Some issues with using teacache to speed up image generation in forge

[Firetheft]

https://github.com/Firetheft/sd-forge-teacache
Here is a teacache plugin for forge. I feel that the image quality is not very ideal. Even if the Relative L1 Threshold value is set to 0.25, it is always very different from the image structure generated by disabling the teacache plugin, and it seems to be far from the effect in comfyui.
You can try it. It would be better if there are technical experts who can solve the image quality problem.
The original author of the plugin seems to have used AI to generate the code, and I only have a vague understanding of the code.

Metadata：
A melancholy old wizard sits alone in a dimly lit stone chamber, the flickering light of a small fire casting long, dancing shadows on the rough walls. His weathered face is framed by a long, flowing gray beard, streaked with white, and his tattered pointed hat droops slightly to one side. The wizard gazes into the distance, his heavy-lidded eyes filled with sorrow and nostalgia, as he puffs on a wooden pipe, the smoke curling lazily upward in ethereal, glowing tendrils. His robe, once resplendent with mystical patterns, is now faded and frayed, its colors muted by time. The background hints at shelves crammed with dusty books, ancient scrolls, and forgotten magical artifacts.

In the middle of the image, overlaid text in an elegant, glowing script reads:
'I Used to Get 5 Stars on GTA Just to Feel Wanted.'
Steps: 25, Sampler: DPM++ 2M, Schedule type: Simple, CFG scale: 1, Distilled CFG Scale: 3.5, Seed: 734200992, Size: 832x1216, Model hash: 7190d3aefe, Model: flux1-dev-fp8-16g, Version: f2.0.1v1.10.1-previous-636-gb835f24a

Using teacache in forge:

Using teacache in comfyui:

[LiewFeng]

See here. TeaCache leverages the timestep embedding modulated noisy input while sd-forge-teacache leverages noisy input directly.