VRAM, it takes a more time by alot (around 15-20 seconds per image with default sampling steps and cfg scale). This become a horrible results in my opinion.

But, Should be 20seconds is okay?

No. Generally we dont use the first prompt of the image. It is often generate a bad results or just want to try alternative results, so rewriting the prompt is a common thing. 20 seconds to generate first image, 10 seconds to edit the prompt and waiting 20 seconds more to generate other result.

I think, the maximum resolution that can be guaranteed to be generated is around 900x900 pixels with default cfg scale and sampling steps. It can be higher, but this might be throw an out-of-memory error due to limited vram.

Mov2Mov

I also tried with video, and its even worst. I'm using mov2mov with reACTOR plugin, it tooks around 25 minutes for 720p video with 25fps, and it took around 40-50 minutes for 1080p video with 25fps.

The FPS of the video is time multiplier here, if you have 30 or 60 fps, then it will take much-much longer.

All that test is for 10 seconds video! So, this GPU is not recommended using reACTOR, probably will much better for others plugin if any, even though i'm pessimist about this.

Surprisingly, the temperature is not that hot, since I saw the GPU usage is like a sinewave for every frame. Its stable around 60 celcius degree.

Training LoRA using koyhass

I also training my LoRA model using koyhass (which is i think it is much faster that built in trainig modules in stable diffusion). Dataset is the key, since LoRA doesnt need big dataset, but we need variatif and well weighted database (not dominant of some style of dataset). It took around 50 minutes for 40 repeat, 4 epochs, with 36 dataset (512x512 px images), the rest you can see at the chart.

The temperature is around 76-80 celcius degree, this will depends on GPU brand you use

Conclusion

RTX 3060 12GB (Palit) is ok-ish to learn/use simple version of stable diffusion. I bold the use word because this not really greate for creations. If you really plan to do an AI thing with this GPU, my suggestion is to skip this, bougth an RTX 4000 series with at least 16GB of VRAM.