Workflows

Fix BLURRY AI Videos Wan 2.1 Video to Video ComfyUI Workflow

0
Please log in or register to do it.

So yeah, last week I tried generating a video of a couple dancing—looked great until their faces turned into blobs and the guy’s hand decided to morph into a claw. Not exactly what I was going for. After way too many failed attempts, I finally figured out how to clean up those messy AI-generated videos using WAN 2.1 in ComfyUI. Here’s what actually worked.

The Models You’ll Need

First, grab these files—no fluff, just the essentials:

  • WAN 2.1 TV2 1.3B BF16 Model → Drop this in ComfyUI/models/diffusion. BF16 runs smoother on mid-range GPUs like my RTX 3070 Ti. → Download
  • VAE File → Goes in ComfyUI/models/vae. Skip this, and your colors end up looking like a faded Instagram post. → Download
  • UM5T5 XXL FP8 Text Encoder → Toss it in text_encoder. Fixes blurry text and random artifacts in the video. → Download
  • LORA File → Save it in the Loras folder. → Download(

Settings That Actually Work

I ran tests on a 12GB GPU, and these settings gave the best balance:

  • K Sampler:
    • Steps: 40 (anything lower gets blurry, higher is overkill)
    • CFG: 6 (keeps details stable without going nuts)
    • Denoise: 0.4 (start here—anything above 0.6 starts erasing motion)

Skip Layer Guidance: The Fix for Melted Faces

Skip Layer Guidance (SLG) is like a precision tool for broken video parts. Here’s how it saved my dance clip:

  • Blocks 9 & 10 are the sweet spots—other layers just added flickering.
  • Start=0.2, End=0.8 keeps transitions smooth.
  • In my test, a character’s face turned into a puddle at the 3-second mark. SLG sharpened the eyes and lips without messing up the movement.

Tile LORA: The Detail Booster

Tile LORA isn’t for fixing disasters—it’s for polishing videos that are almost there.

  • Use it for:
    • Skin textures, fabric wrinkles, or adding depth to flat lighting.
  • Avoid if: Faces or hands are completely glitched—it won’t rebuild missing parts.
  • Workflow tip: Run SLG first, then apply Tile LORA with denoise=0.3 (higher values wipe out details).

I didn’t expect this part, but combining both gave me way cleaner results than starting from scratch. If you’re dealing with blurry AI videos, this setup might save you a ton of time.

Download Workflows
🤖 Hey, welcome! Thanks for visiting comfyuiblog.com
ComfyUI WAN 2.1 Depth Lora Workflow
LTX Video 0.95 Multiple Frame ComfyUI Workflow

Your email address will not be published. Required fields are marked *