Consistency VAE Decoder

In November 2023, at OpenAI's DevDay, OpenAI open-sourced the consistency decoder: a big feedforward network that decodes a Stable Diffusion latent into an image in a single pass, positioned as a drop-in replacement for the SD 1.0+ VAE decoder. A DALL·E 3 team member confirmed in the announcement thread that it was the VAE behind DALL·E 3. The pitch was better text, faces, and straight lines. This node drags that decoder into ComfyUI, so instead of the usual VAE Decode you can decode your SD 1.5 latent through OpenAI's network.

Here's the honest framing: this is a piece of 2023 nostalgia with a genuinely interesting mechanism, and it's worth understanding, but it is not your everyday decoder. More on that in a minute.

How it works

The node loads openai/consistency-decoder from the HuggingFace Hub using diffusers' ConsistencyDecoderVAE, in fp16 on CUDA, and decodes latent["samples"] straight to an image, clamping to [0,1]. You feed it a latent, you get an IMAGE out. That's the whole interface - no VAE input, no settings, because the decoder is baked into the model itself.

The model is roughly 2.5GB and is not downloaded at install. It's pulled from HF into ComfyUI/models/huggingface_cache/ on first run, so expect one long wait and then a fast-ish decode.

Install

It ships in ComfyUI-GlifNodes, the pack from the glif.app team. ComfyUI Manager → search "ComfyUI-GlifNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/glifxyz/ComfyUI-GlifNodes
pip install -r requirements.txt   # diffusers>=0.28.0

Restart ComfyUI, then use it in place of VAE Decode in an SD 1.5 workflow.

The reality check

Three caveats, and they're big ones. First, the consistency decoder only understands SD 1.x/2.x latent space. Feed it SDXL or Flux latents and you'll get noise-flavored garbage. Second, the community verdict was genuinely mixed - for every person who loved the sharper text there was someone reporting faces worse than the classic vae-ft-mse-840000-ema-pruned, plus decoding that took ~20 seconds where the stock VAE took one. Third, the hype died fast: reddit mentions of the consistency decoder basically vanish after mid-2024. It's a curiosity and a way to squeeze better text out of stubborn SD 1.5 models, not a quality lever you should expect to keep.

If the latent is big and your GPU is small, you'll OOM on this node - at which point you want the tiled sibling from the same pack (Patch Consistency VAE Decoder), which swaps the decoder in with tiled decoding so large images survive on limited VRAM.