I evaluated the effects of TeaCache using the Flux.1-dev anime model.

max_skip_step = 1

max_skip_step = 2

max_skip_step = 3

• Graph
  • Horizontal axis: Generation time
  • Vertical axis: Image quality (similarity to the original image)
  • Smaller numbers: rel_l1_thresh
• Terms
  • MAE: Mean Absolute Error
  • SSIM: Structural Similarity Index

From the graphs, there is a tendency for image quality to decrease as generation time shortens, but no significant differences were observed due to the value of max_skip_step.

SSIM Similarity Trend Lines for Each max_skip_step

Comparing the SSIM Similarity trend lines for each max_skip_step, there is no significant difference in the slopes, indicating that the trade-off balance between speed and image quality remains largely the same across settings.

Image Comparison at TeaCache Maximum Settings

At TeaCache's maximum settings (max_skip_step = 3, rel_l1_thresh = 1), the depiction of fingers and text becomes inaccurate, but the generation time is reduced to one-third.

The measurement data and comparison with ComfyUI-WaveSpeed is summarized on my blog site.