How much visual and motion quality do offloading and 4-bit/FP8 quantization cost for large open video models?

The Diffusers guide shows group offloading to ~13 GB and bitsandbytes 4-bit for 14B Wan, and warns quantization may affect quality per model. LTX, Self-Forcing and LongLive also advertise FP8 speedups. Is there a measured, reproducible quality-versus-memory table (motion coherence, flicker, identity) across models and quantization methods? Why this is still open: Sources state the trade-off exists but quantify memory, not quality loss.