Pith. sign in

REVIEW 5 major objections 5 minor 36 references

TRACE: Trajectory-Constrained Concept Erasure in Diffusion Models

T0 review · 5 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 TRACE erases targeted concepts from text-to-image diffusion models by aligning concept tokens with a neutral token in cross-attention and editing only late denoising steps, reporting the best erasure-fidelity trade-off on object…

desk verdict TRACE is a plausible engineering combination of UCE, ANT, and MACE, but its theoretical guarantee is not satisfied by its own closed-form update, and the experimental support is a single unreproducible table. read the letter →

arxiv 2505.23312 v1 pith:KJWA75OE submitted 2025-05-29 cs.CV

Finn Carter This is my paper
classification cs.CV
keywords concepterasurediffusionmodelscross-attentioneditingtrajectory-awarefine-tuninglow-rankadaptationtext-to-imagesafetyStablerectifiedflow
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

TRACE is a method for removing targeted concepts—objects, celebrity identities, artistic styles, and explicit content—from text-to-image diffusion models. The paper argues that a concept is completely erased when, in every cross-attention layer, head, and denoising step, the concept token's key and value projections coincide with those of a neutral token; under that condition, prompting the concept is distributionally identical to prompting the neutral token. TRACE implements this with a closed-form rank-one update to attention weights followed by a trajectory-aware fine-tuning of per-concept LoRA adapters that only changes late denoising behavior. On object, face, style, and NSFW benchmarks, the authors report the best erasure-fidelity trade-off among ESD, UCE, MACE, ANT, and EraseAnything, and show the method transfers to the rectified-flow model FLUX.

What carries the argument

The load-bearing object is the cross-attention nullification condition $W_K e_w = W_K e_u$ and $W_V e_w = W_V e_u$ holding at every layer, head, and timestep. TRACE approximates it by a rank-one update $\Delta W = (W e_u - W e_w) e_w^\top / \|e_w\|^2$, damped by a validation-set factor, using the end-of-sequence token embedding as the neutral $e_u$. It then fine-tunes one low-rank adapter per concept with a trajectory-aware loss: early denoising steps are kept close to the original model, while from the spontaneous semantic breaking point onward the model is pushed away from concept-specific details. A separate integration loss on prompts containing multiple erased concepts prevents the adapters from interfering.

What would settle it

Run a model whose cross-attention layers are edited exactly by the closed-form update, leaving all residual connections and normalization layers intact, and compare the distribution of images from a concept prompt against the distribution from the neutral-token prompt using a sensitive divergence such as feature-space MMD or an object detector's confidence. If the divergence clearly exceeds the sampling noise between two runs of the original model, the induction in Theorem 3.1 does not hold in the real architecture. A second check: measure cosine similarity between the end-of-sequence embedding and each erased concept's embedding; high similarity would mean the 'neutral' substitute is not neutral after all.

Watch

Extended reading notes

Core claim

The paper's central claim is that a diffusion model can be made to forget a concept exactly: if, for every cross-attention layer, head, and denoising step, the key and value projections of the concept token $e_w$ are made equal to those of a neutral token $e_u$, then the edited model's output distribution given a prompt containing $w$ is identical to the original model's output distribution given the same prompt with $w$ replaced by $u$ (Theorem 3.1). The paper derives a closed-form rank-one projection for this nullification (Proposition 3.2), a Lipschitz stability bound relating parameter-edit size to distributional shift on unrelated prompts (Theorem 3.3), a locality result showing late-step edits preserve early image statistics (Corollary 3.4), and a trajectory-constrained objective that minimizes concept-specific divergence in late steps while protecting clean prompts (Theorem 3.5). Empirically, TRACE reports the best harmonic means on four benchmarks—object erasure $H_o=85.6$, celebrity erasure $H_c=92.3$, 95% style removal, and 2% residual nudity detection—with FIDs lower than all compared baselines.

Load-bearing premise

The guarantee of perfect erasure rests on the assumption that matching the concept token's attention projections to a neutral token in every layer and timestep fully determines the generated image, so no residual connection, normalization layer, or other pathway can still carry the concept; it also assumes the end-of-sequence token is genuinely neutral for every erased concept.

Editorial extensions

If this is right

  • If the projection condition holds exactly, no prompt containing the erased token can reproduce the concept: outputs are indistinguishable from those of the neutral-token prompt, so synonym and paraphrase attacks that route through the same embeddings fail.
  • Focusing edits on late denoising steps leaves early layout and global geometry untouched, which is why FID and CLIP scores stay close to the original model on unrelated prompts.
  • Using separate LoRA modules per concept with an integration loss allows dozens of concepts to be erased jointly without the interference that plagues sequential or monolithic fine-tuning.
  • Because the closed-form update depends only on the text encoder and cross-attention structure, the method applies to rectified-flow transformer models such as FLUX with only minor adjustment, not just to U-Net latent diffusion.
  • At the reported numbers, erasure is near-total on the tested benchmarks: near zero classifier accuracy for erased objects and celebrities, 95% style removal, and 2% residual nudity detection.

Reading between the lines

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

  • A testable extension the paper leaves implicit: replace the EOS neutral token with a learned or per-concept neutral vector; for abstract concepts such as artistic style, EOS may not be semantically empty, and a learned neutral would likely improve the closed-form initialization.
  • The theoretical guarantee assumes attention projections are the only channel for concept information; in actual U-Net and transformer stacks, residual connections and normalization layers provide a second pathway, so the exact-equality claim should be read as approximate in practice.
  • If the trajectory-constrained late-stage editing is correct, the same 'edit late, preserve early' principle should transfer to other safety interventions such as watermarking and prompt filtering, since early structure carries layout rather than concept identity.
  • A natural stress test for the multi-concept integration loss is scaling to all pairs of 50 concepts; sampling pairs per batch may miss adversarial conjunctions, so exhaustive or coverage-maximizing pair selection is a concrete robustness extension.
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 introduces TRACE, a concept erasure method for text-to-image diffusion models. The method combines a rank-one closed-form update to cross-attention key/value projections with a trajectory-aware LoRA fine-tuning objective that is intended to constrain edits to late denoising steps. The authors claim a rigorous theoretical framework (Theorems 3.1–3.5) that proves perfect concept suppression and bounded collateral damage, and they report state-of-the-art results on object, celebrity, style, and NSFW erasure benchmarks, including application to both Stable Diffusion and the Flux rectified-flow transformer.

Significance. If the theoretical and empirical claims were established, TRACE would be a useful contribution: the modular per-concept LoRA design, the late-stage trajectory constraint, and the cross-architecture demonstration are all potentially valuable. The ablation study is informative and the benchmark coverage is broad. However, the central theoretical guarantee is not proven, the closed-form initialization does not implement the stated sufficient condition, and the empirical evidence is a single table with no seeds, error bars, statistical tests, or code. As presented, the paper does not support its headline claims of provable erasure and state-of-the-art performance, so its significance is currently limited.

major comments (5)
  1. [Section 4.1, Eq. (1)] The closed-form update does not satisfy the sufficient condition of Theorem 3.1. For the stated update ΔW = (W e_u − W e_w) e_w^T / ||e_w||^2, one obtains (W+ΔW)e_w = W e_u, but (W+ΔW)e_u = W e_u + (W e_u − W e_w)(e_w·e_u)/||e_w||^2, so (W+ΔW)(e_w − e_u) = −(W e_u − W e_w)(e_w·e_u)/||e_w||^2, which is generically nonzero. Thus the theorem's condition (W+ΔW)e_w = (W+ΔW)e_u is not enforced. Moreover, the rank-one minimizer of Proposition 3.2 is ΔW = −W(e_w−e_u)(e_w−e_u)^T / ||e_w−e_u||^2, which is different from Eq. (1). Since this initialization is the claimed bridge between theory and algorithm, the theoretical framework does not cover the implemented method.
  2. [Section 3.2, Theorem 3.1] The proof of Theorem 3.1 is not actually given: the text says 'Identical to the argument given previously,' but no previous argument appears in this manuscript. The asserted induction over attention layers and diffusion timesteps is nontrivial because the outputs of cross-attention layers pass through residual connections, layer normalization, and feedforward blocks; equality of per-head key/value projections for a single token does not by itself imply equality of the entire denoising trajectory. A complete proof, or an explicit statement of the missing assumptions, is needed before this can be called a formal guarantee.
  3. [Section 3.2, Theorems 3.3–3.5] Theorem 3.3 states the bound D_KL(p_θ(x0|q) ∥ p_θ'(x0|q)) ≤ L ||Δθ||^2, but no proof is supplied, and Lipschitz continuity of the score network alone does not by itself imply a distributional KL bound of that form; the statement lacks the necessary regularity, dimension, and discretization conditions. Corollary 3.4 is true by definition: assuming Δθ affects only timesteps t < t_late means that the first T−t_late steps are unchanged, and the corollary does not show that the trajectory-constrained loss actually has this property. Theorem 3.5 observes that gradient descent descends the objective; the 'controlled ascent of size O(η)' is a first-order Taylor statement, not a guarantee, and the theorem does not establish that the preservation term remains controlled at convergence. The claimed rigorous framework in Section 3 is therefore not established.
  4. [Table 1 and Section 5.2] The central empirical claim rests entirely on Table 1, which reports no error bars, no number of seeds, no statistical significance tests, and no confidence intervals. Metrics such as Ho, Hc, FaceRate, and StyleRem are defined only verbally in Section 5.1, and precise formulas are missing. The table does not include EraseAnything or the Flux results that are claimed in the text, and no code or repository is provided despite the promise to release it. Consequently, the state-of-the-art claim and the cross-architecture claim are not supported by the evidence presented.
  5. [Sections 3.1 and 4.1] The choice of the EOS token embedding as the neutral token u is an unvalidated assumption. Theorem 3.1 requires that replacing w with u in a prompt leaves the output distribution identical to that of a prompt that does not invoke the concept, but the EOS embedding is not demonstrated to be semantically neutral or even consistent across prompts, contexts, and timesteps. The theoretical guarantee applies only to the extent this assumption holds, and no experiment or analysis verifies it.
minor comments (5)
  1. [Throughout] Several cross-references are unresolved, including 'Theorem ??', 'Eq. (??)', and 'Figure ??', particularly in Sections 4.1 and 5.2; these placeholders should be filled for the paper to be readable.
  2. [Table 1 and Section 5.2] The introduction and Section 5.2 say the evaluation compares against six baseline methods including EraseAnything, but Table 1 lists only ESD, UCE, MACE, ANT, and TRACE; the missing baseline rows or the text should be reconciled.
  3. [Section 5.1] The metric definitions are incomplete: Ho, Hc, FaceRate, and StyleRem are described in prose but no algebraic definitions are given, making the reported numbers hard to reproduce or verify.
  4. [Section 4.1, after Eq. (1)] The sentence immediately following Eq. (1) is dimensionally garbled: 'ensures W_K(e_w + e_w/||e_w||^2 (e_w)^T) = W_K e_u' is not a well-formed statement; the intended identity is presumably (W+ΔW)e_w = W e_u.
  5. [References] The reference list contains typographical artifacts such as 'Dal ˘00ed' and 'W ACV'24', and the affiliation line appears incomplete; these should be cleaned before final submission.

Circularity Check

4 steps flagged · score 6.0 of 10

Core 'theorems' are restatements or tautologies, and the NSFW metric is literally the training objective.

  1. renaming known result [Section 3.2, Theorem 3.1 and its proof]
    "We restate the original sufficient condition as a first theorem and elevate several further claims that were previously embedded in the prose to formal statements. ... Proof. Identical to the argument given previously; we reproduce only the key idea."

    The paper's first theorem is explicitly a restatement of a previously known sufficient condition, with its proof delegated to 'the argument given previously.' The theoretical framework therefore imports the central erasure criterion rather than deriving it, and the claimed 'provable suppression' rests on an external earlier argument that is not reproduced or verified here. This is a renamed prior result presented as a formal contribution.

  2. self definitional [Section 3.2, Corollary 3.4]
    "Suppose ∆θ affects only the score predictions for timesteps t < tlate. Then global image statistics (captured by any functional that depends only on the first k denoising steps with k = T − tlate) remain unchanged"

    The corollary's conclusion is a restatement of its assumption: any statistic that depends only on the steps unaffected by ∆θ is unchanged by definition. No diffusion-specific mechanism is used, and the 'locality of late-step edits' is imposed by construction by the trajectory-constrained loss rather than derived. The paper cites this tautology as formal support for editing only late denoising steps.

2 more flagged steps
  1. self definitional [Section 3.2, Theorem 3.5]
    "Any parameter update obtained by (stochastic) gradient descent on L satisfies, to first order, ∆θ = −η ∇θ′L(θ′), η > 0, and therefore achieves simultaneous descent of the concept-specific term and controlled ascent (of size O(η)) of the preservation term."

    The 'theorem' states an analytic property of gradient descent on the very loss L it defines: minimizing L moves parameters along its negative gradient. The 'controlled ascent of size O(η)' is just the statement that a small parameter step causes a proportionally small change through the earlier Lipschitz bound. The claimed erasure guarantee is therefore equivalent to the definition of the optimization procedure, not an independent result about concept erasure.

  2. fitted input called prediction [Section 4.2 and Section 5.1, Explicit Content (NSFW) Erasure]
    "for nudity removal we let D(x0∥x∗0) = NudeDetect(x0) which outputs a scalar of how much explicit content is in x0, and we try to minimize that. ... we run a NudeNet detector [3] to count how many images contain nudity or sexual content."

    The training objective for explicit-content erasure is the NudeDetector output, and the reported evaluation metric is the same NudeDetector detection rate. Thus the headline '2% nudity' result is a direct measure of how well the model minimized its own training loss, not an independent assessment of concept erasure. This makes the empirical claim on the NSFW benchmark partially forced by construction.

full rationale

Several load-bearing parts of the paper reduce to their own premises. Theorem 3.1 is explicitly a restatement of a previously known sufficient condition, with the proof replaced by a reference to an earlier argument. Corollary 3.4 is a tautology: a statistic that depends only on timesteps unaffected by the edit is unchanged. Theorem 3.5 is gradient descent restated as a formal guarantee. On the empirical side, the NSFW training loss is the NudeNet detector and the NSFW evaluation metric is the same detector, so that headline number is partly the optimized objective itself. These are genuine circularities because the claimed derivation or prediction is equivalent, in each case, to its input assumption or training signal. The method does retain some independent content: most object, celebrity, and style evaluations use external classifiers or human judgments, and the closed-form update plus LoRA finetuning is a concrete algorithmic proposal. A separate, non-circular correctness concern is that Eq. (1) does not actually satisfy the sufficient condition of Theorem 3.1: it maps e_w to W e_u but does not make the edited projections of e_w and e_u coincide, so the theoretical guarantee does not cover the proposed initialization. That inconsistency affects validity, not circularity, and is reflected in the partial score rather than a higher one.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central method depends on several fitted hyperparameters (t0, eta, loss weights, LoRA rank, neutral embedding) that are not reported, and on unproven domain assumptions about score-network Lipschitzness and semantic factorization of the denoising trajectory. No new entities are introduced.

free parameters (5)
  • t0 = not specified
    The timestep threshold after which the trajectory-constrained loss applies; the paper says 'from some t0 onward' without giving the value or selection procedure (Section 4.2).
  • eta (closed-form damping) = not specified
    Damping factor scaling the rank-one attention update, 'determined on a validation set' (Section 4.1).
  • loss weights (alpha_t, beta_t, lambda) = not specified
    Weights in the trajectory loss L(theta') (Theorem 3.5, Section 4.2) are not reported.
  • LoRA rank = 4 (example)
    Rank of per-concept adapters set to a small value (e.g. 4), with no ablation over rank.
  • neutral embedding e_u = EOS token embedding
    The neutral token used as erasure target; the paper notes it could be a learned vector, but uses EOS without validating that EOS is semantically neutral (Section 4.1).
assumptions (4)
  • domain assumption Score network epsilon_theta is L-Lipschitz in its parameters
    Used in Theorem 3.3 to bound KL divergence for unrelated prompts; no proof or estimate of L given.
  • domain assumption Diffusion trajectory semantics separate into early layout and late details, with a spontaneous semantic breaking point
    The trajectory-constrained loss relies on this (cited to [29], [20]) without verification for SD1.5 and FLUX.
  • domain assumption Induction over attention layers and denoising steps preserves equality of trajectories when key/value vectors coincide
    The proof of Theorem 3.1 is asserted, not carried out; residual connections and normalization are not addressed.
  • ad hoc to paper EOS token embedding is a neutral concept replacement
    Used as the erasure target e_u for all tasks; no evidence that EOS is concept-free or that this choice works across domains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TRACE: Trajectory-Constrained Concept Erasure in Diffusion Models." pith.science (2026). https://pith.science/paper/KJWA75OE

@misc{pith2026250523312,
  author       = {Pith},
  title        = {Pith review of: TRACE: Trajectory-Constrained Concept Erasure in Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KJWA75OE}},
  note         = {Machine review of arXiv:2505.23312}
}
read the original abstract

Text-to-image diffusion models have shown unprecedented generative capability, but their ability to produce undesirable concepts (e.g.~pornographic content, sensitive identities, copyrighted styles) poses serious concerns for privacy, fairness, and safety. {Concept erasure} aims to remove or suppress specific concept information in a generative model. In this paper, we introduce \textbf{TRACE (Trajectory-Constrained Attentional Concept Erasure)}, a novel method to erase targeted concepts from diffusion models while preserving overall generative quality. Our approach combines a rigorous theoretical framework, establishing formal conditions under which a concept can be provably suppressed in the diffusion process, with an effective fine-tuning procedure compatible with both conventional latent diffusion (Stable Diffusion) and emerging rectified flow models (e.g.~FLUX). We first derive a closed-form update to the model's cross-attention layers that removes hidden representations of the target concept. We then introduce a trajectory-aware finetuning objective that steers the denoising process away from the concept only in the late sampling stages, thus maintaining the model's fidelity on unrelated content. Empirically, we evaluate TRACE on multiple benchmarks used in prior concept erasure studies (object classes, celebrity faces, artistic styles, and explicit content from the I2P dataset). TRACE achieves state-of-the-art performance, outperforming recent methods such as ANT, EraseAnything, and MACE in terms of removal efficacy and output quality.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 21 canonical work pages

  1. [1]

    Stable diffusion public release, 2022

    Stability AI. Stable diffusion public release, 2022. 1

    work page 2022

  2. [2]

    Negative prompt

    AUTOMATIC1111. Negative prompt. https : / / github . com / AUTOMATIC1111 / stable - diffusion - webui / wiki / Negative - prompt . 1, 2

    work page

  3. [3]

    Nudenet: Neural nets for nudity classification, detection and selective censoring, 2019

    P Bedapudi. Nudenet: Neural nets for nudity classification, detection and selective censoring, 2019. 8

    work page 2019

  4. [4]

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

    Jinhe Bi, Yujun Wang, Haokun Chen, Xun Xiao, Artur Hecker, V olker Tresp, and Yunpu Ma. Visual instruction tuning with 500x fewer parameters through modality linear representation-steering. arXiv preprint arXiv:2412.12359 ,

    arXiv

  5. [5]

    Prism: Self-pruning intrinsic selection method for training-free multimodal data selection

    Jinhe Bi, Yifan Wang, Danqi Yan, Xun Xiao, Artur Hecker, V olker Tresp, and Yunpu Ma. Prism: Self-pruning intrinsic selection method for training-free multimodal data selection. arXiv preprint arXiv:2502.12119, 2025

    arXiv 2025

  6. [6]

    Cot-kinetics: A theoretical modeling assessing lrm reasoning process

    Jinhe Bi, Danqi Yan, Yifan Wang, Wenke Huang, Haokun Chen, Guancheng Wan, Mang Ye, Xun Xiao, Hinrich Schuetze, V olker Tresp, et al. Cot-kinetics: A theoretical modeling assessing lrm reasoning process. arXiv preprint arXiv:2505.13408, 2025. 3

    arXiv 2025

  7. [7]

    Erasing undesir- able concepts in diffusion models with adversarial preserva- tion

    Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Mon- tague, Tamas Abraham, and Dinh Phung. Erasing undesir- able concepts in diffusion models with adversarial preserva- tion. arXiv preprint arXiv:2410.15618, 2024. 1, 3

    arXiv 2024

  8. [8]

    Erasing concepts from diffusion models

    Rohit Gandikota, Joanna Materzynska, Jaden Fiotto- Kaufman, and David Bau. Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345, 2023. 1, 2, 3, 9

    arXiv 2023

Show all 36 references
  1. [9]

    Unified concept editing in diffusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzy´nska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 5111–5120, 2024. 1, 2, 3, 9

    2024

  2. [10]

    Eraseanything: Enabling con- cept erasure in rectified flow transformers

    Daiheng Gao, Shilin Lu, Shaw Walters, Wenbo Zhou, Ji- aming Chu, Jie Zhang, Bang Zhang, Mengxi Jia, Jian Zhao, Zhaoxin Fan, et al. Eraseanything: Enabling con- cept erasure in rectified flow transformers. arXiv preprint arXiv:2412.20413, 2024. 2, 9

    2024 arXiv

  3. [11]

    Reliable and efficient concept erasure of text-to- image diffusion models

    Chao Gong, Kai Chen, Zhipeng Wei, Jingjing Chen, and Yu- Gang Jiang. Reliable and efficient concept erasure of text-to- image diffusion models. In European Conference on Com- puter Vision, pages 73–88. Springer, 2024. 3

    2024

  4. [12]

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

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt im- age editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3

    2022 arXiv

  5. [13]

    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. arXiv preprint arXiv:2106.09685, 2021. 2

    2021 arXiv

  6. [14]

    Robin: Robust and invisible watermarks for diffusion models with adversarial optimization

    Huayang Huang, Yu Wu, and Qian Wang. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. Advances in Neural Information Processing Systems, 37:3937–3963, 2024. 3

    2024

  7. [15]

    Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux , 2024. Accessed: Febru- ary 21, 2025. 2

    2024

  8. [16]

    Bilateral personalized dialogue generation with contrastive learning

    Bin Li and Hanjun Deng. Bilateral personalized dialogue generation with contrastive learning. Soft Computing, 27(6): 3115–3132, 2023. 3

    2023

  9. [17]

    VPAI Lab at MedVidQA 2022: A two-stage cross-modal fu- sion method for medical instructional video classification

    Bin Li, Yixuan Weng, Fei Xia, Bin Sun, and Shutao Li. VPAI Lab at MedVidQA 2022: A two-stage cross-modal fu- sion method for medical instructional video classification. In Proceedings of the 21st Workshop on Biomedical Language Processing, pages 212–219, Dublin, Ireland, 2022....

    2022

  10. [18]

    Distinct but correct: generating diversified and entity-revised medical re- sponse

    Bin Li, Bin Sun, Shutao Li, Encheng Chen, Hongru Liu, Yixuan Weng, Yongping Bai, and Meiling Hu. Distinct but correct: generating diversified and entity-revised medical re- sponse. Science China Information Sciences, 67(3):132106, 2024

    2024

  11. [19]

    To- wards better chinese-centric neural machine translation for low-resource languages

    Bin Li, Yixuan Weng, Fei Xia, and Hanjun Deng. To- wards better chinese-centric neural machine translation for low-resource languages. Computer Speech & Language, 84: 101566, 2024. 3

    2024

  12. [20]

    Set you straight: Auto-steering denoising tra- jectories to sidestep unwanted concepts

    Leyang Li, Shilin Lu, Yan Ren, and Adams Wai-Kin Kong. Set you straight: Auto-steering denoising tra- jectories to sidestep unwanted concepts. arXiv preprint arXiv:2504.12782, 2025. 2, 3, 9, 10

    2025 arXiv

  13. [21]

    Towards visual-prompt temporal answer grounding in instructional video

    Shutao Li, Bin Li, Bin Sun, and Yixuan Weng. Towards visual-prompt temporal answer grounding in instructional video. IEEE transactions on pattern analysis and machine intelligence, 46(12):8836–8853, 2024. 3

    2024

  14. [22]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceeding...

    2014

  15. [23]

    Tf-icon: Diffusion-based training-free cross-domain image composi- tion

    Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf-icon: Diffusion-based training-free cross-domain image composi- tion. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 2294–2305, 2023. 3

    2023

  16. [24]

    Mace: Mass concept erasure in diffu- sion models

    Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffu- sion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6430– 6440, 2024. 1, 2, 3, 6, 9

    2024

  17. [25]

    Robust watermarking using generative pri- ors against image editing: From benchmarking to advances

    Shilin Lu, Zihan Zhou, Jiayou Lu, Yuanzhi Zhu, and Adams Wai-Kin Kong. Robust watermarking using generative pri- ors against image editing: From benchmarking to advances. arXiv preprint arXiv:2410.18775, 2024. 3

    2024 arXiv

  18. [26]

    On aliased resizing and surprising subtleties in gan evaluation

    Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 2

    2022

  19. [27]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  20. [28]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and 11 Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485–5551, 2020. 9

    2020

  21. [29]

    Spontaneous symme- try breaking in generative diffusion models

    Gabriel Raya and Luca Ambrogioni. Spontaneous symme- try breaking in generative diffusion models. arXiv preprint arXiv:2305.19693, 2023. 2

    2023 arXiv

  22. [30]

    Low-rank adaptation for fast text-to-image 715 diffusion fine-tuning

    Simo Ryu. Low-rank adaptation for fast text-to-image 715 diffusion fine-tuning. https : / / github . com / cloneofsimo/lora. 2

  23. [31]

    Safe latent diffusion: Mitigating inappro- priate degeneration in diffusion models

    Patrick Schramowski, Manuel Brack, Bj ¨orn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappro- priate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522–22531, 2023. 1, 2, 8

    2023

  24. [32]

    Stereo: Towards adversarially ro- bust concept erasing from text-to-image generation models

    Koushik Srivatsan, Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Stereo: Towards adversarially ro- bust concept erasing from text-to-image generation models. arXiv preprint arXiv:2408.16807, 2024. 1

    2024 arXiv

  25. [33]

    Tree-ring watermarks: Fingerprints for diffu- sion images that are invisible and robust

    Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffu- sion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 3

    2023 arXiv

  26. [34]

    Ich-scnet: Intracerebral hemorrhage segmenta- tion and prognosis classification network using clip-guided sam mechanism

    Xinlei Yu, Ahmed Elazab, Ruiquan Ge, Hui Jin, Xinchen Jiang, Gangyong Jia, Qing Wu, Qinglei Shi, and Chang- miao Wang. Ich-scnet: Intracerebral hemorrhage segmenta- tion and prognosis classification network using clip-guided sam mechanism. In 2024 IEEE International Conference...

    2024

  27. [35]

    Ichpro: Intracerebral hem- orrhage prognosis classification via joint-attention fusion- based 3d cross-modal network

    Xinlei Yu, Xinyang Li, Ruiquan Ge, Shibin Wu, Ahmed Elazab, Jichao Zhu, Lingyan Zhang, Gangyong Jia, Taosheng Xu, Xiang Wan, et al. Ichpro: Intracerebral hem- orrhage prognosis classification via joint-attention fusion- based 3d cross-modal network. In 2024 IEEE International ...

    2024

  28. [36]

    Ich-prnet: a cross-modal intracere- bral haemorrhage prognostic prediction method using joint- attention interaction mechanism

    Xinlei Yu, Ahmed Elazab, Ruiquan Ge, Jichao Zhu, Lingyan Zhang, Gangyong Jia, Qing Wu, Xiang Wan, Lihua Li, and Changmiao Wang. Ich-prnet: a cross-modal intracere- bral haemorrhage prognostic prediction method using joint- attention interaction mechanism. Neural Networks , 184...

    2025

Pith tools

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