DragNoise is a point-based interactive editing framework that avoids global adjustments of the latent code/map, a common issue in frameworks like DragGAN and DragDiffusion, facilitating stable and semantically accurate point-based editing. The core idea of DragNoise is to use the predicted noise output of each U-Net as a semantic editor. This approach is based on two key observations: first, the bottleneck features of U-Net inherently possess semantically rich features ideal for interactive editing; second, high-level semantics established early in the denoising process show minimal variation in subsequent stages. DragNoise edits diffusion semantics in a single denoising step and efficiently propagates these changes, ensuring stability and efficiency in diffusion editing. Comparative experiments show that DragNoise achieves superior control and semantic retention, reducing optimization time by over 50% compared to DragDiffusion. DragNoise leverages diffusion semantic propagation, treating predicted noises as sequential semantic editors. It starts editing at a timestep where high-level semantics are well-trained, optimizing the bottleneck feature of the U-Net to reflect user edits. The optimized bottleneck feature learns the intended dragging effect and produces the corresponding manipulation noise. This optimized bottleneck feature is then propagated to all subsequent timesteps, avoiding redundant feature optimization. DragNoise significantly enhances the manipulation effect in a stable and efficient manner. It conducts extensive quantitative and qualitative experiments on the drag-based editing benchmark DragBench and diverse example images, demonstrating its effectiveness. DragNoise significantly reduces optimization time by over 50% compared to DragDiffusion. It also explores the impacts of various initial timesteps on the editing process, the optimization of different layers, and the extent of optimization propagation, highlighting DragNoise's efficiency and flexibility in interactive editing. DragNoise is implemented based on Stable Diffusion 1.5, with a LoRA model trained with the user-supplied image before diffusion semantic optimization. The method uses DDIM inversion to attain the noisy latent map with t = 35. The optimization process is conducted with a learning rate of 0.01 and a maximum number of optimization steps of 80. DragNoise achieves superior editing capabilities, ensuring more accurate dragging even amidst substantial changes while preserving semantics or identities more effectively. It supports multi-point editing, demonstrating its versatility in diverse editing scenarios. DragNoise's optimization efficiency is analyzed, showing that it reduces the number of optimization steps by over 50% compared to DragDiffusion. DragNoise's performance is evaluated on the DragBench dataset, demonstrating its superior dragging accuracy and image fidelity. DragNoise's limitations include the challenge in handling real images while preserving their original fidelity. Future research should aim at creating a universal adapter to maintain diffusion model fidelity and improve editing that requires integrated point understanding.DragNoise is a point-based interactive editing framework that avoids global adjustments of the latent code/map, a common issue in frameworks like DragGAN and DragDiffusion, facilitating stable and semantically accurate point-based editing. The core idea of DragNoise is to use the predicted noise output of each U-Net as a semantic editor. This approach is based on two key observations: first, the bottleneck features of U-Net inherently possess semantically rich features ideal for interactive editing; second, high-level semantics established early in the denoising process show minimal variation in subsequent stages. DragNoise edits diffusion semantics in a single denoising step and efficiently propagates these changes, ensuring stability and efficiency in diffusion editing. Comparative experiments show that DragNoise achieves superior control and semantic retention, reducing optimization time by over 50% compared to DragDiffusion. DragNoise leverages diffusion semantic propagation, treating predicted noises as sequential semantic editors. It starts editing at a timestep where high-level semantics are well-trained, optimizing the bottleneck feature of the U-Net to reflect user edits. The optimized bottleneck feature learns the intended dragging effect and produces the corresponding manipulation noise. This optimized bottleneck feature is then propagated to all subsequent timesteps, avoiding redundant feature optimization. DragNoise significantly enhances the manipulation effect in a stable and efficient manner. It conducts extensive quantitative and qualitative experiments on the drag-based editing benchmark DragBench and diverse example images, demonstrating its effectiveness. DragNoise significantly reduces optimization time by over 50% compared to DragDiffusion. It also explores the impacts of various initial timesteps on the editing process, the optimization of different layers, and the extent of optimization propagation, highlighting DragNoise's efficiency and flexibility in interactive editing. DragNoise is implemented based on Stable Diffusion 1.5, with a LoRA model trained with the user-supplied image before diffusion semantic optimization. The method uses DDIM inversion to attain the noisy latent map with t = 35. The optimization process is conducted with a learning rate of 0.01 and a maximum number of optimization steps of 80. DragNoise achieves superior editing capabilities, ensuring more accurate dragging even amidst substantial changes while preserving semantics or identities more effectively. It supports multi-point editing, demonstrating its versatility in diverse editing scenarios. DragNoise's optimization efficiency is analyzed, showing that it reduces the number of optimization steps by over 50% compared to DragDiffusion. DragNoise's performance is evaluated on the DragBench dataset, demonstrating its superior dragging accuracy and image fidelity. DragNoise's limitations include the challenge in handling real images while preserving their original fidelity. Future research should aim at creating a universal adapter to maintain diffusion model fidelity and improve editing that requires integrated point understanding.