Skip to main content

ae.safetensors

Because your image deserves more than a potato-grade decode.

🔧 File Format​

  • Filename: ae.safetensors

  • File Type: .safetensors

  • Serialization Format: SafeTensors – a blazing-fast, zero-copy format designed for security and speed. It avoids common pitfalls of .ckpt or .pt (e.g. arbitrary code execution on load).

Why it matters:

Loading a VAE in SafeTensors format ensures compatibility, speed, and peace of mind (no malicious payloads embedded in your decoder). Plus, it's what ComfyUI actually prefers when you’re not trying to cook your machine with janky .ckpt files.

📁 Function in ComfyUI Workflows​

In a ComfyUI workflow, the VAE (Variational Autoencoder) is responsible for:

  • Decoding latent tensors into full RGB images.

  • Encoding images back into the latent space (when applicable).

  • Maintaining fidelity, dynamic range, and color integrity between the latent and visual representation.

Think of it as the bridge between your hallucinated dream world (latent) and the visible reality (image). If you don’t load a proper VAE, you're going to get weird outputs—blurry, overly contrasted, or plain old busted.

🧠 Technical Details​

  • Type: Variational Autoencoder

  • Architecture: Likely derived from SD 1.4/1.5’s original VAE backbone

  • Dimensions: Works with latent spaces of shape 4x64x64 for 512x512 inputs

  • Training: Pretrained on large datasets to preserve visual fidelity during decode

  • Latent Precision: 32-bit float, standardized

  • Compression Target: Latent downsampling by factor of 8 (e.g., 512x512 → 64x64)

This VAE is not fancy. It’s the vanilla backbone of stability—not trained for anime, not trained for face beautification, not trained for fantasy—just solid general-purpose image representation.

✅ Benefits​

  • Stable decoding for checkpoints trained on vanilla SD1.4 / SD1.5

  • Lightweight and memory efficient

  • Fast loading due to SafeTensors format

  • Good color retention** compared to baked-in VAEs

  • Compatible with nearly everything not trying to be special

⚙️ Usage Tips​

  • Always explicitly load this VAE if your checkpoint doesn't have a baked-in VAE (or if it's terrible).

  • If you're seeing muddy images, weird contrast, or blocky compression artifacts—your VAE might be misconfigured. Start here.

  • Match the VAE to the checkpoint. If you're using v1-5-pruned-emaonly, revAnimated, Deliberate, etc., this VAE is often a safe fallback.

  • If you’re using a fancy model (like anime, furry, or inpainting specialists), they might require a custom VAE. This one will still work, but possibly not well.

🤝 Best Model Compatibility​

The ae.safetensors VAE works best with Standard 1.x Stable Diffusion checkpoints like:

| Checkpoint                                                                                           | Status                                                  |

| ---------------------------------------------------------------------------------------------------- | ------------------------------------------------------- |

| v1-5-pruned-emaonly.safetensors                                                                    | ✅ Perfect match                                        |

| revAnimated_v122.safetensors                                                                       | ✅ Good default                                         |

| deliberate_v2.safetensors                                                                          | ✅ Compatible                                           |

| dreamshaper_8.safetensors | ✅ Works fine                                           |

| anything-v3.ckpt                                                                                   | ⚠️ Acceptable, but may benefit from anime-tuned VAEs    |

| RealisticVisionV5.safetensors                                                                      | ⚠️ Will decode, but may blunt realism-specific tuning   |

| SDXL-based models                                                                                  | ❌ Incompatible – completely different latent structure |

📍 Setup Instructions in ComfyUI​

  1. Make sure you have the file ae.safetensors placed in your ComfyUI/models/vae/ folder.

  2. In your workflow, add a Load VAE or VAELoader node.

  3. Select ae.safetensors from the dropdown menu (may show up as ae).

  4. Connect it to the appropriate nodes:

  • For CheckpointLoaderSimple, connect VAE output from the loader.

  • For VAE Decode, pipe in the VAE input.

  1. Preview your outputs with an ImageViewer node to verify that decoding looks sharp, accurate, and not like someone threw Vaseline on the lens.

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

  • ❌ Don’t skip loading a VAE entirely and expect usable results.

  • ❌ Don’t use this VAE with SDXL checkpoints – you will get garbage (if anything at all).

  • ❌ Don’t mix VAE expectations—i.e., using a baked-in VAE checkpoint with an external VAE like this can lead to double-decoding weirdness.

  • ❌ Don’t use this with inpainting-specific models unless you know the model was trained with this VAE.

  • ❌ Don’t overwrite this VAE thinking you’re being helpful. This is the "safe fallback" one. You’ll want to keep it.

📚 Additional Resources​

📎 Example Node Configuration​

VAELoader Node:

{ "type": "VAELoader", "vae_name": "ae.safetensors" }

Typical Workflow Connection:


CheckpointLoaderSimple        └──▶ Model, Clip        └──▶ VAE ──▶ KSampler / VAE Decode

📝 Notes​

  • This is not a "fancy" VAE—it’s a baseline. That’s the point. It’s there when your baked VAE fails or is just plain ugly.

  • This VAE does not support SDXL, which uses a different encoder/decoder architecture (with two VAEs, because why use one when you can overengineer?).

  • Keep this one as your default fallback in workflows. It’s reliable, tested, and unlikely to break.

Still seeing overcompressed blobs? Your issue might not be the VAE—it might be your sampler, CFG scale, or ControlNet settings. But starting with a good VAE like ae.safetensors gives your workflow a fighting chance.