LTX Director and Sulphur: ComfyUI setup for 8GB VRAM

This tutorial covers local AI video generation with LTX Director in ComfyUI, using text prompts and multiple reference images. The speaker tests Sulphur FP8 and LTX 2.3 Distilled 1.1 on an RTX 5060 with 8GB VRAM and 32GB system RAM.

Setup starts with ComfyUI nightly and updated nodes. The speaker distinguishes native FP8 support on Ada Lovelace and Blackwell GPUs from use on older architectures, recommending GGUF for older cards with 8GB or 12GB VRAM and limited system RAM. These are recommendations for the demonstrated setup, rather than universal performance guarantees.

The adapted workflow adds a chunk feed-forward node and removes the second rendering pass. The speaker says this resolved end-of-clip artifacts with multiple reference images. Timeline text blocks guide scene changes, while frame rate and duration determine the frame count. For 8GB cards, the suggested limits are eight seconds at 1080p, 10 to 12 seconds at 1600 × 900, and 720p for longer clips.

The speaker prefers Euler Ancestral for stability and lip sync, and LCM for dynamic movement. Hunyuan-Foley supplies background sound; Google AI Studio and DaVinci Resolve handle voiceover and editing. Limitations remain: the speaker reports morphing in wide shots and fast action, repeated attempts for usable takes, and a steep learning curve. The closing discussion proposes combining local generation with cloud models for more demanding scenes.