Official Review · ICLR 2026

This paper draws on the ideas of DDPM and transfers the generation process of image denoising to object detection. Different from the traditional DiffusionDet method, which denoises noisy boxes and then combines them with RCNN for detection, this paper integrates the denoising process into the attention process. Furthermore, based on the deformable DETR and the contrastive denoising queries (CDN) of DINO, it proposes two models, namely DiffuDETR and DiffuDINO, respectively. Comparisons with baselines and some contemporaneous methods on three datasets (COCO2017, LVIS, and V3Det) demonstrate that the proposed models achieve the best accuracy. 1. Integrates the denoising process with the query decoding mechanism of attention. 2. Validates the two proposed models on three sufficient datasets. 1 The flow chart fails to effectively illustrate the network architecture, and it also appears somewhat small. 2 The symbol "f" is used in both Equation 3 and Equation 4 but represents different meanings, which should be avoided. 3. The methods used for comparison are somewhat outdated, as the latest version of DiffusionDet was released in 2023. Q1: Does the generation of initial queries still adopt the top-k approach? In Line 76 of the paper, it is mentioned that "We propose a new query initialization technique that aligns with the objective of denoising diffusion models to sample from the normally distributed reference points. ", yet the illustration in Figure 2 still indicates the use of learnable initial queries. Q2: Is the denoising process reasonable? The method section shows that during training and inference, noisy boxes are not the main body of iteration; instead, they are used in the form of K to compute cross-attention with Q, assisting the iteration of Q. This step encourages the model to generate Q that conforms to the distribution of K, given that K represents noisy boxes. However, during inference, noisy boxes are directly sampled from a Gaussian distribution and iterated step-by-step following the DDPM paradigm. This iteration of noise does not align with the distribution of targets, yet the model still generates Q with a similar distribution to the noise, which may mislead the model. Q3: How are noisy boxes embedded? Are noisy boxes directly used as K for cross-attention computation, or are they mapped to the dimension of Q? Q4: Are the noisy boxes r at different time steps t independently and identically distributed (i.i.d.)? Or is the noise accumulated incrementally like in DDPM? If the latter is the case, I believe this should be more explicitly represented in Equation 4, and the meaning of "q" therein should also be clarified. Q5: In Equation 3, does the time step t embedding have different effects at different positions? Q6: Could you provide a detailed definition of "Decoder Evaluation"?