cui-teacache-lu2

referenced from https://github.com/spawner1145/TeaCache/tree/main/TeaCache4Lumina2

firstly transplanted by @fexli

retransplanted by @spawner

Installation

Manual installation

// switch to your project's root directory
cd custom_nodes
git clone https://github.com/spawner1145/CUI-Lumina2-TeaCache.git

Installation via comfyui-manager

1. Open ComfyUI WebUI
2. Navigate to Manager -> Install Custom Node
3. Enter CUI-Lumina2-TeaCache in the Search field, and click Search
4. Click Install

usage

1. Connect the TeaCache node between the UNet Loader and KSampler in your workflow.
2. Set the rel_l1_thresh parameter to a value greater than 0.
3. to work on low steps, you can set the value below to [393.76566581, -603.50993606, 209.10239044, -23.00726601, 0.86377344] and a small rel_l1_thresh like 0.3 for higher speed or set the value below to [225.7042019806413, -608.8453716535591, 304.1869942338369, 124.21267720116742, -1.4089066892956552] and a very large rel_l1_thresh like 5 for higher speed and better quality, and for higher steps, you can set the value below to [225.7042019806413, -608.8453716535591, 304.1869942338369, 124.21267720116742, -1.4089066892956552] and a rel_l1_thresh<1.1 to get better quality and higher speed.
4. The nodes are configured with different parameters. When using 25 steps or fewer, it is recommended to set the l1 value to approximately 6. For larger step sizes, the l1 should be decreased proportionally. For instance, a value of 0.6 is suggested for 50 steps.([225.7042019806413, -608.8453716535591, 304.1869942338369, 124.21267720116742, -1.4089066892956552])

Note:

• Higher rel_l1_thresh values will improve generation efficiency (manifested as shorter generation times), at the cost of reduced image quality.
• The optimal value should be determined through empirical testing based on your specific quality/efficiency requirements.

reference

TeaCache can speedup Lumina-Image-2.0 without much visual quality degradation, in a training-free manner. The following image shows the results generated by TeaCache-Lumina-Image-2.0 with various rel_l1_thresh values: 0 (original), 0.2 (1.25x speedup), 0.3 (1.5625x speedup), 0.4 (2.0833x speedup), 0.5 (2.5x speedup).

<p align="center"> <img src="https://github.com/user-attachments/assets/d2c87b99-e4ac-4407-809a-caf9750f41ef" width="150" style="margin: 5px;"> <img src="https://github.com/user-attachments/assets/411ff763-9c31-438d-8a9b-3ec5c88f6c27" width="150" style="margin: 5px;"> <img src="https://github.com/user-attachments/assets/e57dfb60-a07f-4e17-837e-e46a69d8b9c0" width="150" style="margin: 5px;"> <img src="https://github.com/user-attachments/assets/6e3184fe-e31a-452c-a447-48d4b74fcc10" width="150" style="margin: 5px;"> <img src="https://github.com/user-attachments/assets/d6a52c4c-bd22-45c0-9f40-00a2daa85fc8" width="150" style="margin: 5px;"> </p>

📈 Inference Latency Comparisons on a single 4090 (step 50)

| Lumina-Image-2.0 | TeaCache (0.2) | TeaCache (0.3) | TeaCache (0.4) | TeaCache (0.5) | |:-------------------------:|:---------------------------:|:--------------------:|:---------------------:|:---------------------:| | ~25 s | ~20 s | ~16 s | ~12 s | ~10 s |

special thanks

fexli

The original TeaCache transplant of Lumina2 in cui

welltop-cn

model patch code design