What the official repo does
--convert_model_dtype converts parameters to config.param_dtype (typically bf16/fp16). Combined with --offload_model True and --t5_cpu it is the supported 80 GB → slightly less path — not 4-bit.
Community compression
- DiffSynth-Studio — FP8, layer-offload, sequence parallel, LoRA.
- LightX2V — step distillation, quantized packs, lightweight VAE, sparse attention.
- FastVideo — distilled Wan with sparse attention.
- Cache-dit — DBCache / TaylorSeer / Cache-CFG on Wan2.2 MoE (skips compute, not bits).
- ComfyUI-GGUF / Kijai WanVideoWrapper — GGUF and kernel-level tricks.
How to choose
Need official sampling quality → bf16 + offload or multi-GPU. Need a 4090 at near-full quality → FP8 (see FP8 vs GGUF). Need 12–16 GB → GGUF or a distilled LightX2V pack, and expect texture loss.
FAQ
- Does quantization add audio?
- No.
- Is cache quantization?
- No. Cache-dit reuses activations across steps. Weights stay full precision unless you also quantize.
Primary sources
Checked against Wan-Video/Wan2.2 README · Aug 28, 2026
