How can I get better face details in ComfyUI?

Custom Nodes
1

I use the Wan.2.2 14b q3 k m fp16 fast - video to video workflows.

you can see in the video, the eyes and teeth look low quality. Other body parts such as fingers are also not very detailed and tend to become distorted or blurry during fast motion.

I have already tried a 2x upscale pass, but it didn’t make any noticeable difference.

there any solution for this? Maybe adding one or more nodes that can reconstruct or stabilize the face based on a reference image? Something similar to Face Detailer, Face Restore, or any workflow that works well with Video-to-Video generation.

current settings: Output resolution: 544×960 16–32 frames After rendering, I upscale the final video to 1080p using an offline application. GPU: RTX 4070 12GB RAM: 32GB

I can’t really increase the workflow quality settings much further or render at higher resolutions such as 640p or 720p because I run into VRAM limitations and overload issues.

Has anyone managed to improve the quality of eyes, teeth, faces, and hands while working with similar hardware limitations? Any tested workflows, nodes, or recommendations would be greatly appreciated

3

You are working with too low of a quantization. I have an older GPU than yours by several years and use full FP8 models fine. On my 2080 Ti with 11GB VRAM (I have same amount of 32GB RAM, and a Ryzen 9 3900x) it takes me anywhere from 4 to 6 minutes to generate at 480p, faster with 360p. If you are set on using Quantized models over everything, than I suggest going to Q5 or Q6, although you should just install the full models. Look into custom versions of WAN 2.2, like SmoothMix and DaSiWa. And also, VisoMaster and Topaz Video AI for pre/post processing. Also, if you are running 4 steps for this fast workflow, double it to 8. “Model Patch Torch Settings” / enable FP16 Accumulation via comfyui-kjnodes, or SageAttention.