A large graphics card on a workbench with two of its three fans covered in strips of masking tape

Installed PyTorch for CUDA 12.8? ComfyUI Is Skipping Its Fast Kernels.

A 5090 owner got a 20-second MiniMax H3 clip that ran out of memory to fit, just by reinstalling PyTorch for a newer CUDA. That is one person’s result. The reason behind it is in ComfyUI’s own code, it applies to every recent NVIDIA card, and you can check your install in under a minute.

November 14, 2026 · 3 min · Aillex / DIY AI
Real-Time Lip-Sync with MuseTalk on an RTX 5090 (Blackwell Survival Guide)

Real-Time Lip-Sync with MuseTalk on an RTX 5090 (Blackwell Survival Guide)

MuseTalk generates lip-synced video faster than real time on a 5090, but getting it to BUILD on Blackwell is dependency hell. Here’s the exact recipe that works.

July 1, 2026 · 4 min · Aillex / DIY AI
As an Amazon Associate, this site earns from qualifying purchases. Referral links are always disclosed.