EraseSAE: Surgical Concept Erasure in Text-to-Video Diffusion Models via Sparse Autoencoders
Xinghao Wang, Dong Li, Wei Yu, Yingwei Pan, Tao Gong, Qi Chu, Nenghai Yu, Ting Yao
cs.CV · cs.AI
Quick verdict
Proposes SAE-based surgical concept erasure for T2V diffusion; strong, well-ablated results, but single-run metrics and loosely defined headline gains.
Summary
The paper proposes EraseSAE, a framework for removing target concepts (nudity, celebrity identities) from DiT-based text-to-video diffusion models while leaving unrelated content intact. It trains a Partitioned Convolutional Sparse Autoencoder (PConvSAE) at an intermediate DiT layer, splitting the latent space into a Context branch (scene background) and a Concept branch (per-concept channel partitions), supervised by cross-attention-derived masks and five loss terms. An offline contrastive attribution step contrasts activations from paired prompts to lock a small set of concept-specific kernels; at inference, per-timestep masks (foreground × concept activation) restrict a spatially-modulated classifier-free guidance to concept-active regions. On CogVideoX-5B and HunyuanVideo, the method reports nudity rates of 2.62/7.13 with SSIM 74.99/65.69 (Table 1), robustness on Ring-A-Bell (lowest detection in all six settings), celebrity erasure to 8.00/19.20 average detection with preservation of non-target identities (58.60/61.50) and the best SSIM (Table 2), plus the lowest I2P explicit count on Flux.1 (160, Table 3). Ablations support each loss term and the dual-branch design. Limitations: all results are single-run without variance or significance testing; key hyperparameters are relegated to a supplementary absent from the paper; the HunyuanVideo VideoEraser baseline is unreported; and per-concept-set SAE training costs are not quantified.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Intervening at monosemantic SAE features enables surgical concept erasure in DiT-based T2V models with better removal and preservation than prior T2V erasure methods (Tables 1-2).
- PConvSAE's dual-branch convolutional design with per-concept channel partitions yields higher-purity concept features (0.76 vs 0.37 for a linear SAE) while preserving spatiotemporal coherence (Table 5).
- Offline one-versus-max contrastive attribution locks concept-specific kernels whose suppression resists adversarial prompt reformulation (Ring-A-Bell, Table 1).
- Timestep-resolved masks with spatially-modulated CFG give the best erasure/preservation trade-off among candidate inference strategies (Table 6).
- The framework transfers to T2I, achieving the lowest explicit-content count on Flux.1/I2P (Table 3).
Strengths
Video-native SAE architecture with ablation-backed gains
PConvSAE replaces activation flattening with 2D convolutions on temporally folded tensors and a spatial-aware channel Top-K (Sec. 3.2); Table 5 shows monotonic improvement from Linear SAE (12.76 nudity rate / 56.97 SSIM) to single-branch (9.06/61.37) to dual-branch (7.13/65.69) on HunyuanVideo.
Strongest preservation results among compared methods
Table 1: SSIM 74.99/65.69 vs next best 47.02/60.67 (+27.97/+5.02 points); Table 2 celebrity SSIM 63.38/77.47, roughly double every baseline, while VBench Object Class (77.80/80.61) and Subject Consistency (94.30/96.04) stay near the unedited model.
Adversarial and cross-concept robustness is measured, not just claimed
Honest inference-strategy ablation
Weaknesses
Zero statistical rigor
Headline percentages are undefined and selectively framed
Reproducibility deferred and baselines incomplete
Narrow concept coverage vs the 'diverse tasks' claim
Per-concept-set training cost unquantified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
EraseSAE (arXiv 2609.03629v2; USTC + HiDream.ai) targets surgical concept erasure — nudity and five celebrity identities — in DiT-based T2V models (CogVideoX-5B, HunyuanVideo), with a transfer check on Flux.1 [dev] (I2P). The claimed contributions (Sec. 1): SAE-mediated intervention at monosemantic features; PConvSAE, a partitioned convolutional SAE; a contrastive attribution mechanism with timestep-resolved spatiotemporal masks; and state-of-the-art erasure 'outperforming the strongest baseline by 34.5% in erasure accuracy and 8.3% in SSIM'. The architectural and empirical claims are largely supported by the paper's own ablations; the headline percentages are the weakest link (see Statistical Evidence).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EraseSAE?
EraseSAE is a method for removing specific concepts — such as nudity or a particular celebrity's likeness — from text-to-video diffusion models. It uses sparse autoencoders to decompose the model's internal video features into interpretable units, identifies which units encode the unwanted concept, and suppresses the concept only in the video regions and timesteps where it appears, leaving backgrounds and other identities intact.
How effective is EraseSAE compared to other concept-erasure methods?
In the paper's experiments on CogVideoX-5B and HunyuanVideo, it reports the lowest nudity detection rates (2.62 and 7.13, versus roughly 2.9-68 for baselines including negative prompting, SAFREE, VideoEraser and T2VUnlearning), the highest structural similarity to the unedited output, and the lowest detection rates under the adversarial Ring-A-Bell prompt benchmark. For celebrities, target detectability dropped to 8-19% on average while four other identities remained recognizable.
Is the EraseSAE code available?
The paper states the code is available at github.com/HiDream-ai/EraseSAE. However, most implementation details (intervention layer choice, sparsity level Top-K, loss weights, thresholds) are deferred to a supplementary document, so reproducing results likely requires that material plus GPUs for training the SAE and generating evaluation videos.
How is EraseSAE different from previous concept-erasure approaches?
Earlier methods either steer inference (negative prompts, embedding edits) or globally edit model weights; both act on entangled representations and apply erasure everywhere. EraseSAE instead decomposes activations into monosemantic features with a convolutional sparse autoencoder, locks concept-specific feature kernels offline via a contrastive procedure, and at inference applies guidance only inside per-timestep masks of concept-active regions, which is why unrelated content is better preserved.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Fig. 1's embedded formulas extracted as placeholder glyphs; the figure caption and all body text and Tables 1-6 extracted cleanly, and every number cited in this review comes from those tables or body text.
- The supplementary material (implementation details referenced in Sec. 4) is not part of the provided text, so hyperparameters described as 'in the supplementary' could not be verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.