This is the in-progress version of me tweaking the default ComfyUI workflow to my liking, benchmarking different stuff along the way. Right now pretty much the only difference I can think of is that I added the ModelAttentionBackend node.

Model used was minimax_h3_ref2va_pruned_int8convrot.safetensors at 20 steps, with Comfy Kitchen Attention enabled.

I wanna really emphasize that this is NOT an optimized workflow, even for my own hardware. I am actively going through and tweaking and benchmarking different stuff, and it is very much in an "in-progress" state.

Also want to say that the prompt was generated by a Language Model, and it most certainly needs a proper human review... that's just not what I'm focused on for right now.

Reference video and other assets can be found at:
https://github.com/CowOfSteel/video_generation_workshop/

Other Tools Used:

Blender 5.2.1 + the official Blender MCP (OpenAI's Luna with Extra High Attention)
ComfyUI + MiniMax H3
DaVinci Resolve 20
Local hardware: RTX 3090 Ti, 64 GB DDR5 RAM