Pith. sign in

REVIEW 5 major objections 5 minor 37 references

FADE: Adversarial Concept Erasure in Flow Models

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read FADE erases concepts from trained diffusion models by making them fool a concept-detecting adversary, proving zero mutual information at equilibrium and reporting 0.0-0.5% residual concept detection in practice.

desk verdict Plausible method and a solid ablation, but the paper's headline claims rest on evidence it doesn't actually include—no FLUX results, missing FID/CLIP columns, and a theory that proves independence from the prompt, not absence of the concept. read the letter →

arxiv 2507.12283 v1 pith:SEGF3345 submitted 2025-07-16 cs.CV

Zixuan Fu , Yan Ren , Finn Carter , Chenyue Wang , Ze Niu , Dacheng Yu , Emily Davis , Bo Zhang This is my paper
classification cs.CV
keywords concepterasurediffusionmodelsadversarialtrainingmachineunlearningmutualinformationtext-to-imagegenerationmodelfidelitygenerativefairness
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

FADE (Fair Adversarial Diffusion Erasure) is a procedure for removing a specified concept — a person's face, an object class, an artistic style, or explicit content — from an already-trained text-to-image diffusion model by fine-tuning rather than retraining from scratch. The core idea is an adversarial game: a small discriminator is trained on the model's own generated images to detect whether a prompt contained the concept, and the diffusion model is fine-tuned so that images from concept prompts become indistinguishable from images from matched neutral prompts. The paper proves that when this game reaches equilibrium against a Bayes-optimal discriminator, the model's output distribution is independent of the concept and the mutual information $I(C;X)$ between the concept and the outputs is exactly zero, making the erasure a formal information-theoretic statement rather than a visual disguise. Empirically, FADE reports concept classification accuracy of 0.0% to 0.5% across object, celebrity, explicit-content, and artistic-style benchmarks on both Stable Diffusion and FLUX, while keeping image quality close to the original model; the harmonic mean of erasure and fidelity is 5-10% higher than the best prior baseline. If correct, this gives model deployers a post-hoc tool to honor privacy and copyright demands — removing a celebrity, a style, or harmful content on request — without rebuilding the model.

What carries the argument

The load-bearing object is the adversarial game between the diffusion model $M_\theta$ and the concept discriminator $D_\varphi(x)$, which outputs the probability that image $x$ contains the concept. The discriminator is trained with binary cross-entropy on pairs $(x_c, x_{\neg c})$ generated by the model itself from concept and neutral prompts, and the diffusion model minimizes the removal loss $L_{\mathrm{rem}}(\theta)=-\mathbb{E}[\log(1-D_\varphi(x_c))]$; since the model generates its own training data, no real images of the concept are ever needed. Theorem 5.1 supplies the guarantee: against the Bayes-optimal discriminator $D^*$, achieving $L_{\mathrm{rem}}=-\log 0.5$ forces $P_{\theta'}(X|C=1)=P_{\theta'}(X|C=0)$, hence $I(C;X)=0$. Two auxiliary mechanisms preserve everything else: the trajectory preservation loss $L_{\mathrm{pres}}$ compares the current model's noise predictions with the original model's on neutral prompts at early denoising timesteps, and the saliency-based weight restriction ranks weights by gradient magnitude on concept images and confines updates to the most concept-relevant parameters.

What would settle it

Run FADE with its early-stopping rule, then evaluate the edited model with detectors the training never saw — a large zero-shot CLIP classifier, a fine-tuned concept classifier, and human raters — on thousands of concept prompts spanning synonyms and paraphrases absent from the training prompt set. If any of these detectors flags the concept in more than a few percent of images generated from concept prompts while the internal discriminator hovers near 50% accuracy, then near-random internal-discriminator accuracy does not by itself establish that the concept is informationally absent. The paper's own quantitative claim could also be checked directly: estimate the total variation distance between the concept-present and concept-absent output distributions on held-out prompt pairs and compare it with the bound implied by the observed discriminator accuracy.

Watch

Extended reading notes

Core claim

The paper's central claim is that concept erasure can be cast as an adversarial min-max game whose equilibrium is exactly the state in which the erased concept carries no information about the model's outputs. Concretely, FADE couples the diffusion model $M_\theta$ with a concept discriminator $D_\varphi$ that learns to classify whether a generated image came from a concept prompt $y_c$ or a matched neutral prompt $y_{\neg c}$; the model is fine-tuned to fool $D_\varphi$, a trajectory-preservation loss keeps its early denoising steps aligned with the original model, and parameter updates are confined to concept-salient weights. Theorem 5.1 is the load-bearing result: for an ideal discriminator attaining the Bayes-optimal likelihood ratio, the generator's optimal loss value $-\log 0.5$ is reachable if and only if $P_{\theta'}(X|C=1)=P_{\theta'}(X|C=0)$, which is equivalent to $I(C;X)=0$; conversely, any positive residual mutual information leaves a discriminator that beats chance, so the objective targets information removal rather than mere visual suppression. The paper also derives a practical bound linking discriminator accuracy to the total variation distance between the concept-present and concept-absent output distributions, interpreting the equilibrium condition as demographic parity in a fairness sense, with the concept's presence in the prompt treated as a protected attribute.

Load-bearing premise

The assumption on which everything rests is that the small, randomly initialized discriminator trained on the model's own generated images is a faithful stand-in for any detector of the concept, so that driving that discriminator to near-random accuracy is taken as proof the concept is gone — and this is relied on even though training stops before the equilibrium where the formal guarantee is proven.

Editorial extensions

If this is right

  • Erased concepts become nearly undetectable: concept classification accuracy drops to 0.0-0.5% across CIFAR-10 objects, celebrities, NSFW terms, and artist styles, roughly halving the residual rate of the best prior baseline (ANT).
  • Removal no longer has to trade away image quality: the harmonic mean of erasure and fidelity reaches 85.9-91.4 across the four task groups, 5-10 points above ANT, with FID and CLIP similarity close to the unedited model.
  • The game scales to mass erasure: a multi-output discriminator erases all ten CIFAR-10 classes at once and extends to 100 concepts, with a reported harmonic mean of 82 vs. MACE's 75.
  • The method is backbone-agnostic: the same objective and preservation loss work on both Stable Diffusion and the 12-billion-parameter FLUX rectified-flow transformer.
  • Residual concept information is quantitatively bounded during training: the paper derives that a discriminator held to roughly 55% accuracy bounds the total variation distance between concept-present and concept-absent output distributions at about 0.1.

Reading between the lines

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

  • A test the paper leaves open: training stops well before the proven equilibrium, so in the deployed early-stopped regime the near-random internal discriminator is a proxy for erasure; probing the edited model with stronger, independently trained detectors on thousands of novel prompt phrasings would measure how good that proxy is.
  • If the mutual-information framing is correct, the same generate-and-detect loop should transfer to other generative domains — the conclusion itself suggests language models — where concept and neutral prompts can be paired and a detector trained on the model's own outputs.
  • The erasure's coverage is bounded by the prompt set: because the discriminator only ever sees the model's own generations for the supplied prompts and synonyms, concepts expressible only through phrasings outside that set could survive, which motivates an automatic prompt-proposing adversary as a natural extension.
  • Specificity is measured against a limited control set (50 prompts per concept), so the fidelity claim would be tested more demandingly on a broad, held-out prompt distribution rather than the fixed neutral set used in training.
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

5 major / 5 minor

Summary. The paper proposes FADE, an adversarial concept erasure method for text-to-image diffusion models. The method combines a discriminative adversary trained to distinguish images generated from concept-containing and neutral prompts with a trajectory-preservation loss and a saliency-based parameter restriction. The authors claim a formal guarantee (Theorem 5.1) that at equilibrium the mutual information between the concept indicator and the model output is zero, and they claim state-of-the-art empirical results on Stable Diffusion and FLUX for object, celebrity, explicit-content, and style erasure, improving the harmonic mean of erasure and fidelity by 5-10% over ANT. The empirical evidence, however, is incomplete: Table 1 omits the FID and CLIP columns it is described as containing, no FLUX results appear anywhere, and the harmonic-mean formula in Appendix B is ill-defined. The theoretical guarantee, moreover, equates independence from the prompt variable with absence of the concept in generated images, a step that relies on an unmeasured assumption about neutral-prompt baselines.

Significance. If the empirical results were fully reported and the theoretical claim were correctly scoped, FADE would be a useful contribution: the adversarial objective is a natural way to target concept-specific information, and the ablation study in Table 2 gives some evidence that each component (adversarial loss, preservation loss, saliency mask) contributes. The paper does not ship code, does not report the promised FID/CLIP or FLUX numbers, and its central 'formal guarantee' is essentially the definition of GAN equilibrium plus an unvalidated baseline assumption. As it stands, the claimed state-of-the-art performance and the claimed theoretical novelty are not supported, so the significance is limited to a plausible algorithmic idea awaiting proper evaluation.

major comments (5)
  1. [§6.1, Table 1; Appendix B] The central empirical claim is unverifiable from the manuscript. Table 1 reports only Acc and H columns, yet its caption states that FID and CLIP similarity measure overall output quality, and §6.2 says 'CLIP similarity (not fully shown in table for brevity)'; the promised 'Full table with FID, CLIP metrics' does not appear in Appendix B. The harmonic mean H is defined in Appendix B with F = 1/2 ((CLIP sim/CLIP*orig) + (max(FID*orig - (FID - FID*orig), 0)/FID*orig)); this expression can exceed 1 when FID improves, and no normalization or baseline values are given. Since the paper's headline '5-10% improvement' is a statement about H, these omissions undermine the main empirical conclusion.
  2. [Abstract; §6.1] The abstract and contributions claim evaluation on FLUX and the title promises flow models, but no FLUX result is reported anywhere. Table 1 is labeled 'on Stable Diffusion', and Appendix B's additional experiments contain only SD multi-concept results and runtime. The claimed generality to FLUX is therefore unsupported.
  3. [§5, Theorem 5.1; §3] The theoretical guarantee does not establish concept erasure as defined. Section 3 defines erasure as requiring that images generated from any prompt entailing c do not contain c. Theorem 5.1 establishes at most that P_{θ'}(X|C=1) = P_{θ'}(X|C=0), i.e., independence of the output from the binary prompt variable; this implies equality of concept-appearance probabilities, not that they are near zero. The paper's bridge claim, that P('c' in output | C=0) is 'essentially 0' for well-chosen neutral prompts (Section 5, Bound on Concept Appearance Probability), is never measured. For concepts with non-negligible baseline rates (e.g., a celebrity from 'a photo of a person'), the equilibrium would match that rate, so the reported near-zero concept accuracy could reflect a low baseline rather than successful erasure.
  4. [§5; §4.3] Theorem 5.1 is not a derived guarantee about FADE; it restates the standard Bayes-optimal discriminator solution and the GAN equilibrium condition. The proof sketch assumes an ideal discriminator D*, but the method trains a small, randomly initialized CNN (§4.3) and stops early once D's accuracy reaches approximately 50% (§4.3 and §6.3). No analysis connects the trained discriminator's accuracy to a bound on residual concept information in the early-stopped model, so the claimed 'formal guarantee' does not apply to the model that is actually deployed.
  5. [§6.2; Appendix B] The multi-concept results are inconsistent and unreported. Section 6.2 reports H = 82 when erasing 100 concepts and compares to 'MACE reported H around 75', while Appendix B's 'Multi-concept results' describes erasing all 10 CIFAR classes simultaneously with a 10-way classifier and reports H = 82.3 for that setting. Neither claim is backed by a table or a detailed protocol, and the relationship between the two numbers is not explained.
minor comments (5)
  1. [Throughout] The manuscript contains numerous typos, e.g., 'Erasre' in the introduction, 'methds' in §1, 'focusig' in §6.3, and 'implmentation' in §6.1.
  2. [§4.2, Eq. (2)] The trajectory-preservation loss Lpres sums over t = 1 to T0, but the text says T0 is a cutoff for early denoising steps; since diffusion timesteps are conventionally indexed from T down to 1, the direction of the sum and the definition of 'early' should be clarified.
  3. [§6.1] For object erasure, the paper states that a ResNet trained on CIFAR-10 is used to classify generated images, but generated images are not 32x32 CIFAR-style images; the transfer of the classifier to high-resolution outputs should be validated or described.
  4. [Appendix B] The formula for F is not invariant to the scale of FID and CLIP and can produce values above 1; if this is intentional, a worked example for the reported H values should be provided.
  5. [References] Some references are incomplete or missing, e.g., 'FLUX.1 (?)' in §6.1, and several cited works appear tangential to the paper's topic; the reference list should be trimmed to works actually used.

Circularity Check

1 steps flagged · score 4.0 of 10

Theoretical 'guarantee' restates GAN equilibrium; empirical SOTA comparison is external and not circular.

  1. self definitional [Section 5, Theorem 5.1 and proof sketch (page 6); objective defined in Section 4.1.]
    "At equilibrium (Nash equilibrium of the minimax game), the only solution is where D is indifferent (outputs 0.5) and Mθ′ cannot reduce loss further, which precisely implies Pθ′(X|C = 1) = Pθ′(X|C = 0). This result formalizes that our adversarial training objective indeed pushes the model towards eliminating concept-specific evidence in X."

    The adversarial objective is defined as a minimax game in which D learns to distinguish images from concept prompts (C=1) from neutral prompts (C=0) while M learns to fool D. The Nash equilibrium of that game is, by the standard optimal-discriminator argument reproduced in the proof, exactly P(X|C=1)=P(X|C=0), which for binary C with prior 0.5 is equivalent to I(C;X)=0. The theorem therefore restates the equilibrium condition of the paper's own objective rather than deriving a property of concept erasure from independent assumptions. The claimed 'formal guarantee' is true by construction of the game; the additional step from I(C;X)=0 to 'images do not contain c' is not established by the theorem.

full rationale

FADE's empirical headline — near-zero concept accuracy on CIFAR-10 objects, celebrities, NSFW, and styles with improved harmonic mean over ESD/UCE/MACE/ANT — is benchmarked against external baselines using pretrained classifiers, so that claim is not circular. The circularity is localized to the theoretical 'guarantee': Theorem 5.1 is a restatement of the GAN equilibrium for the objective defined in Sec. 4.1, not an independent proof of concept absence. A separate, non-circular weakness is the unmeasured bridge assumption in Sec. 5 ('Pθ′['c'|C=0] is essentially 0') that converts conditional independence into the paper's defined erasure criterion; this missing support should be weighed as a correctness/rigor issue. Overall score 4 reflects one definitionally-loaded theoretical result with otherwise independent external evaluation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method and its theoretical guarantee rest on standard GAN equilibrium properties and on an unproven information-theoretic bound. The early stopping in practice means the equilibrium guarantee is never actually reached.

free parameters (3)
  • λ (preservation weight) = not specified (chosen by hand)
    Controls the trade-off between concept removal and trajectory preservation; the paper says 'we typically choose λ such that the preservation loss prevents noticeable quality degradation' (Section 4.2).
  • T0 (early-step cutoff) = not specified (e.g., mid-way)
    Determines which denoising steps are preserved; described inconsistently in Section 4.2 and Algorithm 1.
  • Saliency threshold / Θ_salient size = not specified
    The number of weights updated is determined by ranking gradient magnitudes, but no threshold or fraction is given (Section 4.2).
assumptions (4)
  • standard math At the Nash equilibrium of the adversarial game with an optimal discriminator, the generator's conditional distributions for concept and neutral prompts coincide (P(X|C=1)=P(X|C=0)).
    This is the standard GAN minimax property, used in Theorem 5.1.
  • domain assumption The trained discriminator Dϕ is sufficiently expressive and well-converged to approximate the Bayes-optimal D*.
    The paper states 'by training a sufficiently expressive Dϕ, we approximate this condition' (Section 5).
  • ad hoc to paper The stated bound I(C;X) ≤ KL(P(X|C=1)||P(X|C=0)) holds.
    This inequality is asserted without proof in Section 5 and is not a generally valid information-theoretic bound.
  • domain assumption Neutral prompts are chosen so that P(c in output|C=0)≈0.
    Used to argue that erasure implies concept appearance probability near zero (Section 5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of FADE: Adversarial Concept Erasure in Flow Models." pith.science (2026). https://pith.science/paper/SEGF3345

@misc{pith2026250712283,
  author       = {Pith},
  title        = {Pith review of: FADE: Adversarial Concept Erasure in Flow Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SEGF3345}},
  note         = {Machine review of arXiv:2507.12283}
}
read the original abstract

Diffusion models have demonstrated remarkable image generation capabilities, but also pose risks in privacy and fairness by memorizing sensitive concepts or perpetuating biases. We propose a novel \textbf{concept erasure} method for text-to-image diffusion models, designed to remove specified concepts (e.g., a private individual or a harmful stereotype) from the model's generative repertoire. Our method, termed \textbf{FADE} (Fair Adversarial Diffusion Erasure), combines a trajectory-aware fine-tuning strategy with an adversarial objective to ensure the concept is reliably removed while preserving overall model fidelity. Theoretically, we prove a formal guarantee that our approach minimizes the mutual information between the erased concept and the model's outputs, ensuring privacy and fairness. Empirically, we evaluate FADE on Stable Diffusion and FLUX, using benchmarks from prior work (e.g., object, celebrity, explicit content, and style erasure tasks from MACE). FADE achieves state-of-the-art concept removal performance, surpassing recent baselines like ESD, UCE, MACE, and ANT in terms of removal efficacy and image quality. Notably, FADE improves the harmonic mean of concept removal and fidelity by 5--10\% over the best prior method. We also conduct an ablation study to validate each component of FADE, confirming that our adversarial and trajectory-preserving objectives each contribute to its superior performance. Our work sets a new standard for safe and fair generative modeling by unlearning specified concepts without retraining from scratch.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 8 canonical work pages

  1. [1]

    Visual instruction tuning with 500x fewer pa- rameters through modality linear representation-steering

    Bi, J., Wang, Y ., Chen, H., Xiao, X., Hecker, A., Tresp, V ., and Ma, Y . Visual instruction tuning with 500x fewer pa- rameters through modality linear representation-steering. arXiv preprint arXiv:2412.12359,

    arXiv

  2. [4]

    Offset: Segmentation-based focus shift revision for composed image retrieval, 2025c

    Chen, Z., Hu, Y ., Li, Z., Fu, Z., Song, X., and Nie, L. Offset: Segmentation-based focus shift revision for composed image retrieval, 2025c. URL https://arxiv.org/ abs/2507.05631. Cui, Y ., Fang, Z., Shan, J., Gu, Z., and Zhou, S. 3d ob- ject tracking with transformer. British Machine Vision Conference,

    arXiv

  3. [5]

    Eraseanything: Enabling concept erasure in rectified flow transformers

    Gao, D., Lu, S., Walters, S., Zhou, W., Chu, J., Zhang, J., Zhang, B., Jia, M., Zhao, J., Fan, Z., et al. Eraseanything: Enabling concept erasure in rectified flow transformers. arXiv preprint arXiv:2412.20413,

    arXiv

  4. [8]

    Prompt-to-prompt im- age editing with cross-attention control

    Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y ., and Cohen-Or, D. Prompt-to-prompt im- age editing with cross-attention control. arXiv preprint arXiv:2208.01626,

    arXiv

  5. [9]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance. In NeurIPS 2022 Workshop on Score-Based Methods,

    work page 2022

  6. [11]

    Mvctrack: Boost- ing 3d point cloud tracking via multimodal-guided virtual cues

    Hu, Z., Zhou, S., Zhao, S., and Yuan, Z. Mvctrack: Boost- ing 3d point cloud tracking via multimodal-guided virtual cues. International Conference on Robotics and Automa- tion 2025,

    work page 2025

  7. [12]

    Scaletrack: Scaling and back-tracking automated gui agents

    Huang, J., Zeng, Z., Han, W., Zhong, Y ., Zheng, L., Fu, S., Chen, J., and Ma, L. Scaletrack: Scaling and back-tracking automated gui agents. arXiv preprint arXiv:2505.00416, 2025a. Huang, Q., Chen, Z., Li, Z., Wang, C., Song, X., Hu, Y ., and Nie, L. Median: Adaptive intermediate-grained aggrega- tion network for composed image retrieval. In Proceed- ing...

    arXiv 2025

  8. [13]

    Overview of the nlpcc 2023 shared task: Chinese medical instructional video question answering

    Li, B., Weng, Y ., Guo, H., Sun, B., Li, S., Luo, Y ., Qi, M., Liu, X., Han, Y ., Liang, H., et al. Overview of the nlpcc 2023 shared task: Chinese medical instructional video question answering. In CCF International Conference on Natural Language Processing and Chinese Computing, pp. 233–242. Springer,

    work page 2023

Show all 37 references
  1. [14]

    Distinct but correct: generating diversi- fied and entity-revised medical response

    Li, B., Sun, B., Li, S., Chen, E., Liu, H., Weng, Y ., Bai, Y ., and Hu, M. Distinct but correct: generating diversi- fied and entity-revised medical response. Science China Information Sciences, 67(3):132106, 2024a. Li, B., Weng, Y ., Xia, F., and Deng, H. Towards bet- ter ch...

    arXiv

  2. [15]

    Phy124: Fast physics-driven 4d content generation from a single image

    Lin, J., Wang, Z., Hou, Y ., Tang, Y ., and Jiang, M. Phy124: Fast physics-driven 4d content generation from a single image. arXiv preprint arXiv:2409.07179, 2024a. Lin, J., Wang, Z., Jiang, S., Hou, Y ., and Jiang, M. Phys4dgen: A physics-driven framework for controllable and...

    arXiv

  3. [16]

    Lu, S., Wang, Z., Li, L., Liu, Y ., and Kong, A. W.-K. Mace: Mass concept erasure in diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6430–6440, 2024a. Lu, S., Zhou, Z., Lu, J., Zhu, Y ., and Kong, A. W.-K. Ro- bust wat...

    arXiv

  4. [17]

    Magicstick: Controllable video edit- ing via control handle transformations

    Ma, Y ., Cun, X., He, Y ., Qi, C., Wang, X., Shan, Y ., Li, X., and Chen, Q. Magicstick: Controllable video edit- ing via control handle transformations. arXiv preprint arXiv:2312.03047,

    arXiv

  5. [18]

    Follow your pose: Pose-guided text-to-video generation using pose-free videos

    Ma, Y ., He, Y ., Cun, X., Wang, X., Chen, S., Li, X., and Chen, Q. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp. 4117–4125, 2024a. Ma, Y ., Liu, H., Wang, H., Pan...

    2024 arXiv

  6. [19]

    Langtime: A language-guided unified model for time se- ries forecasting with proximal policy optimization

    Niu, W., Xie, Z., Sun, Y ., He, W., Xu, M., and Hao, C. Langtime: A language-guided unified model for time se- ries forecasting with proximal policy optimization. arXiv preprint arXiv:2503.08271,

    arXiv

  7. [20]

    Lightweight adap- tive feature de-drifting for compressed image classifica- tion

    Peng, L., Cao, Y ., Sun, Y ., and Wang, Y . Lightweight adap- tive feature de-drifting for compressed image classifica- tion. IEEE Transactions on Multimedia, 26:6424–6436, 2024a. Peng, L., Li, W., Pei, R., Ren, J., Xu, J., Wang, Y ., Cao, Y ., and Zha, Z.-J. Towards realistic...

    arXiv

  8. [21]

    A com- prehensive survey of deep learning for multivariate time series forecasting: A channel strategy perspective

    Qiu, X., Cheng, H., Wu, X., Hu, J., and Guo, C. A com- prehensive survey of deep learning for multivariate time series forecasting: A channel strategy perspective. arXiv preprint arXiv:2502.10721, 2025a. Qiu, X., Li, X., Pang, R., Pan, Z., Wu, X., Yang, L., Hu, J., Shu, Y ., L...

  9. [22]

    Backdoor cleaning without ex- ternal guidance in mllm fine-tuning

    Rong, X., Huang, W., Liang, J., Bi, J., Xiao, X., Li, Y ., Du, B., and Ye, M. Backdoor cleaning without ex- ternal guidance in mllm fine-tuning. arXiv preprint arXiv:2505.16916,

    arXiv

  10. [23]

    Ptt: Point-track- transformer module for 3d single object tracking in point clouds

    Shan, J., Zhou, S., Fang, Z., and Cui, Y . Ptt: Point-track- transformer module for 3d single object tracking in point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1310–1316. IEEE,

    2021

  11. [24]

    URL https://arxiv.org/ pdf/2503.19474

    IEEE. URL https://arxiv.org/ pdf/2503.19474. To appear. Shi, Y ., Zhou, S., Wang, W., and Lu, X. Rethinking it- erative stereo matching from a diffusion bridge model perspective. Pattern Recognition, pp. 111737,

    arXiv

  12. [25]

    Text- toon: Real-time text toonify head avatar from single video

    Song, L., Chen, L., Liu, C., Liu, P., and Xu, C. Text- toon: Real-time text toonify head avatar from single video. 15 FADE: Adversarial Concept Erasure in Flow Models In SIGGRAPH Asia 2024 Conference Papers, pp. 1–11, 2024a. Song, L., Liu, P., Chen, L., Yin, G., and Xu, C. Tri...

    2024

  13. [26]

    Divide-and- conquer: Confluent triple-flow network for rgb-t salient object detection

    Tang, H., Li, Z., Zhang, D., He, S., and Tang, J. Divide-and- conquer: Confluent triple-flow network for rgb-t salient object detection. IEEE Trans. Pattern Anal. Mach. Intell., 47(3):1958–1974, 2025a. Tang, Y ., Bi, J., Xu, S., Song, L., Liang, S., Wang, T., Zhang, D., An, J....

    1958

  14. [27]

    URL https: //arxiv.org/abs/2312.17432. Tang, Y ., Bi, J., Huang, C., Liang, S., Shimada, D., Hua, H., Xiao, Y ., Song, Y ., Liu, P., Feng, M., Guo, J., Liu, Z., Song, L., V osoughi, A., He, J., He, L., Zhang, Z., Luo, J., and Xu, C. Caption anything in video: Fine- grained obj...

  15. [28]

    Ascd: Attention- steerable contrastive decoding for reducing hallucination in mllm

    Wang, Y ., Bi, J., Ma, Y ., and Pirk, S. Ascd: Attention- steerable contrastive decoding for reducing hallucination in mllm. arXiv preprint arXiv:2506.14766,

    arXiv

  16. [29]

    Dynamic uncertainty learning with noisy correspondence for text-based person search

    Xie, Z., Ji, H., and Meng, L. Dynamic uncertainty learning with noisy correspondence for text-based person search. arXiv preprint arXiv:2505.06566,

  17. [30]

    Eedit: Rethinking the spatial and temporal redundancy for efficient image editing

    16 FADE: Adversarial Concept Erasure in Flow Models Yan, Z., Ma, Y ., Zou, C., Chen, W., Chen, Q., and Zhang, L. Eedit: Rethinking the spatial and temporal redundancy for efficient image editing. arXiv preprint arXiv:2503.10270,

  18. [31]

    Ich-scnet: Intracerebral hemor- rhage segmentation and prognosis classification network using clip-guided sam mechanism

    Yu, X., Elazab, A., Ge, R., Jin, H., Jiang, X., Jia, G., Wu, Q., Shi, Q., and Wang, C. Ich-scnet: Intracerebral hemor- rhage segmentation and prognosis classification network using clip-guided sam mechanism. In 2024 IEEE Inter- national Conference on Bioinformatics and Biomedi...

    2024 arXiv

  19. [32]

    Spot! revisiting video-language models for event understanding

    Zhang, G., Bi, J., Gu, J., Chen, Y ., and Tresp, V . Spot! revisiting video-language models for event understanding. arXiv preprint arXiv:2311.12919,

    arXiv

  20. [33]

    Kinmo: Kinematic-aware human motion understanding and generation, 2025a

    Zhang, P., Liu, P., Kim, H., Garrido, P., and Chaudhuri, B. Kinmo: Kinematic-aware human motion understanding and generation, 2025a. URL https://arxiv.org/ abs/2411.15472. Zhang, W., Lin, X., Yu, W., Yao, G., Zhong, J., Li, Y ., Han, R., Xu, S., Shi, H., and Luo, C. Dconad: A ...

    arXiv

  21. [34]

    Y ., and Ma, L

    Zhou, S., Tian, Z., Chu, X., Zhang, X., Zhang, B., Lu, X., Feng, C., Jie, Z., Chiang, P. Y ., and Ma, L. Fastpillars: A deployment-friendly pillar-based 3d detector. arXiv preprint arXiv:2302.02367,

    arXiv

  22. [35]

    Instantswap: Fast customized concept swapping across sharp shape differences

    17 FADE: Adversarial Concept Erasure in Flow Models Zhu, C., Li, K., Ma, Y ., Tang, L., Fang, C., Chen, C., Chen, Q., and Li, X. Instantswap: Fast customized concept swapping across sharp shape differences. arXiv preprint arXiv:2412.01197,

    arXiv

  23. [36]

    18 FADE: Adversarial Concept Erasure in Flow Models A. Additional Backgrounds With the advancement of deep learning (Zheng et al., 2024; 2023; 2022; 2025; Yu et al., 2025b;a; 2024a;b; Qiu et al., 2024; 2025d;c; Liu et al., 2025d; Qiu et al., 2025a;b; Wu et al., 2025b; 2024c; L...

    2024

  24. [37]

    Multimodality (Shen et al.,

    and generative models (He et al., 2025c;b; 2023a; 2025d;e;a; 2024a;b; 2023c;b; Xiao et al., 2024; Wu et al., 2025a; Ma et al., 2024b; 2025b;a;c; 2023; 2024a; 2022; Yan et al., 2025; Zhang et al., 2025e; Zhu et al., 2024; Wang et al., 2024; Feng et al., 2025; Chen et al., 2024;...

    2024

  25. [2020]

    MultiRC: Joint learning for time series anomaly prediction and detection with multi-scale reconstructive contrast

    Hu, S., Zhao, K., Qiu, X., Shu, Y ., Hu, J., Yang, B., and Guo, C. MultiRC: Joint learning for time series anomaly prediction and detection with multi-scale reconstructive contrast. arXiv preprint arXiv:2410.15997,

    arXiv

  26. [2022]

    Ad- versarial learning for neural pde solvers with sparse data

    Gong, Y ., Hou, Y ., Wang, Z., Lin, Z., and Jiang, M. Ad- versarial learning for neural pde solvers with sparse data. arXiv preprint arXiv:2409.02431, 2024a. Gong, Y ., Zeng, Q., Xu, D., Wang, Z., and Jiang, M. Cross- modality attack boosted by gradient-evolutionary mul- tifor...

    2023 arXiv

  27. [2023]

    Fedbip: Het- erogeneous one-shot federated learning with personalized latent diffusion models

    Chen, H., Li, H., Zhang, Y ., Bi, J., Zhang, G., Zhang, Y ., Torr, P., Gu, J., Krompass, D., and Tresp, V . Fedbip: Het- erogeneous one-shot federated learning with personalized latent diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference (...

  28. [2024]

    Why reasoning mat- ters? a survey of advancements in multimodal reasoning (v1)

    Bi, J., Liang, S., Zhou, X., Liu, P., Guo, J., Tang, Y ., Song, L., Huang, C., Sun, G., He, J., et al. Why reasoning mat- ters? a survey of advancements in multimodal reasoning (v1). arXiv preprint arXiv:2504.03151, 2025a. Bi, J., Liu, P., V osoughi, A., Wu, J., He, J., and Xu...

  29. [2025]

    Eliminate devia- tion with deviation for data augmentation and a gen- eral multi-modal data learning method

    Gong, Y ., Huang, L., and Chen, L. Eliminate devia- tion with deviation for data augmentation and a gen- eral multi-modal data learning method. arXiv preprint arXiv:2101.08533,

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.