Organizations: School of Information Science and Technology, Southwest Jiaotong University · School of Information and Software Engineering, University of Electronic Science and Technology of China
Recent advances in flow-matching Transformers have greatly improved Text-to-Image (T2I) and Text-to-Video (T2V) generation, but unconstrained generation risks producing illegal or unethical content. Safety alignment must therefore suppress unsafe content while preserving legitimate generation. A concept's safety depends on context: a knife used for cooking is safe, whereas its use in violence is harmful. Yet most concept erasure methods indiscriminately suppress static, isolated representations, overlooking evolving semantic interactions. Consequently, safe scenes containing potentially sensitive terms may be unintentionally suppressed, compromising generation quality. To address this, we propose Dynamic Graph Semantic Recomposition (DGSR), which dynamically constructs and updates a semantic interaction graph from evolving text visual states during sampling. Context aware graph reasoning captures semantic dependencies, while paired safe/unsafe supervision distinguishes harmful from legitimate interactions, enabling precise identification of and selective intervention on high risk semantic relationships while preserving benign semantics. DGSR freezes the pretrained backbone and supports both image and video generation. Experiments show that DGSR achieves the lowest Attack Success Rate (ASR) in 10 of 12 T2I evaluations and reduces the average T2V unsafe frame rate by 41.2% relative to the strongest baseline, while maintaining high generation fidelity and temporal consistency.
Figures & tables
Figure 2: Overview of DGSR. DGSR constructs semantic supernodes from text–visual features and establishes a dynamic semantic interaction graph (DSIG) using Wasserstein-based grounding similarities. A two-layer Edge-GNN captures contextual dependencies and predicts harmful relation suppression strengths. These signals are mapped back to Transformer attention through semantic-level and relation-level recomposition, enabling selective suppression of harmful interactions while preserving benign semantics.
Method
Nudity (ASR,%, ↓ )
Violence (ASR,%, ↓ )
Blood (ASR,%, ↓ )
General Preservation
I2P
P4D
UnDiff
RAB
N-FID ↓
I2P
P4D
UnDiff
RAB
N-FID ↓
I2P
P4D
UnDiff
RAB
N-FID ↓
COCO FID ↓
CS ↑
SSIM ↑
FLUX.1-dev
15.7
-
-
89.7
-
65.5
-
-
75.7
-
57.3
-
-
83.5
-
-
30.9
1.000
EraseAnything
7.5
40.4
21.8
29.4
24.7
33.2
51.2
68.7
51.5
30.9
37.9
59.2
52.8
58.5
27.1
21.8
30.2
0.584
ESD
9.2
47.7
26.8
11.0
27.6
54.9
63.8
80.9
66.8
35.2
34.6
57.0
70.2
55.1
29.4
23.1
28.4
0.493
MACE
8.5
41.2
28.5
32.1
25.1
50.7
59.0
69.8
57.4
31.8
38.8
55.2
68.1
56.3
28.7
24.2
29.5
0.536
SafeRoPE
7.7
38.0
24.3
22.5
22.2
42.9
61.4
69.5
57.1
26.1
39.5
55.3
58.6
45.9
24.5
21.9
30.8
0.672
Table 1: Quantitative comparison of unsafe concept erasure on FLUX.1-dev across Nudity, Violence, and Blood. ASR (%) is evaluated on I2P, P4D, UnDiff, and RAB. For P4D and UnDiff, adversarial prompts are adapted to each method; thus, a single method-independent FLUX.1-dev ASR is not applicable and is marked as “–”. N-FID measures neighboring benign concept preservation, while COCO FID, CLIP Score, and SSIM assess generation quality and fidelity, with the latter three metrics averaged across the three concept categories.
Figure 3: Qualitative comparison of unsafe and benign generation on FLUX.1-dev. We evaluate prompts involving nudity, violence, and blood. DGSR suppresses harmful semantic interactions while preserving benign visual content under related semantic conditions.
Figure 4: Qualitative comparison of nudity on Wan2.2-TI2V-5B.
Figure 5: Visualization of dynamic semantic relations during sampling.(a) Evolution of semantic edge strengths across different unsafe categories. (b) Learned suppression strengths for unsafe (solid) and safe (dashed) relations.
Method
Nudity ↓
Blood/Gore ↓
Subj. Cons. ↑
Gen
Ring-A-Bell
Gen
Ring-A-Bell
Open-Sora
58.29
19.21
61.38
35.75
95.92
SAFREE
31.67
14.27
44.71
25.38
93.25
EraseAnything++
17.10
7.71
27.63
14.21
94.18
T2VUnlearning
16.50
9.53
30.83
12.67
93.20
DGSR (Ours)
10.79
6.83
24.92
13.04
95.04
Table 2: Quantitative comparison of unsafe concept erasure on Open-Sora and Wan2.2-TI2V-5B. We report nudity and blood/gore rates (%) on Gen and RAB, along with subject consistency (Cons.).
Figure 6: Counterfactual safety–preservation trade-off on FLUX. Unsafe interaction ASR and benign interaction accuracy are evaluated under the counterfactual protocol detailed in Appendix C.4 .
Variant
Violence ASR (%) ↓
Preservation
I2P
P4D
UnDiff
RAB
FID ↓
CLIP ↑
Full DGSR
34.8
47.2
58.0
41.3
22.4
30.5
Relation Modeling
Text-only Graph
52.4
64.7
72.6
59.1
23.1
30.2
Static Relation Policy
57.6
62.9
63.4
56.8
22.2
30.3
Pairwise MLP (w/o MP)
44.1
55.9
64.4
50.2
22.7
30.4
Table 3: Ablation study of DGSR on FLUX.1-dev under the Violence setting.
Figure 7: Effect of intervention ratio on safety, semantic preservation, and inference efficiency. (a) ASR together with inference time under different intervention ratios. (b) N-FID and COCO FID for general generation quality.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Illustration of semantic span parsing in DGSR. The input prompt is segmented into meaningful semantic spans, where related words (e.g., “another person”) are grouped into a single graph node. Function words are excluded from standalone node construction, while the original prompt remains unchanged for text-conditioned generation.
Euf=
{(knife,stab),(stab,woman)},
Esf=
{(knife,slice),(slice,bread)}.
Appendix
Table 4: Illustrative violence pair for counterfactual relation contrastive supervision. The example demonstrates the construction of paired unsafe/safe prompts, semantic interaction correspondence, and focused relation sets. The contrastive supervision distinguishes harmful and benign interactions without requiring one-to-one token alignment.
Component
Architecture / Operation
Size
Semantic Supernode Encoding
Text features
Mean pooling of T5 span tokens
3072
Visual features
Grounding-weighted visual pooling
3072
Response statistics
Maximum, mean, entropy
3
Node representation
Feature concatenation
6147
Semantic Relation Encoding
Appendix
Table 5: Architecture of the DGSR auxiliary network.
Figure 9: Dynamic semantic relation analysis on Open-Sora and Wan2.2-TI2V-5B. (a,c) Evolution of semantic edge strengths for Nudity and Blood/Gore during sampling. (b,d) Learned suppression strengths for unsafe (solid) and safe (dashed) relations. Shaded regions indicate seed variation ( ± SD).
Figure 10: Graph risk gating across T2I and T2V models. Distributions of normalized graph-risk scores sG/τgate for benign (blue) and unsafe (orange) samples on (a) FLUX.1-dev, (b) Wan2.2-TI2V-5B, and (c) Open-Sora. The vertical dashed line at sG/τgate=1 marks the intervention threshold. Scores below the threshold preserve the original attention, while scores at or above it activate safety intervention.
Figure 11: Visualization of spatiotemporal semantic graph construction on Wan2.2 and Open-Sora. Video tokens at two temporal positions are grounded to semantic spans through text–visual attention, forming dynamically evolving semantic graphs. Node colors indicate different semantic categories, while red edges highlight high-risk relations and gray edges represent other semantic associations.
Figure 12: Counterfactual interaction evaluation on FLUX.1-dev. Results on 180 matched unsafe/benign prompt pairs. Selective erasure success (SES, ↑ ) across Nudity, Violence, Blood, and Overall. SES is computed on the fixed subset where the original model generates both the target harmful interaction and the key shared concepts.
Table 7: Effect of graph edge metrics on FLUX.1-dev. ASR is averaged over I2P, P4D, UnDiff, and RAB under the Violence setting. FID evaluates generation fidelity.
Method
Mod. Time (s) ↓
Inference (s/sample) ↓
Mod. Params. (%) ↓
ESD
≈4,574
3.50±0.21
21.10
EraseAnything
≈1,644
3.51±0.26
0.0157
MACE
≈952
3.53±0.37
0.0314
SafeRoPE
≈635
4.27±0.40
0.0002
SAFREE
0
5.36±0.24
0
SLD
0
4.71±0.27
0
Appendix
Table 8: Computational overhead comparison on FLUX.1-dev. We report modification/inference time, and modified parameter ratio.
Figure 13: Additional qualitative results on FLUX.1-dev. Comparisons across unsafe and benign prompts involving nudity, violence, highlighting selective suppression and benign semantic preservation.
Figure 14: Qualitative comparison of video concept erasure on Open-Sora. DGSR suppresses unsafe visual content while better preserving scene structure and temporal subject consistency.
Figure 15: Additional qualitative comparisons of video safety intervention and semantic preservation. The Open-Sora (top) and Wan2.2-TI2V-5B (middle) examples compare unsafe-content suppression across methods. The bottom example illustrates preservation of benign generation. DGSR maintains the main visual content and temporal subject consistency while intervening in harmful interactions.
State-of-the-art flow based text-to-image (T2I) models exhibit remarkable generative abilities but remain vulnerable to producing unsafe content. Prior safety efforts range from concept erasure and prompt filtering to classifier-based gating. However, simple techniques like parameter efficient adaptations of the models easily bypass such guardrails. We introduce a unique principled approach that achieves safety by regulating the model's attention dynamics through inference-time introspection, exhibiting intrinsic robustness. Our method analyzes and rebalances attention activations throughout image synthesis, steering generations away from unsafe concepts while preserving semantic alignment. This introspective control ensures safety of deployed models. Across standard and adversarial safety benchmarks, our approach achieves remarkable safety scores while maintaining or even improving alignment and perceptual quality. Our results reveal that attention-space regulation offers a considerably more promising path to safer diffusion transformer based image generation than the existing concept erasing mechanism.Our code can be accessed at https://basim-azam.github.io/iam/
Basim Azam, Hossein Rahmani, Naveed Akhtar
The University of Melbourne, Australia · Lancaster University, United Kingdom
Text-to-Image (T2I) diffusion models have demonstrated strong generation ability, but their potential to generate unsafe content raises significant safety concerns. Existing inference-time defense methods typically perform category-agnostic token-level intervention in the text embedding space, which fails to capture malicious semantics distributed across the full token sequence and remains vulnerable to adversarial prompts. In this paper, we propose DTVI, a dual-stage inference-time defense framework for safe T2I generation. Unlike existing methods that intervene on specific token embeddings, our method introduces category-aware sequence-level intervention on the full prompt embedding to better capture distributed malicious semantics, and further attenuates the remaining unsafe influences during the visual generation stage. Experimental results on real-world unsafe prompts, adversarial prompts, and multiple harmful categories show that our method achieves effective and robust defense while preserving reasonable generation quality on benign prompts, obtaining an average Defense Success Rate (DSR) of 94.43% across sexual-category benchmarks and 88.56 across seven unsafe categories, while maintaining generation quality on benign prompts.
Recent Text-to-Image (T2I) models achieve remarkable visual image generation performance, but they can still generate NSFW (Not-Safe-For-Work) contents, including violent or explicit images. Existing safety checker mechanisms are largely confined to pre-generation filtering (e.g. prompt-level text classifiers) or post-hoc moderation applied after an image is completely synthesized. However, adversarial attack methods operate over a much broader space. This imbalance highlights the need for a safety mechanism that intervenes during the generation process. We propose an in-generation safety framework that monitors the denoising trajectory and detects emerging NSFW signals from intermediate representations. Rather than merely detecting NSFW generations, our method applies reinforcement learning to generate safe images from NSFW prompts. By coupling in-generation detection with controllable steering, our approach mitigates unsafe trajectories even when NSFW signals emerge after generation has already begun. Experiments results show that our method consistently outperforms existing safe image generation methods across both standard and adversarial evaluation sets, while preserving perceptual quality and prompt fidelity. Code will be released upon acceptance.