Pith. sign in

REVIEW 6 major objections 5 minor 54 references

Free-Lunch Augmentation by Revisiting Diffusion-Based Data Generation for Cross-Domain Few-Shot Object Detection

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Weakened noise makes diffusion-generated data usable for few-shot detection

desk verdict The SOTA claim outruns the evidence, but the background-inpainting recipe is cheap, training-free, and has real independent support—worth a serious referee, not a desk reject. read the letter →

arxiv 2608.04394 v1 pith:3IQX5KW6 submitted 2026-08-05 cs.CV

Zijian Zhuang , Yixiong Zou , Yuhua Li , Ruixuan Li This is my paper
classification cs.CV
keywords cross-domainfew-shotobjectdetectiondiffusionmodelsdataaugmentationbackgroundinpaintingnoisestrengthCKAsimilaritysegmentationtraining-freegeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that diffusion-based data augmentation, which currently hurts cross-domain few-shot object detection because synthesized images drift toward the source domain, can be repaired by handling visual and semantic domain gaps separately. For the visual gap, it argues that pretrained diffusion models mistake expert-domain content for noise, and that weakening the forward noising step preserves the original image's useful information. For the semantic gap, it argues that background regions transfer across domains far better than foregrounds, so inpainting the background avoids generating unseen target categories. These two fixes are wrapped into SITN, a training-free pipeline that generates candidate images, filters them by RPN IoU against the ground-truth box, and adds the survivors to the support set. The paper reports that this recipe improves detection on six CDFSOD datasets and segmentation on four CDFSS datasets, achieving the best average results it compares against.

What carries the argument

The load-bearing mechanism is a frozen pretrained diffusion inpainting model steered by two controls rather than retrained. A noise-strength scalar $\epsilon$ caps the forward noising at $T=\epsilon T_{\max}$ (with $T_{\max}=1000$), so the reverse pass reconstructs the expert-domain image instead of re-synthesizing it from Gaussian noise; low noise preserves the target content and mitigates the visual gap. The ground-truth bounding box defines an inpainting mask: background inpainting regenerates everything outside the box, exploiting the claim that background semantics transfer across domains, while foreground inpainting regenerates the box region under an LLM-generated category prompt. A selection module then feeds candidate images through a pretrained RPN and keeps the top-$k$ samples by average IoU between predicted and ground-truth boxes, discarding poor generations before they reach the detector. The paper uses CKA (Centered Kernel Alignment) similarity between original and synthesized images as its diagnostic for choosing these design points.

What would settle it

Take a fixed support set, generate SITN candidates, and rank them by CKA to the original; add only high-CKA images to the support set in one condition and only low-CKA images in another, holding selection size fixed. If the low-CKA condition matches or beats the high-CKA condition in detector mAP, the CKA premise is false and the observed gains must come from the selection module or from the added data volume rather than from the paper's gap analysis.

Watch

Extended reading notes

Core claim

The paper's central claim is that diffusion-based data augmentation can succeed in cross-domain few-shot object detection when the domain gap is decomposed. It claims that standard diffusion pipelines destroy target-domain information because the model mistakes expert-domain content for noise, and that adding only weakened noise lets the frozen generator preserve the semantic content of the original image instead of replacing it with source-domain material. It further claims that unknown target categories live mostly in the foreground, while backgrounds are shared across domains, so inpainting the background with the pretrained model produces images that stay category-consistent without requiring the model to know the novel class. On top of these two mechanisms, an RPN-based IoU filter selects which synthesized images actually enter the support set, because LLM prompts and foreground inpainting produce a mix of useful and useless candidates. If correct, the paper has a training-free, plug-and-play augmentation that raises detection mAP on six cross-domain benchmarks and segmentation accuracy on four cross-domain few-shot segmentation benchmarks beyond existing methods.

Load-bearing premise

The whole design of the Generation Module rests on the assumption that CKA similarity between a synthesized image and its original is a valid proxy for whether that image will raise detector mAP; if CKA rewards only low-level resemblance, the weakened-noise and background-inpainting choices are not independently grounded.

Editorial extensions

If this is right

  • Synthesized images from SITN are safe to add to support sets at 1-shot, 5-shot, and 10-shot settings, and Table 1 shows average mAP gains over every baseline backbone tested.
  • Background inpainting alone, without the selection module, already improves most datasets in the paper's ablation, which supports the claim that background transfer is the main driver rather than selection noise.
  • The method carries over to cross-domain few-shot segmentation, where the paper reports that applying SITN to the FPTrans baseline raises average accuracy from 62.2 to 66.7 across the four evaluated datasets.
  • The generation pipeline is training-free and samples with DDIM in at most 50 steps, giving per-image generation times around 0.02 to 1.5 seconds versus hundreds of seconds for standard DDPM or SDEdit, so the augmentation cost is small enough for routine use.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • My inference: because the method is training-free and needs only bounding boxes, it could be layered onto any few-shot detector that exposes support-set annotations, not just the detectors tested here.
  • My inference: the background-inpainting step effectively re-renders the scene around the few-shot instances, so it may behave as a style regularizer; a direct test would compare SITN-augmented support sets with style-augmented sets at matched data volume.
  • My inference: the selection module's IoU threshold could be tuned per class rather than per dataset, which might extend the gains to the hardest categories where average mAP remains low in the paper.
  • My inference: as stronger open-source generators appear, the weakened-noise schedule and IoU filtering may be the durable part of the method; the paper's own swap from SDv1.5 to Nucleus-Image is a step toward testing that.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. The paper proposes SITN, a training-free diffusion-based augmentation method for cross-domain few-shot object detection (CDFSOD). The method decomposes the domain gap into visual and semantic components, uses weakened noise (epsilon) to preserve target-domain information, performs foreground/background inpainting with LLM-generated prompts and bounding-box masks, and filters the generated samples with a top-k RPN-IoU selection module. Experiments on six CDFSOD datasets report average mAP improvements over prior state of the art in 1-shot, 5-shot, and 10-shot settings, and the method is also applied to four cross-domain few-shot segmentation datasets.

Significance. If the reported gains hold up, this is a useful and timely contribution: it provides the first systematic demonstration that frozen diffusion priors can be adapted to expert domains for few-shot detection and segmentation without training, through weakened noise and background inpainting. The paper gives explicit credit to the difficulty of the problem, and the background-inpainting-only ablation in Table 3 is a genuine piece of independent evidence that the augmentation itself (not only the selection step) can help. The claimed computational efficiency and compatibility with multiple detectors are also valuable. However, the empirical evidence as presented is not yet strong enough to support the headline state-of-the-art claim, because the main results lack variance estimates, the per-dataset hyperparameters are deferred to an absent appendix, and the selection module is not fully separated from the circularity concern of filtering by the same detector that is later fine-tuned.

major comments (6)
  1. [Section 5.1, Table 1] The headline SOTA claim rests on average margins of 0.7, 1.0, and 1.5 mAP over DomainRAG in the 1-shot, 5-shot, and 10-shot columns, but the paper reports no error bars, number of seeds, or statistical significance tests. In few-shot fine-tuning such differences are commonly within seed-to-seed variance. Please report mean +/- standard deviation over at least three random seeds, or provide paired per-dataset comparisons, for the main tables.
  2. [Section 4.3, Eq. (3) and Eq. (6)] The two method hyperparameters, the noise strength epsilon and the top-k count k, are stated to be 'tailored to each target domain dataset' and their values are deferred to the supplementary material, but no appendix or supplement is present in the submitted manuscript. Without a stated validation protocol it is impossible to rule out that these values were selected using the query set. Please report the exact epsilon and k values per dataset, describe the validation split used to choose them, and include a sensitivity analysis over a reasonable range of both hyperparameters.
  3. [Section 4.3, Eq. (6); Section 5.4, Table 3] The selection module keeps generated images whose RPN proposals, produced by the baseline CD-ViTO detector, have high IoU with ground-truth boxes. Because the same detector is later fine-tuned on the selected samples, part of the measured gain may come from choosing images that the detector already recognizes rather than from genuinely new training signal. The background-only ablation partially mitigates this concern, but the paper should quantify the selection bias, for example by comparing with selection based on a different detector or with random selection of the same number of images. The table also needs an explanation for why Fore.+Back. (22.5 on ArTaxOr, and lower than Back. alone in most rows) is consistently worse than Back. alone.
  4. [Section 3.1-3.3, Figs. 2-4] The paper uses CKA similarity as the main diagnostic to justify weakened noise and background inpainting, but it never validates that higher CKA between synthesized and original images translates into higher downstream detection mAP. Without such a validation, the design choices in the Generation Module are not empirically grounded; the observed mAP gains could in principle be explained by the selection module alone. Please add an experiment or scatter plot relating CKA values to mAP across noise strengths and inpainting strategies, or explicitly discuss the limitations of CKA as a proxy.
  5. [Table 2] The reported SDEdit average of 13.2 is inconsistent with the row values (35.5, 41.5, 13.1, 35.7, 9.8, 13.7), whose mean is 24.9. This arithmetic error directly affects the comparison with diffusion-based baselines and must be corrected. Please also verify all average values in Tables 1 and 4, since at least one additional average in Table 2 appears suspect.
  6. [Section 5.4, Table 4] The abstract and Section 5 claim that SITN also achieves state-of-the-art results on cross-domain few-shot segmentation, but the paper never specifies how the method is adapted to segmentation: the bounding-box masks, RPN-IoU selection, and detection fine-tuning protocol are all detection-specific. Please describe the segmentation adaptation protocol, including how masks are obtained and how selected images are integrated into the segmentation fine-tuning, or temper the CDFSS claim to a preliminary transfer result.
minor comments (5)
  1. [Abstract, Keywords] The abstract contains a grammatical error ('Our codes is available') and the keyword list includes 'Objection-detection', which should be 'Object Detection'.
  2. [Section 4.3, Eq. (1)] The LLM used for prompt generation is not named, and no details are given about the prompting strategy or the model version; this makes the generation module hard to reproduce exactly.
  3. [Section 5.2, Section 5.4] The implementation details are partially deferred to an appendix that is not included in the submitted manuscript, and the caption of Table 3 uses the abbreviation 'Sele.' without defining it. Please make the experimental protocol self-contained.
  4. [Fig. 12] The computational-efficiency comparison reports generation times for DDPM, SDEdit, DomainRAG, and the proposed method, but no hardware or implementation details are given, and DomainRAG's time includes retrieval steps that are not directly comparable to pure diffusion generation. Please specify the measurement conditions.
  5. [Section 5.4, 'CKA vs. MMD'] The paragraph states that MMD trends are 'perfectly symmetric' with CKA, but Fig. 11 is not described in enough detail to verify the claim; please report the MMD values and the exact feature representation used.

Circularity Check

1 steps flagged · score 4.0 of 10

Selection Module filters synthetic images using the baseline CD-ViTO detector's own RPN IoU, so a substantial part of the reported mAP gain is self-confirmation of the baseline rather than evidence that the synthesized data add new cross-domain information; background-inpainting-only ablations provide partial independent support.

  1. fitted input called prediction [Section 4.3, Selection Module (Eq. 6); Section 5.4, Selection Module ablation; Table 3]
    "We employ the RegionProposalNetwork (RPN) pre-trained on the CD-ViTO model (Baseline) to generate candidate bounding boxes (A) for the extended support set. We calculate the average IoU between predicted boxes and ground truth ones (Y), and only synthesized images with high IoUs are selected. ... These carefully selected samples, combined with the original training data for object detection model finetuning, lead to a substantial overall performance enhancement."

    The 'quality' of each synthesized image is defined by the RPN of the very detector (CD-ViTO) that is later fine-tuned and evaluated: selected images are, by construction, the ones the baseline already recognizes with high IoU. Feeding those self-selected easy positives back into the same detector can inflate mAP by reinforcing existing predictions rather than by supplying novel cross-domain training signal. Table 3 shows the selection step contributes a large share of the gain (ArTaxOr: Fore.+Back. 22.5 -> Fore.+Back.+Sele. 26.7 versus baseline 21.0), and the background-inpainting-only row (24.9) is the main independent evidence. Thus part of the 'helpful data' claim reduces to the baseline's own recognition score, though not the whole method.

full rationale

The paper's core generation design is motivated by CKA analyses, but the final evidence is mAP on held-out query sets and the background-inpainting-only ablation improves several datasets (e.g., ArTaxOr 21.0 to 24.9), so the generation module is not equivalent to its inputs by construction. The main circularity is confined to the Selection Module: Eq. (6) selects generated images using an RPN pre-trained on the same CD-ViTO baseline that is then fine-tuned and evaluated, making the 'high-quality sample' label a function of the baseline's own detections. This is a partial self-referential loop rather than a full derivation collapse, because selection is not the only contributor and final evaluation is on query mAP. No load-bearing self-citation was found: references to the authors' prior work (e.g., [22], [29], [51]–[54]) are related-work context, not imported proofs or uniqueness theorems. The missing appendix values for per-dataset noise strength epsilon and top-k, and the deferred fine-tuning details, are reproducibility and validation risks (possible selection on query), but not by themselves circularity.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claim rests on empirical modeling premises rather than new theoretical entities. The main premises are that CKA similarity predicts downstream usefulness, that backgrounds transfer better than foregrounds, and that weakened noise preserves task-relevant information. The tuning degrees of freedom are the noise-strength scalar and the top-k selection count, both per-dataset and not specified in the present text.

free parameters (2)
  • noise strength scalar epsilon = not reported in main text; varied in Fig. 8
    Controls the fraction of forward diffusion steps (T = epsilon * T_max). The paper states epsilon < 1 but does not give the chosen value per dataset, deferring to a missing appendix.
  • top-k selection count k = not reported; 'tailored to each target domain dataset'
    Number of generated images kept after RPN-IoU filtering. Without fixed values or a validation-based rule, the reported mAP may reflect per-dataset tuning on the evaluation splits.
assumptions (6)
  • domain assumption CKA similarity between original and synthesized images is a valid proxy for whether synthesized images will help downstream detection.
    Sec 3.1-3.2 uses CKA to diagnose diffusion failure and to select noise strength; the paper does not correlate CKA with final mAP.
  • domain assumption Background regions are semantically more transferable across domains than foreground objects, so background inpainting avoids the semantic gap.
    Sec 3.3 and Fig. 4; load-bearing for the Generation Module's background inpainting strategy.
  • domain assumption Adding weakened noise (epsilon < 1) keeps the diffusion model from confusing noise with useful target-domain information.
    Sec 3.2 and Eq. (3); the central fix for the visual gap.
  • domain assumption Inpainting within the bounding-box mask preserves object location, so original ground-truth labels remain valid for generated images.
    Sec 4.3, Eq. (4)-(5); enables using synthetic images with original y_sup labels.
  • domain assumption A frozen general-domain diffusion model can serve as a generator for expert domains when given an LLM prompt and partially noised input.
    Sec 4.1-4.3; the method is training-free and relies on the pretrained model's prior.
  • domain assumption The LLM-generated text prompt faithfully describes the target image and is compatible with the frozen diffusion model.
    Sec 4.3, Eq. (1); bad prompts are acknowledged as a source of low-quality samples, motivating the selection module.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Free-Lunch Augmentation by Revisiting Diffusion-Based Data Generation for Cross-Domain Few-Shot Object Detection." pith.science (2026). https://pith.science/paper/3IQX5KW6

@misc{pith2026260804394,
  author       = {Pith},
  title        = {Pith review of: Free-Lunch Augmentation by Revisiting Diffusion-Based Data Generation for Cross-Domain Few-Shot Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3IQX5KW6}},
  note         = {Machine review of arXiv:2608.04394}
}
read the original abstract

Cross-Domain Few-Shot Object Detection (CDFSOD) aims to transfer knowledge from data-rich upstream generic domains to downstream expert domains using scarce training data, where the significant domain gap and data scarcity make it an unsolved challenge. To address this problem, we revisit a natural yet underexplored approach in CDFSOD: data augmentation, by directly synthesizing data through diffusion models to supplement limited training samples. However, due to large domain gaps, we find that current diffusion methods cannot produce good results, leading to performance even lower than using the original images. To address these limitations, we divide the domain gaps into visual gaps and semantic gaps for separate analysis. For the visual gap, we find that the diffusion model cannot distinguish noise from useful information on expert domains, which can be mitigated by adding weakened noise. For the semantic gap, we find that the background semantics shows much smaller gaps between domains than foreground semantics, and we can bridge this gap by background inpainting. Based on the above analysis, we propose a method (Selective Inpainting with Tailored Noise, SITN) to dynamically take different strategies for downstream data synthesis based on their different gaps from the general domain, including a Generation Module for adding tailored noise and a Selection Module to dynamically select the inpainting regions. Extensive experiments on 6 datasets of CDFSOD and 4 datasets of cross-domain few-shot segmentation (CDFSS) validate that we can synthesize helpful data, achieving new state-of-the-art performance. Our codes is available at https://github.com/zzzzj311-droid/Free-Lunch-SITN

Figures

Figures reproduced from arXiv: 2608.04394 by the authors.

Figure 1
Figure 1. (Left-top) Cross-Domain Few-Shot Object Detection learns from scarce training data on downstream expert domains, with knowledge generalized from upstream gen￾eral domains, where domain gaps and data scarcity make it challenging. (Left-bottom & Mid) To address it, we revisit a natural but ignored approach in CDFSOD (i.e., data augmentation) to directly supplement the scarce training data with diffusion models. Howeve… view at source ↗
Figure 2
Figure 2. (Left) Synthesized images on source domains well preserve the semantics of the original images, but largely destroy the semantics on the target domains. (Right) To quantify such a difference between domains, we use the CKA similarity to measure the semantic similarity between synthesized and original images. We can see the traditional method (e.g., flip, rotation) shows consistent CKA similarities across domains, ve… view at source ↗
Figure 3
Figure 3. The relationship between noise strength and CKA similarity shows that low noise strength effectively preserves the semantics of the original image (i.e., similarity only slightly decreases), while high noise levels severely disrupt semantics (i.e., gener￾ated images differ greatly from the original one), indicating diffusion models can hardly recover target-domain information from the noise. Specifically, diffusion … view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: (Left) Compared with the foreground relevant to unseen classes, the back￾ground is easier to transfer across domains. (Right) To verify this intuition, we crop the foreground and background for data synthesis and find that the background CKA is consistently higher than…
Figure 5
Figure 5. Figure 5: Inpainting at timestep t, where the intermediate result x known t−1 from the forward process and the prediction x unknown t−1 from the reverse process are jointly synthesized to generate the image xt−1 in the backward process. to sample new data xnew from it [31]. Mean…
Figure 6
Figure 6. Figure 6: Our method operates solely on the target domain and consists of three stages. In the first stage, we feed images into an LLM to obtain prompts, extract bounding boxes from annotations, and compute their masks. In the second stage, the prompts, original images, and boun…
Figure 7
Figure 7. Figure 7: Images generated by foreground and background inpainting, with the bounding box (yellow). mask is determined by the BBox, the inpainted images retain the original object positions without requiring re-annotation(As shown in [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Ablation study of tailored noise. The legends in different colors indicate different noise strengths. sion Transformer-B (ViT-B), and DETR-R101, under 1-shot, 5-shot, and 10-shot configurations. The benchmark approaches include CDMM-FSOD [37], ViTDeT￾FT [25], DE-ViT-FT…
Figure 10
Figure 10. Figure 10: Ablation study of occlusion scenes. Comparision with other open-source diffusion model As shown in Tab. 6, we use Nucleus-Image [41] for generation. Directly applying Nucleus-Image yields poor performance. However, Nucleus-Image with our method performs compa￾rably to…
Figure 12
Figure 12. Figure 12: The computational efficiency of the original diffusion model and our method. CKA vs. MMD Moreover, the CKA similarity comparison is shown in [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 35 canonical work pages

  1. [1]

    arXiv preprint arXiv:2101.091932(3) (2021)

    Bevandic, P., Krešo, I., Oršic, M., Šegvic, S.: Dense outlier detection and open- set recognition based on training with noisy negative images. arXiv preprint arXiv:2101.091932(3) (2021)

    arXiv 2021

  2. [2]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Bou, X., Facciolo, G., Von Gioi, R.G., Morel, J.M., Ehret, T.: Exploring robust features for few-shot object detection in satellite imagery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 430–439 (2024)

    work page 2024

  3. [3]

    Boudiaf, M., Kervadec, H., Masud, Z.I., Piantanida, P., Ben Ayed, I., Dolz, J.: Few-shot segmentation without meta-learning: A good transductive inference is all you need? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13979–13988 (2021)

    2021

  4. [4]

    IEEE Transactions on Knowledge and Data Engineering (2024)

    Cao, H., Tan, C., Gao, Z., Xu, Y., Chen, G., Heng, P.A., Li, S.Z.: A survey on gen- erative diffusion models. IEEE Transactions on Knowledge and Data Engineering (2024)

    work page 2024

  5. [5]

    In: Proceedings of the ieee/cvf international conference on computer vision

    Chan, R., Rottmann, M., Gottschalk, H.: Entropy maximization and meta classifi- cation for out-of-distribution detection in semantic segmentation. In: Proceedings of the ieee/cvf international conference on computer vision. pp. 5128–5137 (2021)

    2021

  6. [6]

    Davari, M., Horoi, S., Natik, A., Lajoie, G., Wolf, G., Belilovsky, E.: Reliability of cka as a similarity measure in deep learning (Oct 2022)

    work page 2022

  7. [7]

    In: European conference on computer vision

    Fan, Q., Pei, W., Tai, Y.W., Tang, C.K.: Self-support few-shot semantic segmen- tation. In: European conference on computer vision. pp. 701–719. Springer (2022)

    work page 2022

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Genera- tive diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9935–9946 (2023)

    work page 2023

Show all 54 references
  1. [9]

    ACM Computing Surveys57(9), 1–66 (2025)

    Foo,L.G.,Rahmani,H.,Liu,J.:Ai-generatedcontent(aigc)forvariousdatamodal- ities: A survey. ACM Computing Surveys57(9), 1–66 (2025)

    2025

  2. [10]

    In: European Conference on Computer Vision

    Fu, Y., Wang, Y., Pan, Y., Huai, L., Qiu, X., Shangguan, Z., Liu, T., Fu, Y., Van Gool, L., Jiang, X.: Cross-domain few-shot object detection via enhanced open-set object detector. In: European Conference on Computer Vision. pp. 247–

  3. [11]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2918–2928 (2021)

    2021

  4. [12]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Go,H.,Lee,Y.,Kim,J.Y.,Lee,S.,Jeong,M.,Lee,H.S.,Choi,S.:Towardspractical plug-and-play diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1962–1971 (2023)

    2023

  5. [13]

    Gui, G., Gao, B.B., Liu, J., Wang, C., Wu, Y.: Few-shot anomaly-driven generation foranomalyclassificationandsegmentation.In:EuropeanConferenceonComputer Vision. pp. 210–226 (2024)

    2024

  6. [14]

    Han,G.,Lim,S.N.:Few-shotobjectdetectionwithfoundationmodels.In:Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28608–28618 (2024)

    2024

  7. [15]

    He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is syn- thetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574 (2022) Revisiting Diffusion-Based Data Generation for CDFSOD 17

    2022 arXiv

  8. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    He, W., Zhang, Y., Zhuo, W., Shen, L., Yang, J., Deng, S., Sun, L.: Apseg: Auto- prompt network for cross-domain few-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23762–23772 (2024)

    2024

  9. [17]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Herzog, J.: Adapt before comparison: A new perspective on cross-domain few-shot segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 23605–23615 (2024)

    2024

  10. [18]

    Neural In- formation Processing Systems,Neural Information Processing Systems (Jan 2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Neural In- formation Processing Systems,Neural Information Processing Systems (Jan 2020)

    2020

  11. [19]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Huang, J., Wu, Q., Ren, Y., Yang, F., Yang, A., Yang, Q., Pu, X.: Sparse bayesian deep learning for cross domain medical image reconstruction. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2339–2347 (2024)

    2024

  12. [20]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Islam, K., Zaheer, M.Z., Mahmood, A., Nandakumar, K.: Diffusemix: Label- preserving data augmentation with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27621– 27630 (2024)

    2024

  13. [21]

    In: European Conference on Computer Vision

    Jabbour, S., Kondas, G., Kazerooni, E., Sjoding, M., Fouhey, D., Wiens, J.: De- pict: Diffusion-enabled permutation importance for image classification tasks. In: European Conference on Computer Vision. pp. 35–51. Springer (2025)

    2025

  14. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Jiang, Y., Zou, Y., Li, Y., Li, R.: Remedying target-domain astigmatism for cross- domain few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19580–19590 (2026)

    2026

  15. [23]

    In: European conference on computer vision

    Lei, S., Zhang, X., He, J., Chen, F., Du, B., Lu, C.T.: Cross-domain few-shot semantic segmentation. In: European conference on computer vision. pp. 73–90. Springer (2022)

    2022

  16. [24]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10965–10975 (2022)

    2022

  17. [25]

    In: European conference on computer vision

    Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European conference on computer vision. pp. 280–296. Springer (2022)

    2022

  18. [26]

    arXiv preprint arXiv:2506.05872 (2025)

    Li, Y., Qiu, X., Fu, Y., Chen, J., Qian, T., Zheng, X., Paudel, D.P., Fu, Y., Huang, X., Van Gool, L., et al.: Domain-rag: Retrieval-guided compositional image gener- ation for cross-domain few-shot object detection. arXiv preprint arXiv:2506.05872 (2025)

    2025

  19. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liu, J., Wang, Q., Fan, H., Wang, Y., Tang, Y., Qu, L.: Residual denoising diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2773–2783 (2024)

    2024

  20. [28]

    arXiv preprint arXiv:2303.05499 (2023)

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)

    2023 arXiv

  21. [29]

    In: Proceedings of the Computer Vision and Pattern Recog- nition Conference

    Liu, Y., Zou, Y., Li, Y., Li, R.: The devil is in low-level features for cross-domain few-shot segmentation. In: Proceedings of the Computer Vision and Pattern Recog- nition Conference. pp. 4618–4627 (2025)

    2025

  22. [30]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Re- paint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11461– 11471 (2022)

    2022

  23. [31]

    arXiv preprint arXiv:2208.11970 (2022) 18 Z

    Luo, C.: Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970 (2022) 18 Z. Zhuang, Y. Zou, et al

    2022 arXiv

  24. [32]

    arXiv preprint arXiv:2108.01073 (2021)

    Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)

    2021 arXiv

  25. [33]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Min, J., Kang, D., Cho, M.: Hypercorrelation squeeze for few-shot segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6941–6952 (2021)

    2021

  26. [34]

    In: Proceedings of the Computer Vision and Pattern Recog- nition Conference

    Pan, J., Liu, Y., He, X., Peng, L., Li, J., Sun, Y., Huang, X.: Enhance then search: An augmentation-search strategy with foundation models for cross-domain few- shot object detection. In: Proceedings of the Computer Vision and Pattern Recog- nition Conference. pp. 1548–1556 (2025)

    2025

  27. [35]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

    2022

  28. [36]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Sain, A., Bhunia, A.K., Chowdhury, P.N., Koley, S., Xiang, T., Song, Y.Z.: Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2765–2775 (2023)

    2023

  29. [37]

    In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Shangguan, Z., Seita, D., Rostami, M.: Cross-domain multi-modal few-shot object detection via rich text. In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 6570–6580. IEEE (2025)

    2025

  30. [38]

    In: Pro- ceedings of the AAAI Conference on Artificial Intelligence

    Shen, J., Kuang, K., Wang, J., Wang, X., Feng, T., Zhang, W.: Cgmgm: A cross- gaussian mixture generative model for few-shot semantic segmentation. In: Pro- ceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 4784–4792 (2024)

    2024

  31. [39]

    arXiv: Learn- ing,arXiv: Learning (Oct 2020)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv: Learn- ing,arXiv: Learning (Oct 2020)

    2020

  32. [40]

    In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition

    Su, J., Fan, Q., Pei, W., Lu, G., Chen, F.: Domain-rectifying adapter for cross- domain few-shot segmentation. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 24036–24045 (2024)

    2024

  33. [41]

    Team, N.A.: Nucleus-image: Sparse moe for image generation (2026)

    2026

  34. [42]

    arXiv preprint arXiv:2302.07944 (2023)

    Trabucco, B., Doherty, K., Gurinas, M., Salakhutdinov, R.: Effective data augmen- tation with diffusion models. arXiv preprint arXiv:2302.07944 (2023)

    2023 arXiv

  35. [43]

    In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Xiong, W.: Cd-fsod: A benchmark for cross-domain few-shot object detection. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2023)

    2023

  36. [44]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Yan, Q., Chen, G., Zou, Y.: Start small, think big: Curriculum-based relative policy optimization for visual grounding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 11550–11558 (2026)

    2026

  37. [45]

    arXiv preprint arXiv:2308.06721 (2023)

    Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compati- ble image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)

    2023 arXiv

  38. [46]

    Advances in neural information processing systems35, 6575–6588 (2022)

    Zhang, J.W., Sun, Y., Yang, Y., Chen, W.: Feature-proxy transformer for few-shot segmentation. Advances in neural information processing systems35, 6575–6588 (2022)

    2022

  39. [47]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3836–3847 (2023)

    2023

  40. [48]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhang, R., Hu, X., Li, B., Huang, S., Deng, H., Qiao, Y., Gao, P., Li, H.: Prompt, generate,thencache: Cascadeoffoundationmodelsmakesstrongfew-shotlearners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15211–15222 (2023) Revisiting...

    2023

  41. [49]

    arXiv preprint arXiv:2305.03048 (2023)

    Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Ma, X., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)

    2023 arXiv

  42. [50]

    arXiv preprint arXiv:2309.12969 (2023)

    Zhang, X., Liu, Y., Wang, Y., Boularias, A.: Detect everything with few examples. arXiv preprint arXiv:2309.12969 (2023)

    2023 arXiv

  43. [51]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Zhao, Y., Zou, Y., Li, Y., Li, R.: Interpretable cross-domain few-shot learning with rectified target-domain local alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 41605– 41615 (June 2026)

    2026

  44. [52]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zou, Y., Liu, Y., Hu, Y., Li, Y., Li, R.: Flatten long-range loss landscapes for cross-domain few-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23575–23584 (2024)

    2024

  45. [53]

    Advances in Neural Information Processing Systems37, 116332–116354 (2024)

    Zou, Y., Ma, R., Li, Y., Li, R.: Attention temperature matters in vit-based cross- domain few-shot learning. Advances in Neural Information Processing Systems37, 116332–116354 (2024)

    2024

  46. [54]

    arXiv preprint arXiv:2405.17022 (2024)

    Zou, Y., Zhang, S., Zhou, H., Li, Y., Li, R.: Compositional few-shot class- incremental learning. arXiv preprint arXiv:2405.17022 (2024)

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.