Wan-AI / Wan2.1 Video Model (Safetensors) - GGUF Quants - i2v 14B

GGUF Quants of Wan 2.1 for lower VRAM usage.

I suggest to use TeaCache for ~2x speedup.

All samples are 640x640 with TeaCache enabled and a Init-Image.

Source: Source (gguf): https://huggingface.co/city96/Wan2.1-I2V-14B-480P-gguf/tree/main from city96

Here are the results of some testing ...

VRAM and RAM calculations:

• Your VRAM should be 1~2 GB higher as the model size

  e.g. 16GB (VRAM): Model 12GB -2GB = 10 -> good to go

• You should always spare 1-2 GB for you OS to operate functional

• Your RAM should at least be 16, better 32 GB

Hints for faster but reasonable good results I2V (Image to Video):

• Use a low resolution inside the maximum pixel count

• Upscale the complete video after generation with a tool of your liking

• Use 20-30 steps, 20 is good, 30 needs significant more time but preserves details

  • The step increase is linear so, 10 more steps results from e.g. 10 min to 15 min

• Use CGF 4-6 (6 are good most of the time), 4 are sometimes better if lighting goes off

• The image as Init-Image you use should have the same ratio as the video resolution

Good resolutions:

9:16/ 16:9 with 480x832 / 832x480 = 480p

3:4 / 4:3 with 480x640 / 640x480 = 480p