Heya! Okay, this is actually more of an actual blog post. I've been experimenting with the H3 Minimax workflow for video generation, and I wanted to share some of my findings and thoughts on it.
Well... This will also feature some funny failed stuff that I don't share too often, so I guess it's a bit of a 'behind the scenes' post as well. Also, this one will have some ecchi and slight nudity in it, grow up!
For starters, I tried to generate a video of Kaori recording herself for Sakaki (you might have gotten this by now if you actually followed my guide for video and read the prompt I left in the workflow file).
Also, I used the following images as reference:
For my first test, I tried using most speed optimizations I could, including lower resolution, steps, turbo lora, etc... I was surprised that I was able to generate a video in about 10 minutes! But uh... well to be quite frank, it sucks ass, just look at it:
Still! Doing it this way is very useful for testing out prompts and seeing how the model reacts to them. Because actually doing it at full quality takes a lot of time with my current setup (RTX3060, 12GB VRAM, 16GB RAM). Like... 2 hours long!
(also, yeah, these have sound! Most of my gens do nowadays but I hate them so much I almost never include them.)
For my next tests, I adjusted the workflow to make it able to include more reference images (ok I lied, I asked gpt to do it for me because I'm dumb and lazy as shit), but I forgot to properly adjust the prompts for it and I would lose an entire day generating more bad clips.
Here is the reference image that I included:
I adjusted the settings trying several levels of quality but most of them were giving me OOM errors at the VERY END... AFTER 2 HOURS OF GENERATION (keep in mind that my pc is taken hostage when it's generating videos, almost impossible to use it for anything else). I was so mad, I just wanted to cry... Fortunately, I found out it was because of the upscaling node. Bypassing it made me able to generate... these:
Hahaha, okay, so these suck as well and it took me almost 2 hours each... there goes my free time of the day I guess, but hey! They're getting better! Just a matter of finding out the right settings and prompts! Well...
I ended up wasting a bunch of time doing just that! But... Well, the results are definitely high quality, but the crushing reality of AI generated content is that it's also VERY random and it can introduce weird stuff by sheer bad luck (which I have in spades). I will let the video speak for itself:
Well, not all is lost though, I managed to make a cute small clip from the last few seconds of the video:
Also, it makes for a nice cutesy loop:
Well, I guess that's it for now! I will keep experimenting with this workflow and see if I can get better results. I will also TRY to make a proper video of Kaori recording herself for Sakaki, but that will take a lot of time and patience. Stay tuned!