TAESD provides fast high-quality latent previews without running the full VAE

By default ComfyUI shows low-resolution latent previews during sampling using a cheap internal decoder. The Tiny AutoEncoder for Stable Diffusion (TAESD) is a lightweight separate model that decodes latents to pixel space with much higher quality at low cost — separate model variants exist for SD1, SDXL, SD3, and FLUX. Enabling —preview-method taesd shows near-final-quality previews at each sampling step. The files go in models/vae_approx/. This matters for live performance: you can abort and adjust a generation early because the preview is readable at step 5–10.

Examples

Place taesd_decoder.pth in models/vae_approx/ and launch with —preview-method taesd to see high-quality previews during a 20-step SDXL run.

Assessment

Explain why TAESD previews can look different from the final VAE decode. When would you prefer —preview-method auto over taesd?