BlehTAEVideoDecode

TAESD has been the go-to trick for fast image previews for a while now: a tiny, fast autoencoder that trades some fidelity for a massive speed win over running the real VAE. This node is that same idea, ported to video. It uses TAE-family models from madebyollin (the same person behind TAESD) built specifically for video latent formats - Wan, Hunyuan, Mochi, and LTX all get their own dedicated tiny decoder - so you can turn a video latent into actual image frames without paying for a full VAE decode every time.

That matters more for video than it ever did for images. Wan and Hunyuan generations are already slow, and a full VAE decode of a whole video clip's worth of frames is its own meaningfully expensive step - it's not a rounding error the way a single-image VAE decode usually is. A fast approximate decoder that's noticeably lower cost, even at some quality loss, is a genuinely different value proposition here than it was for still images.

Inputs and outputs

• latent_type - which TAE model to use, matched to your video model's latent format: hunyuanvideo, ltxv, ltxv23, ltxv23wide, mochi, wan21, or wan22. The node's own tooltip flags one specific gotcha: use ltxv for "LTX-2 AV" - the naming doesn't map one-to-one onto every LTX variant, so read the option list carefully rather than guessing from the model name alone.
• parallel_mode - boolean, default false. The tooltip is direct about the tradeoff: "Parallel mode may be faster but requires more memory." Leave it off unless you've got VRAM to spare and want the speed.
• latent - the LATENT to decode.
• Output: IMAGE - the decoded frames, ready for a save node, a preview node, or a video-combine node to stitch into an actual video file.

Installing it - and the model download step you can't skip

Installing the pack itself is the same as any other Bleh node:

• ComfyUI Manager - search "ComfyUI-bleh", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/blepping/ComfyUI-bleh, restart ComfyUI.

Unlike most of the pack, though, this node won't do anything until you've downloaded the matching TAE model file and dropped it into models/vae_approx in your ComfyUI install, keeping the filename exactly as given:

• Wan 2.2: taew2_2.pth - but note the README's own caveat here: the 14B Wan 2.2 models don't actually use the Wan 2.2 latent format, they use Wan 2.1's. This file is only relevant for the smaller 5B Wan 2.2 model.
• Wan 2.1: taew2_1.pth
• Hunyuan: taehv.pth
• Mochi: taem1.pth
• LTX 2.0: laeltx_2.pth
• LTX 2.3: laeltx2_3.pth
• LTX 2.3 wide: taeltx2_3_wide.pth

All of these are hosted in madebyollin's taehv/taem1 repos on GitHub - check the pack's README for the exact links, since the filenames alone won't tell you which repo/branch each one lives on.

Common issues

Wrong or garbled output. The single most likely cause is a latent_type mismatch against what actually produced the latent. Because Wan 2.2's 14B models secretly use the Wan 2.1 latent format, picking wan22 for a 14B Wan 2.2 generation is a plausible mistake that will hand the decoder a format it wasn't built for - use wan21 there instead, and save wan22 for the 5B model.

Missing model file / node errors on load. These decoders are not bundled with the pack; you have to fetch the .pth file yourself and place it in models/vae_approx under the exact filename the README specifies. A renamed file won't be found.

Out-of-memory with parallel_mode on. That's the documented tradeoff, not a bug - parallel mode trades memory for speed. Turn it off if you're tight on VRAM, especially on longer clips or higher resolutions where a full-length video's worth of frames adds up fast.

Output looks softer or less detailed than a full VAE decode. That's inherent to the whole TAE approach, same as TAESD for images - it's an approximation built for speed, not a drop-in replacement for the real decoder when you need maximum final-output fidelity. Use it where speed matters (fast previews, iteration, quick exports) and fall back to the model's real VAE decode for your final pass if quality is the priority.