Some data, for those interested, and because I'm in that kind of mood:
I am using a 5060ti (16GB VRAM) with 64GB RAM.
Original image was made using SDXL @ 1024x1024. I have a prompt with random elements revolving around these Sci-Fairy girls doing nasty things with combinations of robots, tentacles, etc. I'll run a batch of 4, 8, 16, whatever, which takes about 20 seconds each once it gets cooking, see if I get one I think I can make work with the loras I've found, and then move over to a Wan 480p img2vid workflow I've been slowly optimizing.
Both the 480p and 720p were run with the same fp8 model, 81 frames, 16fps, 20 steps, 6 CFG, 1 denoise. Blockswapped as needed to get it to work, TeaCache on both, same seed. Exactly the same workflow other than changing resolution and Blockswap numbers. Blockswap was optimized to hit roughly 95% VRAM in both cases.
1024p was just an upscale/interpolate workflow I found floating around.
>>28968261 512x512 - 812s, or 13.5 minutes.
>>28968415 720x720 - 1921s, or 32 minutes.
>>28968462 1024x1024 - 286s, or 4.75 minutes.
Because I keep spreadsheets of all of this:
480x480 without TeaCache typically runs me about 15 minutes.
512x512 without TeaCache ~19 minutes.
720x720 without TeaCache ~47 minutes.
So TeaCache is roughly saving me about 40% of my time. I just started using it, so not sure how happy I'll be with the results long term.
I was not happy with CausVid, btw. It just kind of made things vibrate in place.
>I was not really happy with the results of running the same seed again at higher resolution>I am fairly happy with the result of the upscale/interpolate>the time for a lower res run + an upscale were about 75% better to get even more res and those sick fps gainsThus, going forward, my workflow will probably be to TeaCache multi-batches of random-seed 480p or 512p videos until I hit one I like then upscale/interpolate.
I think that's everything. Storytime over.
