Skip to main content

vae-ft-mse-840000-ema-pruned.safetensors

The vae-ft-mse-840000-ema-pruned.safetensors is a fine-tuned, lightweight VAE trained using Mean Squared Error (MSE) loss and Exponential Moving Average (EMA) smoothing. It’s optimized for SD 1.4/1.5 checkpoints and is widely used in ComfyUI workflows to decode latent images into high-quality pixel outputs.

This isn’t just a VAE—it’s the VAE you swap in when you’re tired of the default one giving your images the emotional range of a potato.

🔧 File Format​

  • Filename: vae-ft-mse-840000-ema-pruned.safetensors
  • File Type: .safetensors
  • Purpose: VAE for decoding latent representations into image space
  • Optimization: EMA + MSE for smooth, stable, high-fidelity output
  • Model Family: Stable Diffusion 1.x compatible

📁 Function in ComfyUI Workflows​

This VAE connects to key nodes like VAE Decode, KSampler, and CheckpointLoaderSimple. It improves final image output by ensuring that decoded images reflect the full potential of your latent vectors—without barfing color gradients or introducing bizarre lighting.

Typical hook-up points:

  • VAE Decode
  • KSampler (when a VAE is required alongside a checkpoint)
  • Auto-decoding pipelines

🧠 Technical Details​

PropertyValue
Training LossMean Squared Error (MSE)
Training Steps840,000
EMA Applied✅ Yes
Format.safetensors
Model Pruned✅ Reduced unnecessary parameters
Ideal Checkpoint Pairv1-5-pruned-emaonly.safetensors
ComfyUI Compatibility✅ Fully Supported

✅ Benefits​

  • Crisp edges and smoother skin rendering
  • Improved detail fidelity without bloating your workflow
  • Faster loading thanks to pruning
  • Compatible with virtually all 1.x-based checkpoints

⚙️ Usage Tips​

  • Best for: Realism workflows, portrait renders, concept art, style transfers
  • Avoid: Mixing with SDXL checkpoints (unless you're trying to reinvent glitch art)
  • Compare against vae-ft-mse-560000 or baked VAEs to assess output quality per use case

📍 ComfyUI Setup Instructions​

  1. Drop in your CheckpointLoaderSimple node.
  2. Set your checkpoint to v1-5-pruned-emaonly.safetensors (or similar).
  3. Set vae_name to vae-ft-mse-840000-ema-pruned.safetensors.
  4. Wire the VAE output to your VAE Decode node.
  5. Bask in your gloriously reconstructed images.

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

Let’s avoid turning your beautiful workflow into a dumpster inferno. Here’s what not to do:

  • ❌ Do not use with SDXL checkpoints
    This is a VAE for SD 1.x. If you connect it to SDXL models, don’t come crying when your render looks like a Dali painting got dunked in a microwave.
  • ❌ Do not pair with baked VAE checkpoints
    If your checkpoint already includes a baked-in VAE, this one will just argue with it. It’s like wearing two pairs of glasses at once—confusing and blurry.
  • ❌ Do not skip connecting the VAE in VAE Decode
    The node literally exists to decode using the VAE. Skipping it is like ordering sushi and forgetting the fish.
  • ❌ Do not rename the file with typos and wonder why it won’t load
    Misspell it as vae-ft-mse-840000-ema-purned.safetensors and watch ComfyUI have a quiet meltdown.
  • ❌ Don’t plug it into SDXL-specific nodes expecting magic
    This VAE doesn’t understand SDXL’s latent dimensions. It’s like trying to watch Netflix on a toaster.

📚 Additional Resources​

📎 Example Node Configuration​

json


{   "type": "CheckpointLoaderSimple",   "inputs": {     "ckpt_name": "v1-5-pruned-emaonly.safetensors",     "vae_name": "vae-ft-mse-840000-ema-pruned.safetensors"   } }

📝 Notes​

  • Do not mix incompatible VAE and checkpoint versions (e.g., SDXL VAEs with SD 1.5 base models).
  • If you notice unusual banding or oversaturation, test alternate VAEs or adjust the decoding step strength.