How it's going gguffers, and yes we are going to use a video model for generating images. So I should to repeat it again - this is text to image workflow. Why? Because suddenly WAN is capable of making very cool photorealistic images. Also, this workflow is different from others because it's very simple and uses minimal plugins (modded nodes). Now let's get these goofy GGUFs and make things work.
Requirements:
16GB of VRAM (or 12GB, if you will choose lower quants) and 32GB of RAM
Installation:
Download model files:
Main model - Wan2.2-T2V-A14B-LowNoise-Q6_K (drop into ComfyUI\models\unet). Options: choose Q6_K for 16GB VRAM or Q5_K_S for 12GB. Q8 will fit into 24GB.
Text encoder model - umt5-xxl-encoder-Q6_K (place into ComfyUI\models\text_encoders). Options: choose Q6_K or something.
VAE - Wan2.1_VAE (throw into ComfyUI\models\vae)
Lora - Wan2.2-Lightning_T2V-v1.1-A14B-4steps-lora_LOW_fp16 (move into ComfyUI\models\lora). Attention attention: download the correct model, don't mix up low noise and high noise and t2v with i2v. Over.
Download and open this workflow in ComfyUI.
Go to "Manager" - "Custom Nodes Manager" and install "ComfyUI-GGUF" v1.1.3 or above (older versions can blow an error "Unexpected text model architecture type"). Restart the ComfyUI.
Usage:
Choose some resolution: 720p and 1080p (1088) are ok.
Choose number of steps and a sampler:
10-12 steps with 1.0-1.3 cfg is recommended. You can increase the steps up to 20 for more contrast and hallucinations removal. Samplers: euler, euler_cfg_pp, dpmpp_2m, dpmpp_sde, dpmpp_2m_sde, dpmpp_2s_a; scheduler: karras or expotential.
20-30 or more steps with ~3.5 cfg if you bypass Lora; euler simple, karras or something.
If you experience crashes and disconnects, try using lower quants.
If you have any other errors during GGUF node installation, do a clean install of ComfyUI and Manager and repeat.
Things tested on ComfyUI Windows portable edition v0.3.49 with 32RAM and 16VRAM 5060Ti.