In this video, I will introduce the features of "Detailer Hook" and the newly added "cycle" feature in Detailer. These features are applicable across DetailerForeach, MaskDetailer, and FaceDetailer, except DetailerForAnimateDiff. The "cycle" is a simple concept of repeatedly applying the Detailer. The key difference from splitting the Detailer is that it occurs at the Latent stage without going through VAE Encode/Decode. To change the armor part of the super girl to a metal material and enhance the damage effect, a mask was applied, and prompts were set. Let's generate with a "cycle" set to 1. Looks good.
Now, let's increase the "cycle" to 3 and apply the Detailer. The damage effect on the armor appears more pronounced.
Using Detailer Hook allows for more diverse changes. Using DenoiseSchedulerDetailerHook with target_denoise set to 0.3, the denoise gradually drops from 0.5 to 0.3 during the cycle. This strategy allows for strong changes early on and gradually refining details. NoiseInjectionDetailerHook injects noise into the latent during the cycle. It requires ComfyUI_Noise to be installed as it uses BNK_InjectNoise internally. Similarly, "start_strength" and "end_strength" control the amount of noise injected progressively during the cycle. Using "from_start" injects noise from the first cycle, while "skip_start" skips noise injection in the first cycle. Let's combine these two hooks and generate an image. The details in the abdomen of the armor have improved slightly. "UnsamplerDetailerHook" inverses the sampling process in the latent during the cycle. Unlike injecting random noise globally, unsampling allows focused removal of specific elements through prompts. I applied unsampling for prompts like perfect "armor, clean, glassy, blue". Also, "end_at_step" is increased from 16 to 24. Unlike KSampler, a higher end_at_step reduces the unsample steps as it approaches the total steps. Adjust the Switch using UnsamplerDetailerHook and generate the image again. The damage effects and details are much more prominent.