ComfyUI-TaylorSeer
October 7, 2025 · View on GitHub
This project is the ComfyUI implementation of the TaylorSeer project https://github.com/Shenyi-Z/TaylorSeer.
Important Note
Please ensure your ComfyUI version is newer than commit c496e53.
Project Updates
- update:
2025/10/07: Based on TaylorSeer-Lite, support Qwen-Image. - update:
2025/09/25: Support TaylorSeer-Lite, reducing cache quantity with negligible VRAM increase. Based on TaylorSeer-Lite, now supporting WAN 2.1/2.2 models. - update:
2025/05/25: Support block swap, now you can run it with low VRAM - update:
2025/05/13: Support Hidream, force VRAM purge when checkpoint is changed - update:
2025/04/30: First release, supporting FLUX.
Quick Start
Installation
- Navigate to
ComfyUI/custom_nodes - Clone this repository
- Run ComfyUI after installation is complete!
Running the Workflow
Workflow for FLUX-TaylorSeer-Lite
Workflow for WAN-2.2-TaylorSeer-Lite
Workflow for Qwen-Image-TaylorSeer-Lite
Usage Instructions
Using TaylorSeer Standard Version
Memory Requirements Flux: The cache needs to be stored in GPU memory for efficient computation. For a 1024*1024 image using FLUX FP8 precision model:
- Order 0: Increases VRAM usage by 2GB
- Order 1: Increases VRAM usage by 4GB
- Order 2: Increases VRAM usage by 6GB
VRAM usage increases linearly with resolution and number of images.
Memory Requirements Hidream: The cache needs to be stored in GPU memory for efficient computation. For a 1024*1024 image using Hidream-full FP8 precision model:
- Order 0: Increases VRAM usage by 5GB
- Order 1: Increases VRAM usage by 10GB
- Order 2: Increases VRAM usage by 15GB
VRAM usage increases linearly with resolution and number of images.
Acceleration Ratio: The first_enhance parameter can adjust when Taylor Cache intervenes. When first_enhance = 10, with 30 iteration steps, the results are almost lossless compared to the original results, and the acceleration ratio can reach 2x.
Using TaylorSeer-Lite
Nearly zero VRAM increase.
Exciting Performance on WAN 2.2: For 81 frames generation on RTX 5090, TaylorSeer-Lite achieves remarkable acceleration - 386s vs 1176s (3.05x speedup) compared to the original implementation!
Comparison with teacache
Compared to TeaCache, TaylorSeer maintains a higher acceleration ratio and preserves more consistent composition and elements relative to the original image.
| Prompt | Original (steps = 50, 2.42it/s) | TaylorSeer (steps = 50, order = 1, first_enhance = 10, 5.73it/s) | TeaCache (steps = 50, rel_l1_thresh = 0.25, 4.08it/s) | TeaCache (steps = 50, rel_l1_thresh = 0.40, 5.32it/s) |
|---|---|---|---|---|
| fashion photo of a model wearing black draped plastic fabric designed by Demna Gvasalia, standing in an apocalyptic room, sunglasses, synthetic light, 4K, photoreal, 3D render, |  |  |  |  |
| This picture depicts an anime scene that revolves around a young girl. She is characterized by purple hair tied into two ponytails, each decorated with a bow and pink earmuffs. The girl is wearing dark clothes, a vest-style top and shorts. She has a happy expression, a big smile on her face, and a pearl or something delicate in her mouth, as if she is jokingly feeding herself. |  |  |  |  |
License
The code in this repository is released under the GNU General Public License v3.0.