Skip to main content

wan_2.1_vae.safetensors

Welcome to the documentation for wan_2.1_vae.safetensors, a finely tuned VAE (Variational Autoencoder) model tailored for use with Stable Diffusion 2.1 checkpoints in ComfyUI. If you’ve ever had your generations come out blurry, desaturated, or looking like they were filtered through a fog machine, this VAE might just be your secret sauce.

🔧 File Format​

  • Filename: wan_2.1_vae.safetensors
  • Format: .safetensors
  • Serialization: safetensors is a secure and fast model serialization format designed to prevent pickle-based exploits.
  • Precision: Likely in fp16 or fp32 depending on your source — most .safetensors VAE files default to fp16 for lower VRAM usage.
  • Size: Usually between 300MB–500MB (depending on the optimizer and precision used)

✅ Pro Tip: Never rename .safetensors files to .pt or .ckpt. They're not the same thing and your workflow will break harder than a dollar store tripod.

📁 Function in ComfyUI Workflows​

In a ComfyUI workflow, the VAE decodes the latent outputs of the Stable Diffusion model into full-resolution images. Without it, your precious diffusion process would result in ugly latent garbage no one wants to look at (trust me, we’ve all been there).

Where it's used:​

  • Attached to the Load Checkpoint node OR the standalone Load VAE node.
  • Works in tandem with the VAE Decode node to produce final images from latent tensors.
  • Required to visualize any latent output in a human-readable format (aka an actual image).

🧠 Technical Details​

wan_2.1_vae.safetensors is a VAE fine-tuned specifically for Stable Diffusion v2.1. It's a significant visual quality upgrade over the default vae-ft-mse-840000-ema VAE that ships with most SD 2.1 checkpoints.

Features:​

  • Trained on higher-resolution data: Better detail preservation during decoding.
  • Higher dynamic range: Less crushed blacks and more color fidelity.
  • Minimized over-blurring: Avoids the overly smoothed look that some other VAEs introduce.
  • Better for photorealism and fine details.
  • Improved latent space alignment with 2.1-style models.

Architecture:​

  • Same VAE backbone as default SD VAE.
  • Trained with improved reconstruction loss and perceptual tuning.

✅ Benefits​

  • Sharper outputs: Crisper image decoding from latent to RGB.
  • More accurate color mapping: Color tones and gradients appear closer to prompt intent.
  • Preserves fine details: Textures and tiny facial features won't turn into mush.
  • Drop-in compatible: Works in all SD 2.1-compatible pipelines with no special finagling.
  • Low VRAM footprint: Suitable for mid-range GPUs (8GB+ VRAM).

⚙️ Usage Tips​

  • Use with SD 2.1 base or derivative checkpoints like realisticVision, revAnimated_v2Rebirth, or dreamshaper_8 (if adapted to 2.1).
  • Use in tandem with high CFG values (7–12) and moderate Denoise (~0.5–0.7) for best results.
  • To test output changes, toggle VAEs mid-workflow. ComfyUI makes this easy with node swapping.
  • Pair with Ultimate SD Upscale for super-detailed renders at high resolution.
  • Use KSampler with samplers like DPM++ 2M Karras for refined quality when paired with this VAE.

🎯 Ideal batch size? 1–2 for mid-tier cards; go nuts if you’re rocking an RTX 4090.

🎯 Which Model Types This Works Best For​

✅ Recommended Models⚠️ Not Recommended Models
stable-diffusion-v2.1stable-diffusion-v1.5
realisticVision-v5.1anything-v4, anime models
revAnimated_v2RebirthpastelMix, cetusMix
dreamshaper_8 (2.1)darkSushiMix, chilloutMix
analog-diffusion-2.1Any SDXL model (use SDXL VAE!)

TL;DR: This VAE speaks fluent 2.1. Don’t toss it into a 1.5 party — they’ll just argue all night.

🛠️ Setup Instructions​

  1. Download the VAE:

    • You can find it on CivitAI, HuggingFace, or your favorite model repo.

  2. Place the file in the ComfyUI models folder:

    bash

    CopyEdit

    ComfyUI/models/vae/

  3. Load it via one of two methods:

    • Standalone Load VAE node:
      • Add it manually and connect it to VAE Decode
    • Or embed it inside your Load Checkpoint node:
      • Select it under the vae dropdown (only available in updated versions)

  4. Hit run. If you're seeing artifacts, blurry textures, or washed-out color, check your sampler, VAE match, and resolution.

🔥 What-Not-To-Do-Unless-You-Want-a-Fire​

  • ❌ Don't use with SDXL models. This isn’t their VAE. You’ll get garbage-tier outputs.
  • ❌ Don't use with anime checkpoints. Unless you want your anime to look like watercolor soup.
  • ❌ Don't rename the file to .pt or .ckpt. Just... no.
  • ❌ Don't skip decoding. Forgetting the VAE Decode node = zero image output. Just raw tensors. Not helpful.
  • ❌ Don’t pair with the wrong latent resolution. This VAE is meant for 512x512 base latent shapes, as used in SD 2.1. If you’re upscaling from Empty Latent Image, use 64x64 latent with 8x scaling unless otherwise modified.

📚 Additional Resources​

📎 Example Node Configuration​

{
  "id": 5,
  "type": "VAELoader",
  "pos": [200, 1000],
  "size": [320, 60],
  "inputs": [
    {
      "name": "vae_name",
      "type": "COMBO",
      "widget": { "name": "vae_name", "default": "wan_2.1_vae.safetensors" },
      "link": null
    }
  ],
  "outputs": [{ "name": "VAE", "type": "VAE", "links": [999] }],
  "properties": {
    "Node name for S&R": "VAELoader",
    "models": [
      {
        "name": "wan_2.1_vae.safetensors",
        "url": "https://huggingface.co/waaayoff/vae"
      }
    ]
  }
}

📝 Notes​

  • Not all VAEs are created equal. wan_2.1_vae stands out as a top-tier VAE for realism-heavy SD 2.1 workflows.
  • Always match your VAE to your checkpoint. Mixing 1.5 checkpoints with a 2.1 VAE is like mixing oil and water and calling it a smoothie.
  • If your generations look off, try swapping VAEs. It's one of the fastest ways to troubleshoot rendering quality.
  • If you're building workflows with reusable nodes, create a dedicated Load VAE subgraph to swap VAEs quickly.

Need help troubleshooting? Just remember the holy trinity:
Checkpoint ➜ VAE ➜ Decoder

Get one of those wrong and you’re just diffusing into the void.