[Creative Diary] Testing MiniMax H3

Regarding the issue where ComfyUI wouldn't run and I was getting frustrated:

After trying various things, it finally started up once I set the ID to only use the GeForce RTX 5060 Ti at startup. Perhaps the RTX 3050 was indeed acting up. Or maybe it's because after I missed out on installing the V100, the RTX somehow became GPU ID=2 in Windows (0: RTX 5060 Ti, 1: Ryzen 5700G integrated graphics), or it could even be the motherboard's PCIe slot... I'm not entirely sure and haven't been able to isolate the problem, so I'll just proceed as is. It's a bit painful not being able to use the 8GB of VRAM on the 3050, but it can't be helped. LM Studio, on the other hand, runs the 3050 just fine using Vulkan...

Anyway, by referring to the articles by these people, I was able to get the new video generation AI, MiniMax H3, running as well:

Many thanks to Mayuhira-a and Kongou-jun!

I've compiled some results created from image references or text using the standard ComfyUI templates:

The prompts vary, from those I typed directly in Japanese (the first video) to those created for me by Gemini. Overall, it looks cleaner than LTX-2.3 even at low resolutions, and the audio is excellent. It can also speak in Japanese. However, on my setup, the generation time takes many times longer, and it seems impossible to create dozens of seconds at once (it's originally 15 seconds max, but 10 seconds seems to be the limit on my machine). Still, even though it takes time, it seems better to upscale at a 0.4 ratio (half HD) rather than creating a low-resolution one from scratch and upscaling it. This time, I used an app called REAL Video Enhancer to upscale everything to full HD or higher, and then converted that to full HD. It seems great for making short clips of about 10 seconds!

Regarding reducing generation time, I also tried Turbo LoRA, but for some reason, the Patch Sol-Attn node wouldn't load (even after applying it as 'installed,' it remained grayed out). When I ran it without that, the visuals were fine, but the audio distortion was severe, so it felt unusable with sound.

Since it just came out, I'll try experimenting more once it's been optimized a bit further! See you later!