Pith. sign in

REVIEW 4 major objections 7 minor 43 references

Transformed synthetic images lift drone-view human detection by up to +14.1 mAP@50 through a three-stage diffusion pipeline that aligns both global style and local content.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 16:18 UTC pith:RHMCTHNV

load-bearing objection A useful three-stage synthetic-to-real recipe whose headline +14.1 mAP gain is likely inflated by unexamined selection bias in the hallucination-removal stage. the 4 major comments →

arxiv 2512.13869 v3 pith:RHMCTHNV submitted 2025-12-15 cs.CV

Coarse-to-Fine Hierarchical Alignment for UAV-based Human Detection using Diffusion Models

Wenda Li , Meng Wu , Liangzhao Chen , Sungmin Eum , Heesung Kwon , Qing Qu This is my paper
classification cs.CV
keywords Sim2Real domain adaptationUAV human detectiondiffusion modelsstyle transferhallucination removalCLIP-guided filteringsynthetic datahierarchical alignment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the Sim2Real gap for UAV human detection is not one shift but two: a global style gap (lighting, color, texture) and a local content gap (the size, density, and visual realism of the people themselves). It proposes CFHA, a three-stage pipeline that first transfers global style with a diffusion model, then refines small human patches with a one-step super-resolution diffusion module, then filters out synthetic-looking people using a CLIP-guided similarity score. Trained on 500 synthetic plus 20 real images, detectors using the transformed data improve mAP@50 by up to +14.1 on Semantic-Drone over the untransformed baseline. The paper's core message is that object-level realism around the instances matters more than whole-image realism for detection.

Core claim

The central discovery claimed is that jointly aligning global style and local content is what closes the Sim2Real gap, and that diffusion models can do both while leaving annotations intact. The ablation shows that Global Style Transfer alone can hurt detection by introducing artifacts on small people, Local Refinement recovers and improves it, and Hallucination Removal delivers the largest jump, especially on Semantic-Drone. The paper concludes that patch-wise FID around human instances is more predictive of detection performance than image-wise FID.

What carries the argument

CFHA's load-bearing mechanism is the explicit decomposition into three stages. Stage I uses latent-space style transfer with adaptive instance normalization and cross-attention to align color, illumination, and haze while preserving geometry. Stage II applies a one-step super-resolution diffusion model, fine-tuned with low-rank adapters and a caption prompt, to object-centered patches, restoring fine detail and removing artifacts. Stage III builds a CLIP-based prototype vector t* = normalize(mean real-human embedding + λ text anchor), scores each synthetic person crop by cosine similarity to it, and stochastically removes low-scoring instances. The detector is then trained on the transformed

Load-bearing premise

The load-bearing premise is that a synthetic person crop whose similarity to a prototype 'real person' vector (built from a few real examples plus a text description) is low counts as an artifact, and deleting those crops from the training set — rather than transforming them — is what improves detection; this filter carries most of the headline gain, and its two tuning knobs are set by hand.

What would settle it

Run the same detector-training setup but delete the same number of person instances at random from each synthetic image instead of using CLIP-guided deletion. If mAP@50 stays at the CFHA level, the HR gain is explained by density/content alignment rather than by the CLIP prototype's realism judgment. Alternatively, re-run Stage II with a standard non-diffusion super-resolution upscaler; if the LR gain vanishes, the gain is specific to the fine-tuning and data recipe, not to super-resolution per se.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Detectors trained on CFHA-transformed synthetic data beat the untransformed baseline on four UAV benchmarks, with the largest gain on Semantic-Drone (+14.1 mAP@50, +6.7 mAP@50-95).
  • Because transformations preserve the original bounding-box labels, CFHA is a few-shot Sim2Real recipe: it needs only about 20 real images to anchor style and content.
  • The ablation indicates that global style transfer alone can degrade detection, while local patch refinement is the stage that recovers performance — pointing to instance-level realism as the key factor.
  • Interpolating the background between the original render and the style-transferred version peaks at α=0.2, meaning a modest amount of style transfer is better than full transfer.
  • Combining real and transformed synthetic images during detector training stabilizes learning and yields the reported gains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The Hallucination Removal stage may be doing two things at once: removing visually synthetic people and reducing the extreme person-density mismatch between synthetic scenes and real drone views; a random-subsampling control experiment would separate these effects.
  • Because the pipeline is label-preserving, CFHA could transfer to other small-object detection tasks (vehicles, animals) with the same three stages, provided the CLIP prototype and caption prompts are re-anchored to the new category.
  • The drop from full style transfer and the peak at α=0.2 suggests a regularization story: target-style background helps only up to a point, after which geometric distortion hurts; this could be formalized as a bias-variance trade-off over style strength.
  • The one-step super-resolution stage's contribution is currently entangled with the choice of the super-resolution checkpoint and the undescribed training data for its fine-tuning; an ablation that replaces it with a standard non-diffusion upscaler would clarify what the diffusion formulation adds.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes CFHA, a three-stage generative translation pipeline for UAV human detection. Stage I uses DDIM inversion and AdaIN in Stable Diffusion latents to transfer global style from a small set of real images; Stage II applies a one-step super-resolution diffusion model with LoRA and caption prompting to refine local human crops; Stage III uses a CLIP-based prototype filter (Eqs. 7-9) to probabilistically delete human instances whose cosine similarity to a target-domain prototype is low. Detectors (YOLO11x) are then trained on 20 real images plus 500 transformed SynPlay images. The authors report +14.1 mAP@50 and +6.7 mAP@50-95 on Semantic-Drone, +3.6/+2.9 on Okutama, and marginal gains on VisDrone and Manipal-UAV. Ablations show that GST alone can hurt, LR recovers performance, and HR provides the dominant Semantic-Drone gain.

Significance. If substantiated, CFHA would provide a practical few-shot recipe for UAV person detection: generate realistic synthetic training data with a diffusion pipeline while preserving labels, then train a standard detector on those data. The paper's distinction between global style and local content gaps is well motivated, and the FID analysis in Table 2—showing that patch-level realism tracks detection performance better than image-level FID—is a useful observation. The diffusion/AdaIN formulation is standard and coherent, the ablation design is informative, and the code release is a positive step for reproducibility. However, the current evidence is not sufficient to support the central claim. The headline Semantic-Drone gain is dominated by an uncontrolled stochastic deletion filter, the LR stage's training protocol is not described, and one ablation result in Table 3 exceeds the full pipeline, creating an internal inconsistency that must be resolved before the hierarchical-alignment explanation can be accepted.

major comments (4)
  1. [§4.4, Table 2] The HR module is responsible for the majority of the reported Semantic-Drone gain: mAP@50 rises from 21.2 (GST+LR) to 32.7 (GST+LR+HR), i.e., +11.5 of the total +14.1. Yet the paper provides no control that isolates CLIP-similarity-based deletion from simply discarding hard examples. Low CLIP similarity plausibly correlates with small, occluded, or unusual human instances, and dropping them from the training set can mechanically improve test mAP by removing hard positives. The sharpness α in Eq. (9) and the fraction of deleted instances are never reported. The authors should report the deletion rate and add at least three controls: random deletion of the same number of instances, deletion of the highest-scoring instances (inverse filter), and a detector trained on the removed crops. Without these, the headline result cannot be attributed to distribution alignment.
  2. [§4.4, Eq. (9)] The stochastic retention rule is not well defined. The text says a crop k is retained with probability p_k = exp(α s_k) / Σ_j exp(α s_j), but the denominator is not specified: is the sum over all instances in the dataset, over the current mini-batch, or over the current image? If the sum is over the whole dataset, then p_k is not a marginal probability and the procedure cannot be implemented online as described; if it is over a mini-batch, the retention probability of a crop changes with batch composition. The paper should give the exact sampling scheme, including how normalization is performed and how many instances are removed in practice. This matters because the HR deletion rule is the load-bearing component of the Semantic-Drone improvement.
  3. [§4.3] The Local Refinement stage is not reproducible as specified. The authors say the super-resolution diffusion model is fine-tuned with rank-4 LoRA and a VSD loss, but they never describe the training data: which synthetic/real image pairs are used, how many pairs, how many iterations, what learning rate, what the data term and VSD loss are, or how the mask restricting edits to human instances is obtained. The up-sampling factor is stated as s=2, but the relation to the reported detection experiments is unclear. Because LR is claimed to recover the damage caused by GST (Table 2), the missing training protocol is a load-bearing gap for the 'global then local refinement' claim.
  4. [§5.4, Table 3] Table 3 reports that background interpolation with α=0.2 yields mAP@50 of 73.1 and mAP@50-95 of 27.0 on Okutama, which is higher than the full CFHA pipeline's 68.2 and 25.3 in Table 2. This is an internal inconsistency that needs explanation. Is the interpolation applied to the original synthetic data, to GST output, before or after LR/HR? Why is this configuration not part of the final CFHA if it outperforms it on Okutama? As presented, Table 3 undermines the claim that all three stages are complementary and that the full pipeline is the best configuration. The authors should clarify the experimental setup and either reconcile the numbers or explain the difference.
minor comments (7)
  1. [§2.1] The framework is accidentally referred to as 'HSRA' in the last paragraph of Section 2.1; it should be CFHA.
  2. [Table 1] Manipal-UAV results are reported as decimals (0.478, 0.538, 0.183, 0.205) while all other results are percentages. Use a consistent scale.
  3. [§4.3, Eq. (5)] The one-step update defines z_H = z_L - β_T ε(z_L;T,c_y)/α_T without stating the relationship between α_T, β_T and the DDPM/DDIM schedule. Also, the symbols α_T/β_T can be confused with the later interpolation weight α in Table 3.
  4. [§5.1] The paper states that 20 real images from the target domain are used for training, but it does not state how those images are selected or whether they are disjoint from the test split. This is essential for interpreting the reported gains.
  5. [§4.2/§4.4] Several components are named but not defined: DAPE, the VSD loss, the exact mask-guided deletion procedure using AttentiveEraser, and the interpolation operation in Table 3. Please provide formal descriptions or precise references.
  6. [§2, Related Work] PTL [10] is described as a relevant prior method, but no quantitative comparison against PTL or other Sim2Real baselines is reported. A comparison table would strengthen the state-of-the-art claim.
  7. [§5.3] The VisDrone result (+0.5 mAP@50) is reported as 'marginal improvement' with no analysis of variance or statistical significance. Please report results over multiple seeds or justify the single-run protocol.

Circularity Check

0 steps flagged

No circularity found: the mAP gains are empirical downstream measurements, not quantities equivalent to the method's fitted inputs.

full rationale

I walked the claimed derivation chain. Stage I (GST) is a training-free diffusion/AdaIN transform (Eqs. 1-4); Stage II (LR) is a LoRA fine-tuned one-step SR (Eqs. 5-6); Stage III (HR) uses a CLIP prototype t* = normalize(ū + λ t_anchor) (Eqs. 7-8) only to decide which synthetic crops to retain during training (Eq. 9). The headline +14.1 mAP@50 is obtained by training YOLO11x on the transformed set and evaluating on the benchmark; mAP is not algebraically determined by the filter equations, and the filter does not optimize mAP. The prototype is estimated from real images used for adaptation, which is a standard domain-adaptation input, not a fitted 'prediction' of the test labels. The self-citations (Synplay, PTL, model-collapse) provide data/baselines/context, not the core derivation. The main concerns—unreported sharpness α, no random-deletion control for HR, and possible selection of interpolation weights from test mAP—are correctness/reproducibility risks, not circular reductions, and do not make any equation equal to another by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The method is mostly assembly of external tools (diffusion prior, OSEDiff, CLIP), so there are no invented physical entities; the invented quantity is the prototype vector, which is calibrated on the target set and validated on the target benchmark. The effective free parameters (λ, two α's, data sizes) are the ones that turn the headline +14.1 on/off, and several are not reported.

free parameters (6)
  • λ (text-anchor weight in HR prototype) = 0.2
    Hand-set hyperparameter in Eq. (7)-(8); controls how much the real-data prototype is pulled toward the text anchor 'a photo of a person taken from a drone'; no sensitivity analysis or selection protocol reported.
  • α (HR sampling sharpness) = not reported
    Controls how greedily the stochastic filter deletes low-similarity humans (Eq. 9); its value is never given although it determines how much of the training set is discarded, which drives the Semantic-Drone gain.
  • α (background interpolation weight, Table 3) = 0.2 (peak)
    Ablation sweep 0→1, peak at 0.2, with no validation protocol; this configuration reportedly beats the full pipeline on Okutama (73.1 vs 68.2) yet is not used in the final method.
  • λ_orig, λ_tran (detector dual-training weights) = 1, 1
    Weights in Eq. (10); authors note they 'could also be tuned' but do not report a selection procedure.
  • LoRA rank / upsampling factor for LR = rank 4; s=2
    Architecture choices for Stage II (§4.3), stated without justification or ablation.
  • Training set sizes = 500 synthetic + 20 real images
    Chosen per experiment (§5.1); the real-reference images used in GST and HR are drawn from the same pool, conflating several roles.
axioms (6)
  • standard math Diffusion/DDIM background (forward SDE, score matching) is valid for image translation.
    Invoked throughout §3.3-§4.3; inherited from Ho et al. [17], Song et al. [31].
  • domain assumption DDIM inversion + AdaIN in latent space preserves object geometry so bounding-box labels survive style transfer.
    Central premise of Stage I (§4.2, Eq. 1-4): the paper asserts labels are unchanged with no label-consistency measurement.
  • domain assumption CLIP cosine similarity to a text-anchored prototype measures whether a human instance looks 'real'.
    Premise of Stage III (§4.4, Eqs. 7-9); the filter's deletion decisions and, hence, the headline gain rest on this semantic-similarity proxy.
  • domain assumption Patch-wise FID around instances, rather than image-wise FID, is predictive of detection performance.
    Used to justify LR despite image-wise FID worsening (§5.4); the correlation is asserted from two data points rather than demonstrated.
  • domain assumption One randomly sampled real image is a sufficient style reference for each synthetic image.
    Stage I samples a random style example per content image (§4.2); no experiments vary the number/selection of style references.
  • ad hoc to paper The text anchor 'a photo of a person taken from a drone' is an appropriate regularizer for the real-human prototype.
    Introduced for this paper (§4.4, Eq. 7-8) with no evidence that this anchoring improves the filter relative to other prompts.
invented entities (1)
  • Prototype reference vector t* for human 'realness' no independent evidence
    purpose: Central tendency of real human crops in CLIP space (with text-anchor regularization) used to score and delete synthetic humans.
    t* = normalize(ū + λ t_anchor) is computed from 20 target images and validated only by downstream mAP on the same benchmarks; it has no falsifiable handle outside this paper's evaluation loop.

pith-pipeline@v1.3.0-alltime-deepseek · 11632 in / 19591 out tokens · 157154 ms · 2026-08-03T16:18:56.457833+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Coarse-to-Fine Hierarchical Alignment for UAV-based Human Detection using Diffusion Models." pith.science (2026). https://pith.science/paper/RHMCTHNV

@misc{pith2026251213869,
  author       = {Pith},
  title        = {Pith review of: Coarse-to-Fine Hierarchical Alignment for UAV-based Human Detection using Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RHMCTHNV}},
  note         = {Machine review of arXiv:2512.13869}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Training object detectors demands extensive, task-specific annotations, yet this requirement becomes impractical in UAV-based human detection due to constantly shifting target distributions and the scarcity of labeled images. As a remedy, synthetic simulators are adopted to generate annotated data, with a low annotation cost. However, the domain gap between synthetic and real images hinders the model from being effectively applied to the target domain. Accordingly, we introduce Coarse-to-Fine Hierarchical Alignment (CFHA), a three-stage diffusion-based framework designed to transform synthetic data for UAV-based human detection, narrowing the domain gap while preserving the original synthetic labels. CFHA explicitly decouples global style and local content domain discrepancies and bridges those gaps using three modules: (1) Global Style Transfer -- a diffusion model aligns color, illumination, and texture statistics of synthetic images to the realistic style, using only a small real reference set; (2) Local Refinement -- a super-resolution diffusion model is used to facilitate fine-grained and photorealistic details for the small objects, such as human instances, preserving shape and boundary integrity; (3) Hallucination Removal -- a module that filters out human instances whose visual attributes do not align with real-world data to make the human appearance closer to the target distribution. Extensive experiments on public UAV Sim2Real detection benchmarks demonstrate that our methods significantly improve the detection accuracy compared to the non-transformed baselines. Specifically, our method achieves up to $+14.1$ improvement of mAP50 on Semantic-Drone benchmark. Ablation studies confirm the complementary roles of the global and local stages and highlight the importance of hierarchical alignment. The code is released at \href{https://github.com/liwd190019/CFHA}{this url}.

Figures

Figures reproduced from arXiv: 2512.13869 by Heesung Kwon, Liangzhao Chen, Meng Wu, Qing Qu, Sungmin Eum, Wenda Li.

Figure 1
Figure 1. Figure 1: Overview of our proposed three-stage pipeline for synthetic-to-real image translation. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of transformation on Okutama-Action. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of Hallucination Removal (HR) on Okutama. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Human detection results comparison of Semantic Drone dataset. (a) Baseline: Detection results trained on original Synplay data. (b) Ours: Detection results trained on CHFA-transformed Synplay data. • Object(Human) removal further enhances the detector by aligning the content distributions. We further investigate how patch-wise FID and image-wise FID correlate with detection per￾formance. As shown in [PITH… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 3 linked inside Pith

  1. [1]

    Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C

    Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In European Conference on Computer Vision, 2014

    2014

  2. [2]

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale.International journal of computer vision, 2020

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale.International journal of computer vision, 2020

    2020

  3. [3]

    Training deep networks with synthetic data: Bridging the reality gap by domain randomization

    Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. InIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2018

    2018

  4. [4]

    Dreamsim: Learningnewdimensionsofhuman visual similarityusingsynthetic data

    Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and PhillipIsola. Dreamsim: Learningnewdimensionsofhuman visual similarityusingsynthetic data. InAdvances in Neural Information Processing Systems, 2023

    2023

  5. [5]

    Syn2real: A new benchmark forsynthetic-to-real visual domain adaptation.CoRR, abs/1806.09755, 2018

    Xingchao Peng, Ben Usman, Kuniaki Saito, Neela Kaushik, Judy Hoffman, and Kate Saenko. Syn2real: A new benchmark forsynthetic-to-real visual domain adaptation.CoRR, abs/1806.09755, 2018

    Pith/arXiv arXiv 2018

  6. [6]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    2022

  7. [7]

    SDXL: improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. InInternational Conference on Learning Representations, 2024

    2024

  8. [8]

    Convolution meets transformer: Efficient hybrid transformer for semantic segmentation with very high resolution imagery

    Yuji Wang, Ruojun Zhao, Shicai Wei, Jingchen Ni, Meng Wu, Yang Luo, and Chunbo Luo. Convolution meets transformer: Efficient hybrid transformer for semantic segmentation with very high resolution imagery. InIEEE International Geoscience and Remote Sensing Symposium, 2024

    2024

  9. [9]

    Uav- yolo: Small object detection on unmanned aerial vehicle perspective.Sensors, 2020

    Mingjie Liu, Xianhao Wang, Anjian Zhou, Xiuyuan Fu, Yiwei Ma, and Changhao Piao. Uav- yolo: Small object detection on unmanned aerial vehicle perspective.Sensors, 2020

    2020

  10. [10]

    Progressive transformation learning for leveraging virtual images in training

    Yi-Ting Shen, Hyungtae Lee, Heesung Kwon, and Shuvra S Bhattacharyya. Progressive transformation learning for leveraging virtual images in training. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  11. [11]

    Sim2real in reconstructive spectroscopy: Deep learning with augmented device-informed data simulation.APL Machine Learning, 2024

    Jiyi Chen, Pengyu Li, Yutong Wang, Pei-Cheng Ku, and Qing Qu. Sim2real in reconstructive spectroscopy: Deep learning with augmented device-informed data simulation.APL Machine Learning, 2024

    2024

  12. [12]

    A closer look at model collapse: From a generalization-to-memorization perspective.Advances in Neural Information Processing Systems, 2025

    Lianghe Shi, Meng Wu, Huijie Zhang, Zekai Zhang, Molei Tao, and Qing Qu. A closer look at model collapse: From a generalization-to-memorization perspective.Advances in Neural Information Processing Systems, 2025. 14

    2025

  13. [13]

    Sim2air - synthetic aerial dataset for UAV monitoring.IEEE Robotics and Automation Letter, 2022

    Antonella Barisic, Frano Petric, and Stjepan Bogdan. Sim2air - synthetic aerial dataset for UAV monitoring.IEEE Robotics and Automation Letter, 2022

    2022

  14. [14]

    Bi-directional domain adaptation for sim2real transfer of embodied navigation agents.IEEE Robotics Automation Letter, 2021

    Joanne Truong, Sonia Chernova, and Dhruv Batra. Bi-directional domain adaptation for sim2real transfer of embodied navigation agents.IEEE Robotics Automation Letter, 2021

    2021

  15. [15]

    Qiranul Saadiyean, S. P. Samprithi, and Suresh Sundaram. Learning multi-scale context mask- rcnn network for slant angled aerial imagery in instance segmentation in a sim2real setup. In IEEE International Conference on Robotics and Automation, 2024

    2024

  16. [16]

    The emergence of reproducibility and consistency in diffusion models

    Huijie Zhang, Jinfan Zhou, Yifu Lu, Minzhe Guo, Peng Wang, Liyue Shen, and Qing Qu. The emergence of reproducibility and consistency in diffusion models. InInternational Conference on Machine Learning, 2024

    2024

  17. [18]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  18. [19]

    Vip: Versatile image outpainting empowered by multimodal large language model

    Jinze Yang, Haoran Wang, Zining Zhu, Chenglong Liu, Meng Wu, and Mingming Sun. Vip: Versatile image outpainting empowered by multimodal large language model. InAsian Conference on Computer Vision, 2024

    2024

  19. [20]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. InIEEE/CVF conference on computer vision and pattern recognition, 2022

    2022

  20. [21]

    Sdedit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022

    2022

  21. [22]

    Exploring low- dimensional subspace in diffusion models for controllable image editing.Advances in neural information processing systems, 2024

    Siyi Chen, Huijie Zhang, Minzhe Guo, Yifu Lu, Peng Wang, and Qing Qu. Exploring low- dimensional subspace in diffusion models for controllable image editing.Advances in neural information processing systems, 2024

    2024

  22. [23]

    Lee, Jonathan Ho, Tim Salimans, David J

    Chitwan Saharia, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. InACM Special Interest Group for Computer Graphics and Interactive Techniques, 2022

    2022

  23. [24]

    ILVR: conditioning method for denoising diffusion probabilistic models

    Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: conditioning method for denoising diffusion probabilistic models. InIEEE/CVF International Conference on Computer Vision, 2021

    2021

  24. [25]

    Blended diffusion for text-driven editing of natural images

    Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    2022

  25. [26]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022

    2022

  26. [27]

    Compressible dynamics in deep overpa- rameterized low-rank learning & adaptation

    Can Yaras, Peng Wang, Laura Balzano, and Qing Qu. Compressible dynamics in deep overpa- rameterized low-rank learning & adaptation. InInternational Conference on Machine Learning, 2024. 15

    2024

  27. [28]

    Domain adaptive faster R-CNN for object detection in the wild

    Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster R-CNN for object detection in the wild. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018

    2018

  28. [29]

    Simultaneous deep transfer across domains and tasks

    Judy Hoffman, Eric Tzeng, Trevor Darrell, and Kate Saenko. Simultaneous deep transfer across domains and tasks. InDomain Adaptation in Computer Vision Applications. 2017

    2017

  29. [30]

    ifan: Image-instancefull alignment networks for adaptive object detection

    ChenfanZhuang,XintongHan,WeilinHuang,andMatthewR.Scott. ifan: Image-instancefull alignment networks for adaptive object detection. InAAAI Conference on Artificial Intelligence, 2020

    2020

  30. [31]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. InAdvances in Neural Information Processing Systems, 2019

    2019

  31. [32]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, 2020

    2020

  32. [33]

    Ganstrainedbyatwotime-scaleupdateruleconvergetoalocalnashequilibrium

    MartinHeusel,HubertRamsauer,ThomasUnterthiner,BernhardNessler,andSeppHochreiter. Ganstrainedbyatwotime-scaleupdateruleconvergetoalocalnashequilibrium. InAdvances in Neural Information Processing Systems, 2017

    2017

  33. [34]

    Wilson, Meriem Ghrib, and Thomas Oberlin

    Estelle Chigot, Dennis G. Wilson, Meriem Ghrib, and Thomas Oberlin. Style transfer with diffusion models for synthetic-to-real domain adaptation.CoRR, abs/2505.16360, 2025

    arXiv 2025

  34. [35]

    One-step effective diffusion network for real-world image super-resolution

    Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. InAdvances in Neural Information Processing Systems, 2024

    2024

  35. [36]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InInterna- tional Conference on Machine Learning, 2021

    2021

  36. [37]

    Synplay: Importing real-world diversity for a synthetic human dataset.arXiv preprint arXiv:2408.11814, 2024

    Jinsub Yim, Hyungtae Lee, Sungmin Eum, Yi-Ting Shen, Yan Zhang, Heesung Kwon, and Shuvra S Bhattacharyya. Synplay: Importing real-world diversity for a synthetic human dataset.arXiv preprint arXiv:2408.11814, 2024

    arXiv 2024

  37. [38]

    Okutama-action: An aerial view video dataset for concurrent human action detection

    Mohammadamin Barekatain, Miquel Martí, Hsueh-Fu Shih, Samuel Murray, Kotaro Nakayama, Yutaka Matsuo, and Helmut Prendinger. Okutama-action: An aerial view video dataset for concurrent human action detection. InIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2017

    2017

  38. [39]

    Semantic Drone Dataset – ICG, TU Graz.http://dronedataset.icg.tugraz.at, 2019

    Institute of Computer Graphics and Vision, TU Graz, Team Fraundorfer. Semantic Drone Dataset – ICG, TU Graz.http://dronedataset.icg.tugraz.at, 2019

    2019

  39. [40]

    Vision meets drones: A challenge.arXiv preprint arXiv:1804.07437, 2018

    Pengfei Zhu, Longyin Wen, Xiao Bian, Haibin Ling, and Qinghua Hu. Vision meets drones: A challenge.arXiv preprint arXiv:1804.07437, 2018

    Pith/arXiv arXiv 2018

  40. [41]

    Manipal-uav person detection dataset: A step towards benchmarking dataset and algorithms for small object detection.ISPRS Journal of Photogrammetry and Remote Sensing, 2023

    KR Akshatha, AK Karunakar, B Satish Shenoy, K Phani Pavan, V Dhareshwar Chinmay, et al. Manipal-uav person detection dataset: A step towards benchmarking dataset and algorithms for small object detection.ISPRS Journal of Photogrammetry and Remote Sensing, 2023. 16

    2023

  41. [42]

    BenleiCui,WenhaoSun,Xue-MeiDong,JingqunTang,andYiLiu.Attentiveeraser: Unleashing diffusion model’s object removal potential via self-attention redirection guidance, 2025

    2025

  42. [43]

    Data filtering networks, 2023

    AlexFang,AlbinMadappallyJose,AmitJain,LudwigSchmidt,AlexanderToshev,andVaishaal Shankar. Data filtering networks, 2023

    2023

  43. [44]

    Openclip, 2021

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 17

    2021