spacepxl

~If anyone wants to contribute, that would be welcome. I didn't intend to develop this much further, it was mainly just a proof of concept for some colleagues that I...

I added Load EXR and reworked Save EXR to be (hopefully) much more robust and useful. No support for multi-layer beyond RGBA, because OpenCV's EXR support is extremely limited, but...

Double check that your EXR is actually RGB or RGBA. That error message seems to imply that it's only loading a single channel, which could mean that the render data...

Should be fixed now with https://github.com/spacepxl/ComfyUI-HQ-Image-Save/commit/71e242c746e9170e53e00919b12625317a7571e5

#13 is merged now, check that you actually have opencv installed, hopefully manager should ensure that it installs now, or you can do it manually: 

I agree that it should be multiple deterministic timesteps instead of fixed at a single timestep, but the issue was that the dataloader didn't have an option to disable shuffling,...

I think all of those options are valid, because the goal is just to measure relative change. Sampling timesteps mostly from the middle of the schedule will show the largest...

> It would be interesting to see this correlation graph for a timestep shifted model, because there we train mostly higher timesteps but still validate at 500: The loss/timestep correlation...

@dxqbYD If you want to get into picking individual timesteps or a weighted average, you would also want to consider the content you're training on. Style is mostly high frequency,...

I think choosing multiple timesteps is fine, and reporting them separately is fine, but you should absolutely still report the average as well. For larger datasets I still think it...