Understanding the Concepts of render_scale and crop_size

[catkitchen721]

Hi, thank you for your wonderful work!

I want to train photos with a resolution of 4K (4943 x 3284), and I hope to truly train and calculate the loss at this resolution. My current GPU is an RTX 4090 with 24GB VRAM.

However, I'm confused about the difference between render_scale and crop_size. I think render_scale is just the scale of the viewer's screen that humans see, and not the resolution used for training. (For example, when render_scale=0.5, the training is still in 4K, but the displayed image is in 2K.) As for crop_size, I don't understand its meaning at all.

When I actually train, if I set render_scale to 1, the VRAM is insufficient, but if I set it to 0.8 or below, there are no problems. But shouldn't the training always be in 4K (only the rendered output is different), or am I misunderstanding something?

In summary, I hope to train a 4K scene, as long as it can be trained successfully. Are there any parameters I can adjust for my RTX 4090?

I would be immensely grateful for any suggestions or guidance.

[catkitchen721]

Hi!
Additionally, these are my TensorBoard results. The only difference in these three experiments is the render_scale, which are 0.5, 0.75, and 0.8, respectively, while the photo resolution is all 4K (4943 x 3284). Why would their metrics (PSNR, SSIM, LPIPS) differ?