🥮 Tiny AutoEncoder for Hunyuan Video

What is TAEHV?

TAEHV is a Tiny AutoEncoder for Hunyuan Video (and other similar video models). TAEHV can encode and decode latents into videos more cheaply (in time & memory) than the full-size video VAEs, at the cost of slightly lower quality.

Here's a comparison of the output & memory usage of the Full Hunyuan VAE vs. TAEHV during decoding:

Decoder AE	Full Hunyuan VAE	TAEHV
Decoded Video (converted to GIF)
Runtime (in fp16, on GH200)	~2-3s for decoding 61 frames of (512, 320) video	~0.5s for decoding 61 frames of (512, 320) video. Can be even faster with the right settings
Memory (in fp16, on GH200)	~6-9GB Peak Memory Usage	<0.5GB Peak Memory Usage

See the profiling notebook for details on this comparison or the example notebook for a simpler demo.

What video models does TAEHV support?

To use TAEHV with different video models, you can load the different model weight files from this repo:

• For MiniMax H3, load the taeh3 weights (currently supported in ComfyUI nightly, ComfyUI-KJNodes and ComfyUI-bleh)
• For Hunyuan Video 1.5, load the taehv1_5 weights (example notebook).
  • You can also use taehv1_5_super to get higher-quality output at the cost of slightly more compute.
• For Wan 2.1, load the taew2_1 weights (see the Wan 2.1 example notebook).
• For Wan 2.2, load different files depending on model scale:
  • For Wan 2.2 5B, load the taew2_2 weights (example notebook).
    • You can also use taew2_2_super to get higher-quality output at the cost of slightly more compute.
  • For Wan 2.2 14B, load the taew2_1 weights since Wan 2.2 14B still uses the older Wan 2.1 VAE.
• for Qwen Image, load the taew2_1 weights (since Qwen Image uses the Wan 2.1 VAE encoder).
• For CogVideoX, load the taecvx weights (example notebook).
• For Hunyuan Video 1, load the taehv weights (example notebook).
• For Open-Sora 1.3, load the taeos1_3 weights.
• For LTX-2, load the taeltx_2 weights (example notebook).
• For LTX-2.3, load the taeltx2_3 weights
  • The standard TAE decoder for LTX2.3 has blurry outputs (see thread) so I also trained a larger, less-blurry taeltx2_3_wide variant. Those taeltx2_3_wide weights are here and ComfyUI-bleh has taeltx2_3_wide-compatible previewing code here.
• For Mochi 1 and SVD (which use different architectures), see the other repos TAEM1 and TAESDV.

The main model weight .pth files are in the repository root directory. Converted .safetensors files are located in the safetensors subdirectory.

If there's another open video model that would benefit from a TAEHV version, please file an issue (or, worst-case, try training your own).

Where can I get TAEHV?

TAEHV is available:

• In ComfyUI
  • In main, thanks to this PR from Kijai
  • Via the ComfyUI-WanVideoWrapper + VideoHelperSuite nodes thanks to Kijai and AustinMroz
  • Via the ComfyUI-Bleh nodes thanks to blepping
• In stable-diffusion.cpp thanks to this PR from stduhpf
• In SDNext thanks to vladmandic
• In the Wan2.1 Self-Forcing demo thanks to Guande He and Xun Huang

If you've added TAEHV support elsewhere, LMK and I can add a link here.

What are the limitations of TAEHV?

TAEHV is fast and well-suited for live previewing or interactive video, but TAEHV doesn't yet match the quality of the full-size VAEs.

You can see detailed quality comparisons of TAEHV vs full-size VAEs here.

How do I use TAEHV with 🧨 Diffusers?

You can use TAEHV with Diffusers by applying a small bit of wrapper code (example notebook). If you're writing new code involving both TAEHV and Diffusers, keep the following conventions in mind:

1. TAEHV stores image values in the range [0, 1], whereas Diffusers uses [-1, 1].
2. TAEHV stores videos in NTCHW dimension order (time, then channels), while Diffusers stores videos in NCTHW dimension order.
3. TAEHV does not use any latent scales / shifts (TAEHV encodes / decodes exactly what diffusion models use), whereas Diffusers requires explicitly applying a latents_mean and latents_std each time you encode or decode something.

How can I make TAEHV even faster?

You can disable TAEHV's temporal or spatial upscaling to get even-cheaper decoding.

TAEHV(decoder_time_upscale=(False, False), decoder_space_upscale=(True, True, True))

TAEHV(decoder_time_upscale=(False, False), decoder_space_upscale=(False, False, False))

If you have a powerful GPU or are decoding at a reduced resolution, you can also set parallel=True in TAEHV.decode_video to decode all frames at once (which is faster but requires more memory).

How can I use TAEHV to encode and decode live video?

TAEHV's default encode_video and decode_video methods process the entire input sequence before returning the entire result. For live / real-time scenarios (like video-to-video or world modeling), you'll want to use the StreamingTAEHV wrapper, which processes each input timestep sequentially and streams intermediate outputs as soon as they're ready.

You can see example StreamingTAEHV usage in the streaming demo notebook or in taehv.py's StreamingTAEHV docstrings.

How can I make TAEHV decode video at higher-quality?

For taehv1_5 and taew2_2, you can use the larger taehv1_5_super and taew2_2_super weights to get higher-quality video output at the cost of slightly increased compute (and higher parameter count). These super models are intended for realtime world modeling and video usecases where TAEHV output serves as the final output (rather than a temporary preview).

Here's a side-by-side comparison using Overworld's Waypoint-1.5; though neither VAE is perfect, TAEHV1.5-Super does a better job at decoding fine details and keeping them visually-stable.

taehv1_5_super_demo_trimmed.mp4

How can I cite TAEHV in a publication?

If you find TAEHV useful in your research, you can cite the TAEHV repo as a web link:

@misc {BoerBohan2025TAEHV,
  author = {Boer Bohan, Ollin},
  title = {TAEHV: Tiny AutoEncoder for Hunyuan Video},
  year = {2025},
  howpublished = {\url{https://github.com/madebyollin/taehv}},
}

The TAEHV repo contents change over time, so I recommend also noting the latest commit hash and access date in a note field, e.g.

note = {Commit: \texttt{5ce7381}, Accessed: 2025-09-05}