~/shanegraffiti.com/research/forceforget
Shane Graffiti Inc. Semantic Adversarial Research Division 2026

FORCE
FORGET
CONCEPT REMOVAL
BY REWARD

Text-to-image models can still produce unsafe content, and existing concept-erasing methods tend to over-correct suppressing the benign concepts that share a prompt with a harmful one and degrading the model's ability to depict people at all. ForceForget reframes erasure as reward optimisation under reinforcement learning: it eliminates an unsafe concept while preserving safe semantic meaning, by maximising a Concept Erasing Reward that balances a safety signal against an alignment signal. A lightweight Safe Adapter regulates only part of the text embedding inside cross-attention for efficient, targeted removal. This is safety research the goal is a model that refuses to generate the unsafe concept while remaining useful everywhere else.

DivisionSemantic Adversarial Research
DomainT2I Safety / Concept Erasing
PublishedICML 2026 arXiv 2606.14351
Key Result100% removal under Ring-A-Bell & P4D
Concept Erasing Reinforcement Learning Concept Erasing Reward Safe Adapter Cross-Attention Red-Teaming Robustness Benign Preservation I2I Transfer DDPO LoRA Fine-Tuning Artistic Style Removal Object Removal Concept Erasing Reinforcement Learning Concept Erasing Reward Safe Adapter Cross-Attention Red-Teaming Robustness Benign Preservation I2I Transfer DDPO LoRA Fine-Tuning Artistic Style Removal Object Removal
§ 1.0The Over-Erasure Problem

Removing an unsafe concept is easy to overdo. Strong erasers tend to take collateral with them, and because a concept like nudity is entangled with "human," aggressive removal degrades the model's ability to depict people at all. Three failure modes recur across prior methods.

Failure 01
Benign Suppression
Edited models also mitigate the safe concepts that happen to appear in a harmful prompt, stripping out neutral meaning along with the target.
Failure 02
Human-Content Drop
Because the unsafe concept is tied to people, the strongest erasers lose utility on ordinary human-oriented generation including recognisable faces.
Failure 03
I2I Leakage
Methods tuned for text-to-image often fail in image-to-image, where an unsafe input image can re-introduce the concept the eraser was meant to block.
§ 2.0Erasure as Reward Optimisation

Most erasers alter model weights through supervised fine-tuning, where the unsafe concept is hard to define crisply. ForceForget instead frames the problem as reinforcement learning: the model generates images, a reward scores them, and policy-gradient updates push generation toward safe content. Borrowing the DDPO recipe, it uses importance sampling with trust-region clipping to keep updates stable no ground-truth target images required.

RL objective maximise expected reward J(θ) = Ec∼p(c), x₀∼pθ(x₀|c) [ r(x₀, c) ] // pθ(x₀|c): sample distribution under the training-prompt distribution p(c) // policy-gradient update via importance sampling, clipped by a trust region (PPO-style) // the reward r(x₀, c) is the Concept Erasing Reward, defined next
§ 3.0The Concept Erasing Reward

The reward is a balance of two opposing forces. A safety signal pushes generation away from the unsafe concept; an alignment signal anchors it to safe, human-oriented content so the model does not collapse into arbitrary output. Tuned together, they erase without over-erasing.

r_safe
Safety Signal
A safety evaluator scores each generated image on safe-versus-unsafe classes; weighted scores form the reward. Positive means safe content, negative flags a likely violation steering updates toward the safe domain.
r_align
Alignment Signal
An image captioner describes the output; sensitive keywords are filtered, and a CLIP score to the cleaned caption plus a target condition ("a photo of a person wearing clothing") keeps generation anchored to benign, human-oriented content.
CER = λ₁ · r_safe + λ₂ · r_align
// both rewards rescaled to [0, 1] before summing, preventing collapse toward one objective
// balanced weights (λ₁ = λ₂ = 1) gave the best erasing in practice
§ 4.0The Safe Adapter

Reward fine-tuning alone is slow to erase concepts thoroughly. The Safe Adapter accelerates it by regulating only part of the text embedding inside the cross-attention layers. The last few tokens are routed through a small learned linear layer while the rest pass through the usual LoRA projection letting the adapter dominantly absorb the unsafe concept so the bulk of the text feature stays focused on safe content.

// CLIP text sequence (77 tokens, truncated above) only the last 4 tokens route through the Safe Adapter
// front tokens carry the prompt's core meaning; transforming the tail erases broadly while preserving safe alignment

§ 5.0Erasing & Robustness

On the I2P benchmark, residual detections are counted by an automated detector across eight body-part classes; robustness is the removal rate under three red-teaming attacks; utility is held against the original SD v1.4 on COCO-30K. Lower residual and FID are better; higher attack-removal and CLIP are better.

I2P residual detections · attack-removal rate (%) · benign utility
MethodResidual ↓Ring-A-Bell ↑MMA ↑P4D ↑CLIP ↑FID ↓
SD v1.4 (base)8100.00.036.831.3319.59
ESD13363.596.383.529.8923.63
RECE9295.473.186.030.4922.12
MACE11773.199.997.828.8524.00
DuMo4599.796.497.830.5928.96
Co-Erasing5373.397.285.330.3526.97
ForceForget (this work)38100.0100.099.630.5326.73
Image-to-image transfer removal rate (%) with nude / non-nude input
MethodNude input ↑Non-nude input ↑
ESD12.487.2
SafeGen18.859.2
RECE28.660.6
MACE8.491.0
Co-Erasing9.893.8
ForceForget (this work)96.4100.0
100%
Removal under Ring-A-Bell & P4D
38
Residual detections lowest of all
90.7%
Celebrity-identity retention
96.4%
I2I removal with unsafe input
§ 6.0Beyond One Concept

Dropping the safety signal and keeping only alignment turns the same machinery toward other erasure targets showing the approach is not specific to a single concept.

Artistic Style
Erasing "Van Gogh" or "Kelly McKernan" for copyright protection. ForceForget leads the overall trade-off score (LSd 0.32 and 0.26), removing the target style while leaving untargeted styles like Picasso largely intact.
Objects
Removing whole object classes from Imagenette. Rather than scrambling the target to random content, the method gently steers it toward a related object replacing "church" with a prayer-rug-like image, "English Springer" with a generic dog while preserving unrelated classes.
Multiple Concepts
Fine-tuning to erase both "nudity" and "violence" generalises to lower inappropriate content across the broader I2P categories hate, harassment, self-harm, shocking, illegal activity measured by a general inappropriateness classifier.
Other Backbones
The adapter is architecturally agnostic to the text encoder and applies to SDXL's concatenated dual-encoder embedding, and transfers to SD v1.5 as a plug-in LoRA without retraining. Multimodal-attention models like FLUX would need further modification.

The stated limitation: a fixed reward weighting treats every unsafe prompt with the same erasing strength, which can over-erase mildly unsafe prompts or under-penalise highly unsafe ones a natural target for adaptive weighting in future work.

~/conclusion
$ query: what breaks existing erasers // they over-erase taking benign concepts and human content with them. // and they leak in image-to-image settings. $ query: what does ForceForget change // erasure becomes a balanced reward: safety + alignment. // a Safe Adapter regulates only the last few text tokens. $ query: what is the result // 100% removal under Ring-A-Bell and P4D; lowest residual detections. // benign content, faces, and I2I robustness preserved.

ERASE
THE
CONCEPT.
KEEP
THE
MEANING.