Wan 2.2 I2V — Dual GGUF Cascade (4 Steps, VRAM Friendly)

Fast Wan 2.2 Image-to-Video workflow using a dual GGUF cascade strategy: the first 2 sampling steps run on the q5_High quantization for strong structural quality, then hands off to q5_Low for the final 2 steps — giving you speed without sacrificing fidelity.

What you get:

• 4 total steps (2+2 split: q5High then q5Low)

• - Portrait 480x832 @ 49 frames / 16fps

• - Runs on VRAM-constrained GPUs (GGUF keeps memory low)

• - UMT5 XXL fp8 text encoder + Wan 2.1 VAE

• - euler sampler, CFG 1

Models needed:

• DasiwaWAN22I2V14BBoundbiteV10_q5High.gguf

• - DasiwaWAN22I2V14BBoundbiteV10_q5Low.gguf

• - umt5_xxl_fp8_e4m3fn_scaled.safetensors

• - wan_2.1_vae.safetensors

How it works:

The trick is using the higher-fidelity quantization for the initial denoising pass (where structure and composition are established), then letting the faster q5_Low carry it home. Near-q5High quality at closer to q5_Low speed.

Drop your start image into LoadImage, write your prompt, and queue.