# taeh3.safetensors

> TAEH3 · Tiny AutoEncoder for MiniMax H3 · Fast preview decoder for MiniMax H3 sampling

taeh3.safetensors is madebyollin’s Tiny AutoEncoder for MiniMax H3. It decodes H3 latents into rough video previews while you sample, at a fraction of the full VAE’s cost. Put it in ComfyUI/models/vae_approx and pick it in a preview node such as KJNodes’ Model Preview Override.

Kijai’s older 2D preview model is also called taeh3.safetensors but is a different file (9.8 MB). Check the size and SHA256 below.

## Facts

- **Type:** Tiny VAE (preview decoder)
- **Parameters:** 11.3M · fp16 · temporal decoder
- **Format:** Safetensors · fp16
- **Size:** 22.7 MB (22,709,752 bytes)
- **SHA256:** 4fd022bfcab08772fe0536b17ea1a3bbb5625be11e397868d1c5d891863d4c13
- **ComfyUI folder:** ComfyUI/models/vae_approx/
- **Loader node:** Model Preview Override
- **License:** MIT

## What is taeh3.safetensors?

Watching an H3 video take shape normally means decoding every preview through the 5-billion-parameter video VAE, which is too slow to do each step. TAEH3 is an 11.3M-parameter decoder trained by Ollin Boer Bohan (author of TAESD) to approximate that VAE. It turns latents into low-resolution previews quickly enough to update every couple of steps.

It is temporal: 5n+2 latent frames decode to 17n+5 video frames, the same frame count the full VAE produces, so previews play as motion rather than separate stills.

Two different files share this name. Kijai trained a quick 2D preview decoder first and also named it taeh3.safetensors (9.8 MB, SHA256 f0f60fa0…). Kijai now points to madebyollin’s version as the better one. Community packs rename Kijai’s file taeh3_2d_kijai.safetensors so both can live in the same folder.

## Download and install

- [GitHub · madebyollin/taehv](https://github.com/madebyollin/taehv/raw/62f7591f59dfbb4c3c02b7a621d180a9eeaba26c/safetensors/taeh3.safetensors) · saves as `taeh3.safetensors` · Author’s release. Recommended.
- [Hugging Face mirror · OzzyGT/taeh3](https://huggingface.co/OzzyGT/taeh3/resolve/main/taeh3.safetensors) · saves as `taeh3.safetensors` · Unmodified copy, same SHA256.

```sh
cd ComfyUI/models/vae_approx
wget -O taeh3.safetensors \
  https://github.com/madebyollin/taehv/raw/62f7591f59dfbb4c3c02b7a621d180a9eeaba26c/safetensors/taeh3.safetensors
sha256sum taeh3.safetensors
# 4fd022bfcab08772fe0536b17ea1a3bbb5625be11e397868d1c5d891863d4c13
```

Safetensors files hold tensors only and cannot run code. Still compare the SHA256: a failed download often leaves an HTML error page saved under the model’s filename.

## How to load it in ComfyUI

TAEH3 does not touch your final video. It only decodes sampler previews. Model Preview Override from KJNodes reads it from models/vae_approx and patches the MODEL before your sampler.

- **Turn previews on:** Set ComfyUI’s preview method to auto or TAESD so the sampler shows them.
- **Final output unchanged:** The full video VAE still decodes the result. TAEH3 only speeds up previews.

## Files a MiniMax H3 video workflow needs

Use the Ref2VA diffusion model instead of FL2VA for reference-to-video workflows.

- [`minimax_h3_fl2va_pruned_int8_convrot.safetensors`](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors) · Diffusion model (FL2VA) · `models/diffusion_models/` · 20.97 GB
- [`qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors) · Text encoder (Qwen3-VL 32B) · `models/text_encoders/` · 15.69 GB
- [`minimax_h3_video_vae_fp16.safetensors`](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors) · Video VAE · `models/vae/` · 5.21 GB
- [`minimax_h3_audio_vae_fp32.safetensors`](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors) · Audio VAE · `models/vae/` · 605 MB
- [`minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors`](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors) · Turbo LoRA, 4 steps · `models/loras/` · 1.96 GB

## Workflows: Uses this exact file

These workflows already load taeh3.safetensors. Open one and run it as is.

- [MiniMax H3 Long Video Generation | Motion Context V2](https://www.runcomfy.com/comfyui-workflows/minimax-h3-long-video-in-comfyui)
- [MiniMax H3 8-Step Acceleration | All-in-One Video Workflow](https://www.runcomfy.com/comfyui-workflows/minimax-h3-8-step-acceleration-all-in-one-comfyui-workflow)

## Same file, other names, and alternatives

- `taeh3.safetensors` · This page · madebyollin, temporal · 22.7 MB · Smooth video previews
- `taeh3_2d_kijai.safetensors` · Different model · Kijai, 2D (also shipped as taeh3.safetensors) · 9.8 MB · Per-frame stills; older setups
- `taeh3_decoder.safetensors` · Different file · Civitai preview pack · 39.5 MB · The MiniMax H3 Preview Override custom node

## Common errors and fixes

### Value not in list: tiny_vae: 'taeh3.safetensors' not in [...]

Why: The workflow asks for this exact filename, but ComfyUI cannot find it in models/vae_approx. Usually the file is missing, sits in a different folder, or was saved under another name such as minimaxH3Preview_v10.safetensors.

Fix: Move the file into ComfyUI/models/vae_approx/, rename it to match the workflow (or pick it again in the dropdown), then refresh the node list or restart ComfyUI.

### Missing node type: ModelPreviewOverrideKJ

Why: Model Preview Override is not built into ComfyUI. It comes from the ComfyUI-KJNodes custom node pack.

Fix: Install ComfyUI-KJNodes with ComfyUI Manager (or git clone it into ComfyUI/custom_nodes), then restart ComfyUI.

### SafetensorError: Error while deserializing header

Why: The download is incomplete, or the browser saved an HTML sign-in or error page under the model’s filename.

Fix: Check that the file is exactly 22,709,752 bytes and that its SHA256 matches the value above. Download it again from GitHub · madebyollin/taehv if not.

## FAQ

### I have two files called taeh3.safetensors. Which is which?

Check the size. 22,709,752 bytes (SHA256 starting 4fd022bf) is madebyollin’s temporal TAEH3 on this page. 9,791,388 bytes (starting f0f60fa0) is Kijai’s 2D version. They are different networks; keep both by renaming Kijai’s to `taeh3_2d_kijai.safetensors`.

### Does TAEH3 change the final video?

No. It only decodes previews during sampling. Your output still goes through the full MiniMax H3 video VAE.

### Where do I put taeh3.safetensors in ComfyUI?

In `ComfyUI/models/vae_approx/`. Then refresh the node list or restart ComfyUI so Model Preview Override lists the file.

### Which ComfyUI node loads taeh3.safetensors?

**Model Preview Override**. It comes from the [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes) custom node pack, so install that first.

### Can I use taeh3.safetensors commercially?

Yes. TAEH3 is released under the MIT license; keep the copyright notice if you redistribute it.

## Sources

- [madebyollin/taehv on GitHub](https://github.com/madebyollin/taehv)
- [OzzyGT/taeh3 mirror](https://huggingface.co/OzzyGT/taeh3)
- [Kijai/MiniMax-H3-TAE](https://huggingface.co/Kijai/MiniMax-H3-TAE)

Canonical page: https://www.runcomfy.com/comfyui-workflows/model-files/taeh3.safetensors
