DiffUHaul: A Training-Free Method for Object Dragging in Images (2406.01594v2)

Abstract: Text-to-image diffusion models have proven effective for solving many image editing tasks. However, the seemingly straightforward task of seamlessly relocating objects within a scene remains surprisingly challenging. Existing methods addressing this problem often struggle to function reliably in real-world scenarios due to lacking spatial reasoning. In this work, we propose a training-free method, dubbed DiffUHaul, that harnesses the spatial understanding of a localized text-to-image model, for the object dragging task. Blindly manipulating layout inputs of the localized model tends to cause low editing performance due to the intrinsic entanglement of object representation in the model. To this end, we first apply attention masking in each denoising step to make the generation more disentangled across different objects and adopt the self-attention sharing mechanism to preserve the high-level object appearance. Furthermore, we propose a new diffusion anchoring technique: in the early denoising steps, we interpolate the attention features between source and target images to smoothly fuse new layouts with the original appearance; in the later denoising steps, we pass the localized features from the source images to the interpolated images to retain fine-grained object details. To adapt DiffUHaul to real-image editing, we apply a DDPM self-attention bucketing that can better reconstruct real images with the localized model. Finally, we introduce an automated evaluation pipeline for this task and showcase the efficacy of our method. Our results are reinforced through a user preference study.

DiffUHaul: A Training-Free Method for Object Dragging in Images

"DiffUHaul: A Training-Free Method for Object Dragging in Images" introduces a novel approach for the challenging task of seamlessly relocating objects within an image. The apparent simplicity of moving objects belies the intricate spatial reasoning required, which current generative models often fail to deliver reliably. This paper leverages the spatial understanding of localized text-to-image models, particularly BlobGEN, to develop a robust, training-free solution named DiffUHaul.

Methodology

The proposed method addresses key issues in object dragging, including model entanglement, maintaining object appearance, and adapting the method for real images. Here is a breakdown of the approach:

1. Entanglement in Localized Models: The authors identify a crucial problem of entanglement in BlobGEN, specifically within the Gated Self-Attention layers. The leakage of attention across different objects compromises the disentanglement. To resolve this, the paper introduces an inference-time attention masking mechanism, ensuring that textual tokens only attend to their respective visual regions, significantly improving the disentanglement.
2. Consistency in Generated Images: To preserve the high-level object appearance during the dragging task, a self-attention sharing mechanism is employed. This mechanism replaces the keys and values in the target image with those from the source image across different denoising steps. Additionally, a novel soft anchoring technique interpolates self-attention features adaptively over the denoising process, promoting smooth fusion between the object's appearance and the target layout. The latter denoising steps involve finer adjustments using a nearest-neighbor copying strategy based on attention features.
3. Adaptation for Real Images: For real-image scenarios, the method overcomes inversion challenges seen with traditional DDIM inversion. Instead, a DDPM self-attention bucketing technique is used, which involves adding noise to the reference image independently at each diffusion step, allowing for better preservation of image details. Additionally, Blended Latent Diffusion is integrated to seamlessly blend the generated edits with the original background.

Numerical Results and User Studies

The authors validate their method against several baselines: Paint-By-Example, AnyDoor, Diffusion Self-Guidance, DragDiffusion, DragonDiffusion, and DiffEditor. Both qualitative assessments and three automatic metrics—foreground similarity, object traces, and realism—show that DiffUHaul consistently outperforms these baselines. Notably, DiffUHaul achieves higher foreground similarity and minimal object traces, all while maintaining high realism.

• Foreground Similarity: Measures the alignment between the source and target blobs after the drag.
• Object Traces: Evaluates if residual artifacts remain at the original object location.
• Realism: Assesses the perceptual quality of the generated images using KID scores.

Results from a user paper conducted on the Amazon Mechanical Turk platform reinforce these findings, showing that DiffUHaul is preferred over other methods across various quality dimensions including object placement, trace removal, realism, and overall quality.

Implications and Future Work

The implications of this method are significant for both practical applications and theoretical advancements. Practically, it offers a powerful tool for digital content creation, enabling artists and designers to manipulate images with higher fidelity and less effort. Theoretically, it pushes the boundaries of what training-free generative methods can achieve, particularly in tasks requiring intricate spatial reasoning.

Future developments might explore enhancing the capabilities of DiffUHaul to handle more complex scenarios, such as rotating objects, resizing them proportionally, and managing interactions between moving objects. Furthermore, integrating 3D spatial understanding could further improve the robustness and applicability of this method.

Conclusion

"DiffUHaul: A Training-Free Method for Object Dragging in Images" represents a significant step forward in image editing, addressing key challenges with a sophisticated yet efficient approach. By leveraging and modifying localized text-to-image models, the authors present a highly effective solution that blends practicality with theoretical innovation, marking a notable contribution to the field of computer graphics and machine learning.