Summary
Fahd Mirza walks through the full installation and first live generations of LTX-2.5, the latest video model from Lightricks, running inside ComfyUI on local hardware. LTX-2.5 is notable for generating video and synchronized audio in a single forward pass — no separate audio pipeline — and supports image-conditioned generation, multi-shot sequences that maintain consistent character and lighting across cuts, and a 22-billion-parameter distilled checkpoint designed for faster output at a modest quality trade-off.
The tutorial covers every step of the setup: installing the required custom ComfyUI node via terminal, accepting the gated-model terms on Hugging Face, and placing checkpoint files into the correct ComfyUI subdirectories for diffusion models, text encoders, and the variational autoencoder. Mirza uses the official example workflow and runs two live generations — a trampoline scene and a traffic-light sequence — showing results without editing, including visible artifacts: reversed car motion, incorrect traffic light states, and imperfect human movement despite solid audio sync.
On hardware, Mirza measures 66 GB VRAM consumption in full precision, recommending a single A100 or H100 for reliable performance. Quantized variants are available for systems with 30 to 48 GB of VRAM. He also points viewers toward Mass Compute for cloud GPU rental. For anyone wanting to run local multimodal video generation today, this is a no-hype practical reference covering the install, the gotchas, and what real first-generation output actually looks like.
📺 Source: Fahd Mirza · Published August 16, 2026
🏷️ Format: Tutorial Demo
