Decoupled Safety Control (DSC): a safety control algorithm for training-free safety steering
Decoupled Safety Control: A Safety-Control Algorithm for Training-Free Safety Guidance
DSC changes safety guidance to route based on the current state; on SD-v1.4's MMA-Diffusion, SLD+DSC has an ASR of 0.169.
Open-source text-to-image diffusion models may still produce unsafe content during sampling and are vulnerable to adversarial prompts. DSC does not modify weights, intervening at denoising steps of existing training-free safeguards; c− denotes unsafe concept, unsafe subspace, or unsafe reference.
- Text-to-image diffusion models still generate unsafe content, and open-source models are also vulnerable to adversarial prompts. Training-free guidance often binds avoidance targets, current signals, intervention construction, and stopping timing to the same unsafe condition; semantic entanglement may leave residual risk or harm quality and prompt fidelity.
- DSC does not change base weights. At each step, it extracts evidence from the current denoising state, estimates control states for three routes—suppress, balance, release—and selects a route, then constructs an update executed with strength and release coefficients. Instantiated on SLD and SAFREE.
- On SD-v1.4 and MMA-Diffusion, the ASR for SLD+DSC is 0.169 and for SAFREE+DSC is 0.167, lower than the original safeguard and SGF. CoPro average IP is 0.131 and 0.088. The authors state FID is better relative to SGF.
- The authors state that evaluation uses published prompt sets, automatic evaluators supplemented by human annotation, which cannot fully cover all deployment settings; safety mechanisms may still be probed or bypassed. Experimental backbone is SD-v1.4 and SD-v3, with DSC attached to SLD and SAFREE.
- Evaluated models
- Stable Diffusion v1.4 (SD-v1.4)SD-v3 (SD3-medium)
- Benchmarks
- Ring-A-BellUnlearnDiffMMA-DiffusionCOCO-30KCoProCOCO-3K
- Metrics
- ASRTRIPFIDCLIP score
Decoupled Safety Control (DSC) is a plug-and-play safety control algorithm for training-free safety steering that makes interventions determined by the current generation state. It targets the problem that existing coupled safety control chains tend to leave unsafe content or degrade benign generation quality when harmful and benign semantics are entangled. On standard safety benchmarks, it improves representative latent-space and text-space defenses, suppressing unsafe content more strongly while maintaining benign generation quality and efficient inference.
Paper analysis
What problem does this paper attempt to solve?
Training-free safety guidance binds multiple control responsibilities to the same unsafe condition; DSC instead intervenes based on the current generation state.
Training-free safety guidance often binds the avoidance target, current signal, intervention construction, and stopping timing to the same unsafe condition; when semantics are entangled, it can leave residual risk or damage benign generation.
- Scenario: The authors state diffusion models have strong text-to-image generation capabilities but still generate unsafe content, and open-source models remain vulnerable to adversarial prompts (Section 1). Weight-modifying methods require additional optimization or direct parameter changes and may reduce benign generation quality.
- Limitations of existing training-free methods: SLD, SAFREE, Safe Denoiser, SGF, etc., intervene at different sites, but the authors state many safeguards still use the same unsafe condition to do all four things simultaneously. When factors are explicit and aligned with that condition, they can be effective; when harmful and benign semantics are entangled in the current trajectory, the same condition is insufficient to decide what to preserve, which intervention to use, and how long to intervene.
- Authors' judgment: Interventions may be too weak, leaving residual unsafe content, or too broad, harming benign quality or prompt fidelity. The authors relate this to quality degradation under hard filtering, expected task performance decline, quality loss from overly long negative guidance, and mode shift on complex prompts, and consider the control process insufficiently explicit.
- This paper proposes: Decoupled Safety Control (DSC), which writes training-free safety guidance as explicit control dependent on the current generation state, attached to SLD and SAFREE, without changing base weights.
- Threat model:
- Goal: Suppress unsafe content related to safeguard condition c− and preserve benign semantics in prompt condition p that are unrelated to unsafe factors.
- Knowledge: The paper does not use white-box, black-box, gray-box, or no-box.
- Capability: Access to an existing training-free pipeline P, without changing weights; can read xt, p, and c− at each step. c− is an unsafe concept, unsafe subspace, or unsafe reference set.
- System: Experiments on SD-v1.4 and SD-v3, targeting adversarial nude prompts and multi-category inappropriate content.
What related research exists?
Related work is divided into weight modification, latent or prompt space guidance, and trajectory-level negative guidance; the authors place the difference in the organization of the control chain.
The authors discuss three categories of related work based on whether weights are modified and whether intervention falls on representation or trajectory.
Training-based safety methods
- Concept erasure and weight editing: ESD (Gandikota et al., 2023), RECE (Gong et al., 2024), UCE (Gandikota et al., 2024), R.A.C.E. (Kim et al., 2024), MACE (Lu et al., 2024), machine unlearning for image-to-image generation models (Li et al., 2024), STEREO (Srivatsan et al., 2025), Selective Amnesia (Heng and Soh, 2023). The authors claim they can remove target unsafe concepts but require retraining or direct parameter modification, and may reduce benign generation quality. This paper instead uses inference-time control without modifying the base.
Training-free latent and prompt space methods
- SLD and SAFREE: SLD (Schramowski et al., 2023) modifies classifier-free guidance with predefined safety prompts; SAFREE (Yoon et al., 2025) filters unsafe prompt representations and adaptively reweights attention during sampling.
- Other variants: The paper lists latent guarding, embedding guidance, prompt purification, prompt-side steering, without per-paper comparison to this work.
- Focus rationale: The authors claim they focus on these two because they represent latent and text space safeguards respectively, and their interventions are still derived from pre-specified unsafe concepts, subspaces, or references.
Trajectory-level safety guidance
- Modifying sampling paths: Safe Denoiser (Kim et al., 2026), Shielded Diffusion (Kirchhof et al., 2025), Safety-Guided Flow (SGF, Kim et al., 2026) directly modify sampling paths using unsafe references, repulsion fields, or time-varying negative guidance.
- Negative guidance analysis: Ban et al. (2024), Koulischer et al. (2025), Li et al. (2025) analyze how negative prompts or negative guidance affect denoising. The introduction also cites hard filtering, expected task performance degradation, overly long negative guidance, and mode shift on complex prompts as existing observations.
- Authors' distinction: These works handle training-free safety by altering the denoising trajectory; the authors claim this paper focuses on how the safety control chain is organized after introducing training-free interventions.
Baselines and benchmarks
- Baseline methods: Training-based ESD, RECE; training-free SLD, SAFREE; SGF as a recent enhancement baseline on the same underlying safeguard, compared with DSC.
- Benchmarks: Adversarial nudity with Ring-A-Bell, UnlearnDiff, MMA-Diffusion; benign generation with COCO; multi-category with CoPro.
The authors place the difference in the control process: unsafe specifications remain fixed by safeguard conditions, while runtime evidence, control state, intervention routes, update construction, and execution timing are resolved from the current generation state, not directly inherited from a single unsafe condition.
How does the paper address this problem?
DSC extracts evidence, estimates route state, selects intervention, constructs update, and executes by stage at each step, without replacing the original safeguard conditions.
Decoupled Safety Control (DSC) does not change the base model or replace the original safeguard condition, but instead turns each protected denoising step into control determined by the current generation state (Figure 2, Algorithm 1).
Problem Setting
- Protected step: Initial noise xT comes from standard normal. Under prompt condition p and safeguard condition c−, one step is written xP t-1 = Ht(xt, p, c−; P). c− is the unsafe concept, unsafe subspace, or unsafe reference set used by that safeguard.
- Goal: Suppress unsafe content related to c− while preserving benign semantics in p that are unrelated. The authors state that when semantics are entangled, one must also decide whether to intervene, what to suppress, what to preserve, and how strong at the current step.
- Decoupled quantities: The unsafe specification remains fixed; runtime evidence, control state, route, update, and execution timing are resolved from the current state.
Evidence and Route State
- Evidence: et = Φt(xt, p, c−; P). The authors state c− specifies what to avoid, but is not necessarily the most informative observation at the current step. Evidence can come from denoising branch differences, prompt-side filtering cues, predicted clean sample relationships, or current support patterns.
- Quadruple: Route r's state includes unsafe pressure πt(r), retention demand κt(r), harmful–benign entanglement ηt(r), and support scope ωt(r), i.e., st(r)=(πt(r),κt(r),ηt(r),ωt(r)). It summarizes the current control needs of that route.
- How computed: π is the sim between the suppression direction and unsafe-side evidence, κ is the sim between the retention direction and benign retention evidence, η is the leak between the two directions (only non-negative overlap). When spatial or token support maps exist, ω is their mean; otherwise 1. The denominator includes a stability term ε; the paper does not give a value for ε.
Routing, Update, and Execution
- Three routes: rsup is direct suppression, rbal is balanced control that also preserves, rrel is release, returning to baseline dynamics. The acceptable set requires κ≥τκ, η≤τη, ω≥τω, and excludes rrel; among these, choose the one with largest π; if empty, choose rrel. The same c− can thus lead to suppression, fidelity, or release.
- Update: After selecting a route, construct ũt in the safeguard space. In Appendix D, rsup uses a support-map-modulated suppression direction; rbal additionally subtracts a term along the retention direction, scaled by entanglement; rrel's update is 0. The authors state corrections are stronger when unsafe pressure is high, and milder when retention demand is high or leakage is stronger.
- Execution coefficients: αt is clipped from the ratio of baseline correction to ũt norm, ρt depends on the selected state and sampling progress. The executed correction is ut=(1−ρt)αtũt, then fused with the baseline protected step. Larger α means stronger correction, larger ρ means closer to baseline. On the release route, update is 0 and ρt=1, thus preserving the baseline step. Before the release start, ρt is 0 for non-release routes.
Attaching to SLD and SAFREE
- SLD: Evidence comes from three denoising branches: prompt, unconditional, and safety concept. The unsafe side is the difference between prompt branch and safety concept branch; the retention side is the difference between prompt branch and unconditional branch. Route quantities are computed in latent denoising space. SLD's safeguard condition is not replaced.
- SAFREE: Evidence includes prompt-to-subspace proximity, filtered prompt-side representation, resulting denoising difference, and support patterns. The unsafe side is the difference between prompt prediction and filtered text prediction; the retention side is the difference between filtered branch and unconditional branch. SAFREE pipeline or weights are not changed.
- Fixed parameters (Table D.2): Fixed per safeguard on a calibration split disjoint from test benchmarks, not tuned per benchmark. τκ=0.80, τη=0.56, τω=0.08, cref=1.0, ρmax=0.85, release start jrel=0.72T. Strength intervals: SLD [0.98, 1.28], SAFREE [0.94, 1.22]. The paper does not specify the calibration set name or size.
External Judgments
- Not scored within the algorithm: Nude tasks use NudeNet ASR and TR, multi-category uses IP. Thresholds and reviews are in Question 4.
What experiments did the paper conduct?
In Table 1, SLD+DSC and SAFREE+DSC have lower ASR than original safeguards and SGF.
Main experiments compare adversarial nude and benign generation on SD-v1.4, and measure multi-category inappropriate content on CoPro; DSC is attached to SLD and SAFREE.
Experimental Setup
- Backbone and sampling: Main experiments use Stable Diffusion (SD) v1.4, DDPM, 50 steps; compatibility experiments use SD-v3 (official SD3-medium). Unless otherwise stated, methods share the same benchmark assets and evaluation settings (Section 4, Appendix B).
- Scale: Ring-A-Bell uses 79 selected prompts; UnlearnDiff uses 116 nude-related prompts; MMA-Diffusion uses 1000 prompts. COCO-30K measures benign generation; FID and CLIP in Table 1 use the first 10K pairs. CoPro uses CoPro-10K with seven categories: harassment, hate, illegal activity, self-harm, sexual, shocking, violence. The timing protocol is COCO-3K.
- Baselines: Training-based ESD, RECE; training-free SLD, SAFREE; SGF on the same safeguard. DSC is denoted as SLD+Ours and SAFREE+Ours in tables.
- Metrics: ASR is the proportion of images where NudeNet nude probability exceeds 0.6; TR is the mean of that probability. IP uses the existing Q16 protocol. FID uses pytorch-fid, CLIP uses ViT-B/32.
- Supplementary and repetition: Appendix E reports prompt-cluster bootstrap 95% intervals for MMA-Diffusion, as well as NudeNet, Q16, ShieldGemma2, Llama Guard 4, and human annotation. Human annotation involves 5 PhD students, 1000 images per method arm, with method identity hidden. The paper does not report repetition counts for main tables or annotator agreement.
- Ablation scope: Five-phase replacement is done only on SAFREE and MMA-Diffusion, replacing one DSC phase at a time with SAFREE default behavior (Appendix B.4).
Main Results
The table below is from Table 1 (SD-v1.4; lower ASR is better). COCO FID and CLIP are on the first 10K. ESD omitted (ASR 0.456, 0.422, 0.628 respectively).
Swipe horizontally to read the table
Method Ring-A-Bell ASR UnlearnDiff ASR MMA-Diffusion ASR COCO FID COCO CLIP SD-v1.4 0.797 0.809 0.962 25.04 31.38 RECE 0.177 0.284 0.651 33.94 30.29 SLD 0.481 0.629 0.881 36.47 29.28 SLD+SGF 0.228 0.353 0.297 36.83 28.13 SLD+Ours 0.089 0.060 0.169 32.68 28.65 SAFREE 0.278 0.353 0.601 25.29 30.98 SAFREE+SGF 0.051 0.164 0.423 23.73 30.36 SAFREE+Ours 0.025 0.137 0.167 20.95 30.36 - The authors state that relative to the original safeguard, ASR on MMA-Diffusion decreases by 80.8% (SLD) and 72.2% (SAFREE), and on UnlearnDiff by 90.5% (SLD) and 61.2% (SAFREE); these are relative reductions, not percentage points. TR trends are the same. TR for SLD+Ours is 0.095, 0.088, 0.201; for SAFREE+Ours is 0.078, 0.180, 0.225 (order as in table, Table 1).
- The authors state that relative to SGF, FID is better for both safeguards, and CLIP is equal or higher. Relative to unprotected SD-v1.4, SLD series have higher FID and lower CLIP; SLD+Ours CLIP is also lower than original SLD. In the table, RECE ASR on MMA-Diffusion is still higher than both Ours.
- The authors state that SD-v1.4 is fragile on these benchmarks, and original SLD and SAFREE remain fragile on MMA-Diffusion and UnlearnDiff. Figure 1 provides qualitative comparison; the authors state that attaching to SLD or SAFREE avoids generating inappropriate images. Prompts are not reproduced here.
Multi-category IP (CoPro)
Table 2, same SD-v1.4 and category split. Lower IP is better.
Method Avg. IP CLIP SD-v1.4 0.223 29.81 SLD 0.170 29.65 SLD+Ours 0.131 29.02 SAFREE 0.153 28.91 SAFREE+Ours 0.088 28.38 - The authors state that all seven categories decrease. SAFREE+Ours per-category IP: harassment 0.101, hate 0.062, illegal activity 0.062, self-harm 0.116, sexual 0.045, shocking 0.085, violence 0.146.
- The authors consider CLIP still competitive, with limited prompt alignment loss. In the table, adding DSC lowers CLIP relative to the corresponding original safeguard.
- Figure C.4 shows qualitative examples for illegal activity; the authors state outputs are steered to safer images while preserving the general scene. Prompts are not reproduced.
Mechanism Analysis
- Runtime organization: Figure 3 x-axis is denoising steps 0–49, y-axes are step-wise mean and selection rate. The authors state that in the SAFREE default chain, prompt-to-subspace proximity and token filtering ratio are determined by prompt-side concept analysis and fixed across steps, with filtered embeddings used according to the denoising schedule; under DSC, selected routes dominate most sample-step decisions, and default selection remains small. The figure has annotated numbers, but plain text cannot reliably map to legend, so these numbers are not assigned to curves.
- State and update: Figure 4 annotates r=-0.56 between retention demand and unsafe pressure on the selected route; the authors state the negative correlation indicates they are not interchangeable. Routing margin is positive in 82.4% of explicit control decisions (relative to SAFREE default behavior under the same evidence). The text states route conditional update norm ranges from 1.13 to 2.01 across unsafe pressure quartiles, and from 1.95 to 1.15 across retention demand quartiles; Figure 4 annotates 2.01 and 1.15. The authors state this is consistent with the intended routing and update construction.
- Are phases necessary?: The authors state that replacing any DSC phase with SAFREE default behavior results in higher ASR and TR than the full method; per-item numbers are below.
SD-v3, Timing, and Supplementary Evaluations
- SD-v3 (Table 4): Ring-A-Bell ASR/TR: SD-v3 0.304/0.330, SAFREE 0.278/0.298, SAFREE+Ours 0.038/0.073. COCO-30K FID/CLIP from 22.99/31.24 to 17.54/30.82. The authors state SAFREE alone improves little; adding the method reduces ASR and TR, improves FID, and CLIP is 30.82.
- Timing (Table 5): Single NVIDIA H100 80GB HBM3, COCO-3K. SAFREE 4.938 s/image, SAFREE+SGF 5.181, SAFREE+Ours 5.096. The authors state an increase of 0.158 s/image relative to SAFREE.
- Intervals and other reviewers: Table E.1 MMA-Diffusion ASR: SLD+Ours 0.169 [0.147, 0.192], paired difference vs SGF −0.128 [−0.156, −0.101]; SAFREE+Ours 0.167 [0.145, 0.190], paired difference −0.256 [−0.288, −0.224]. The authors state both paired intervals are below 0. Table E.2: DSC relative to SGF on the same safeguard has lower unsafe rates under NudeNet, Q16, ShieldGemma2, Llama Guard 4, and human; human rates are SLD+Ours 0.153, SAFREE+Ours 0.131 (SGF 0.441, 0.479). On the same SAFREE+Ours images, ShieldGemma2 0.072, Llama Guard 4 0.069, NudeNet 0.167; the authors do not explain differences between evaluators. Table E.3: CoPro seven-category average is lower for DSC under all four protocols; human rates are SLD 0.183 vs 0.119, SAFREE 0.130 vs 0.107, with the latter gap smaller than the original IP 0.153 vs 0.088. Full COCO-30K (Table C.1): SAFREE+Ours FID/CLIP 17.14/30.49, SLD+Ours 26.54/29.05.
Other Ablations and Analysis
- Removing runtime evidence extraction: ASR 0.394, TR 0.448 (absolute increase vs SAFREE+Ours ΔASR +0.227, ΔTR +0.223; Table 3).
- Removing control-state estimation: ASR 0.419, TR 0.470 (ΔASR +0.252, ΔTR +0.245).
- Removing intervention routing: ASR 0.452, TR 0.494 (ΔASR +0.285, ΔTR +0.269).
- Removing update construction: ASR 0.462, TR 0.507 (ΔASR +0.295, ΔTR +0.282).
- Removing phase-aware execution: ASR 0.492, TR 0.525 (ΔASR +0.325, ΔTR +0.300), highest among the five.
- Baseline: SAFREE+Ours ASR 0.167, TR 0.225; SAFREE itself 0.601, 0.618. All on MMA-Diffusion, SAFREE pipeline. The paper does not report threshold sweeps.
What are some points that can be further explored?
The authors state that existing protocols cannot cover all deployment settings, and DSC should be seen as a layer of inference-time control in text-to-image generation.
The authors mainly write limitations in that evaluation protocols cannot cover all deployment settings, and mechanisms may still be probed or bypassed.
Limitations and Future Directions Noted by the Authors
- Task scope: DSC is an inference-time safeguard for text-to-image generation (Appendix A).
- Evaluation materials: Like other compared safety methods, empirical evaluation uses published prompt sets, with automated safety evaluators supplemented by human annotation (Appendix A).
- Deployment coverage: The authors state these protocols provide a consistent comparison basis but cannot fully cover all deployment settings (Appendix A).
- Method positioning: The authors state the method should be understood as a layer of practical safety control in text-to-image systems (Appendix A).
- Possibility of bypass: The authors state safety mechanisms may still be probed or bypassed in practice, and sensitive deployment scenarios should be paired with benchmark-aware evaluation and appropriate review (Appendix A).
- Future experiments: The Conclusion restates results under the evaluated settings and does not list specific future experiments.
Experimental Coverage
- Generation backbones are SD-v1.4 (DDPM, 50 steps) and SD-v3 (official SD3-medium); DSC is attached to SLD and SAFREE, and compared with ESD, RECE, SGF (Section 4, Table 1, Table 4).
- Adversarial nude prompts are Ring-A-Bell 79, UnlearnDiff 116, MMA-Diffusion 1000; COCO-30K for FID/CLIP, Table 1 uses first 10K; timing on single H100, COCO-3K (Appendix B, Table 5).
- CoPro uses CoPro-10K with seven categories, IP via Q16 protocol (Appendix B.2, Table 2).
- Five-phase ablation only on SAFREE and MMA-Diffusion; thresholds fixed on a calibration split disjoint from test sets and shared across benchmarks; the paper does not specify the calibration set name or size (Section 4.3, Appendix B.4, D).
- Appendix E reports prompt-cluster bootstrap 95% intervals for MMA-Diffusion, and evaluates with NudeNet, Q16, ShieldGemma2, Llama Guard 4, and 5 PhD students (1000 images per arm, method identity hidden); CoPro has additional seven-category average protocols. The paper does not report main table repetition counts, annotator agreement, full annotation instructions, or per-category sample sizes for CoPro human evaluation.
Summarize the main content of the paper
DSC does not modify diffusion weights, routes SLD/SAFREE interventions based on current state, and reduces unsafe rates on evaluated benchmarks.
Decoupled Safety Control (DSC) is an inference-time control algorithm attached to training-free text-to-image safeguards, without modifying base weights.
- Problem: The authors argue that many training-free methods bind what to avoid, current unsafe signals, how to construct corrections, and how long corrections last to the same unsafe condition. When harmful and benign semantics are entangled in the current trajectory, interventions may be too weak, leaving unsafe content, or too broad, harming quality and prompt fidelity.
- Approach: At each step, extract runtime evidence, estimate unsafe pressure, preservation needs, entanglement, and support range across three routes—suppression, balance, release—then select a route based on acceptability, construct updates, and execute with strength and release coefficients. Instantiated on SLD and SAFREE; thresholds in Table D.2 are shared across benchmarks.
- Nudity benchmark: On SD-v1.4, Table 1, SLD+DSC ASR: Ring-A-Bell 0.089, UnlearnDiff 0.060, MMA-Diffusion 0.169; SAFREE+DSC: 0.025, 0.137, 0.167, all lower than corresponding original safeguards and SGF. The authors claim relative ASR reductions on MMA-Diffusion of 80.8% (SLD) and 72.2% (SAFREE) compared to original safeguards.
- Quality and multi-category: Relative to SGF, the authors claim better FID, CLIP equal or higher; SAFREE+DSC COCO FID 20.95, CLIP 30.36. CoPro average IP from SLD 0.170 to 0.131, from SAFREE 0.153 to 0.088, while CLIP lower than original safeguard.
- Other settings: On SD-v3, SAFREE+DSC Ring-A-Bell ASR/TR 0.038/0.073 (Table 4). On COCO-3K, +0.158 s/image relative to SAFREE (Table 5). Removing any control stage increases ASR and TR on MMA-Diffusion (Table 3).
- Authors' conclusion: The authors claim that organizing safety guidance into explicit runtime control can strengthen evaluated training-free safeguards without modifying the base, and preserve benign generation quality in evaluated settings. Appendix A also states that existing protocols cannot fully cover all deployment settings.