Today, I ran a few tests using different video resolutions and step counts. The conclusion is that for 1024x1024 video at 24 FPS, it’s best to use 20 steps. Sure, generation takes 12 minutes, but the quality turns out absolutely stunning!
Video specifications and generation times: Video 1 – 1024x1024, 24 FPS, 20 steps (12 minutes) Video 2 – 800x800, 24 FPS, 20 steps (6 minutes) Video 3 – 1024x1024, 24 FPS, 10 steps (6 minutes)
If you also draw or generate my Neka, please add these tags: neka, neka_dream, nekadream, neka_(nekadream), nika, nika_(nekadream) So that I could admire her too 😻😻😻
A fun detail I've noticed with H3- higher res improves instruction following. And step count definitely needs to stay around 20 if you want high quality, like you saw here. But if you go down the path of madness I just did and run the reference model using timestamps to control actions (a good way to lose most of a week to one video of questionable quality), I lower the res down to .1 or .2 MP and set step count to 8 to rough in the prompts- it will have severe artifacting, but it can let you see if the prompt was understood at all in a fraction of the time. That way you can have several rounds of fixing the most glaring issues before you commit to those really long gens.
Wholesome and cute tests- I like it. A fun detail I've noticed with H3- higher res improves instruc
That’s exactly what I do—I also set it to 0.1 or 0.2 MP to check how well the AI understands the prompt, and only then do I start the full video generation. I also noticed that higher quality follows the prompt better
Thanks! 🥰 That’s exactly what I do—I also set it to 0.1 or 0.2 MP to check how well the AI unders
![[2 🎞️] Neka (Nika) Dream 😻](https://sf.ib.metapix.net/thumbnails/medium/6169/6169323_NekaDream_minimax-2026-09-22_13-11-47_00001-audio.jpg "[2 🎞️] Neka (Nika) Dream 😻")
![[🎞️] Neka & Bunny 😻](https://sf.ib.metapix.net/thumbnails/medium/6174/6174461_NekaDream_minimax-hq-2026-09-27_16-13-07_00001-audio.jpg "[🎞️] Neka & Bunny 😻")