لخّصلي

4.2.In addition, weJust train ResENorm to extrapolate accuracy for the betterAdapting statistical distribution of feature map for high-resolution images.For example, ResCLoRA has been integrated into the diffusion model and still produces higher resolution images with poor framing and composition.Extrapolation of decision: In Sec.4.2, we introduce ResENorm, which compensates for the lack of capability about ResCLoRA in resolution extrapolation Our initial experience found that there was a significant gap between LoRA and full fine-tuning, meaning that only ResCLoRA did so, not enabling the resolution extrapolation capability of the custom model.

4.2. Extrapolation of decision: In Sec.4.2, we introduce ResENorm, which compensates for the lack of capability about ResCLoRA in resolution extrapolation Our initial experience found that there was a significant gap between LoRA and full fine-tuning, meaning that only ResCLoRA did so, not enabling the resolution extrapolation capability of the custom model. For example, ResCLoRA has been integrated into the diffusion model and still produces higher resolution images with poor framing and composition. However, we find that if all normalization layers are present UNet blocks are trained on the public dataset, the trained weights are not compatible with other personal publishing models, resulting in low-quality images with poor color style, we only open the block normalization to the Resnet layer, which It's called ResENorm. He. SheNot only reduce the gap about the previous decision between themResCLoRA and full fine-tuning, but it also helps maintainThe style field of the custom form. In addition, weJust train ResENorm to extrapolate accuracy for the betterAdapting statistical distribution of feature map for high-resolution images. After that, ResENorm withResCLoRA can improve bad framework and configurationFrom pictures of the generation. Especially, ResENorm only occupies0.1 million parameters but it creates effects to reduce the gap in itExtrapolation of decision.