Need Advice: Local LTX Q4/Q8 Workflow + Cloud Final Rendering

A Reddit user, No-Train-5892, is seeking advice on a serious hybrid workflow combining local and cloud video generation using open-source LTX models in ComfyUI. Their planned laptop setup is an MSI Vector 16 HX AI with an NVIDIA GeForce RTX 5090 Laptop GPU (24GB VRAM), Intel Core Ultra 9 275HX, 64GB system RAM, and a 1TB SSD. The key idea: run quantized Q4 or Q8 models locally for low-resolution previews (240–360p, ~10 seconds, 24–25 fps, 8–12 steps) to iterate quickly, then use the same model, workflow, and seed on cloud hardware for final renders at higher resolution and more steps (30–40+). This approach aims to keep previews visually close to the final output while drastically reducing cloud costs.

The critical element of the workflow is VRAM management. No-Train-5892 plans to use sequential execution—only one heavy model (text encoder, video model, or VAE) stays in VRAM at a time, with inactive models offloaded to the 64GB system RAM. This ensures the 24GB VRAM is reserved solely for active computation. The user asks whether this architecture is stable long-term on laptop hardware, especially for aggressive offloading in long ComfyUI sessions. They also seek recommendations on whether to use base Q4, base Q8, or distilled Q4/Q8 quantizations, and what preview resolution and step range work best for fast iteration. Their priority is workflow stability, predictable previews, and efficient iteration over raw rendering speed.