SAE Debias: Mitigating Gender Bias in T2I Models

• SAE Debias is a model-agnostic approach that leverages k-sparse autoencoders to isolate and suppress gender bias in text-to-image diffusion models.
• The method intervenes in the latent space by identifying gender-relevant directions and subtracting their influence to maintain generation fidelity.
• Empirical evaluations show significant bias reduction with minimal quality loss, demonstrating efficiency across various Stable Diffusion model variants.

SAE Debias denotes a class of model-agnostic approaches that mitigate representational bias—principally occupational gender bias—in text-to-image (T2I) diffusion models by leveraging the properties of k-sparse autoencoders (SAEs). Rather than manipulating prompt formulations or retraining the underlying generative architecture, SAE Debias directly intervenes in the feature space, identifying and suppressing interpretable directions associated with protected attributes across various diffusion models. This framework achieves substantial bias reduction with minimal impact on generation fidelity and requires neither model-specific fine-tuning nor architectural modifications (Wu et al., 28 Jul 2025).

1. Sparse Autoencoder Architecture and Training

SAE Debias employs a linear autoencoder with enforced sparse activations in its hidden layer to analyze and control latent features related to protected attributes. The SAE is trained on CLIP-derived text embeddings corresponding to annotated prompts (e.g., bios labeled for gender and profession).

• Encoder: Given a residual CLIP feature vector $z \in \mathbb{R}^d$ (dimension $d$ dependent on the Stable Diffusion (SD) variant), the encoder performs an affine transformation followed by a ReLU nonlinearity, expanding to $m$ latent dimensions ($m/d$ set as expansion factor $f$):

$$\tilde{h} = \mathrm{ReLU}(W_{\mathrm{enc}} z + b_{\mathrm{enc}}), \quad W_{\mathrm{enc}} \in \mathbb{R}^{m \times d},\ b_{\mathrm{enc}} \in \mathbb{R}^m$$

• k-sparsity: Hard top-$k$ sparsity is imposed:

$$h = \mathrm{Top}_k(\tilde{h}), \qquad \text{typical} \ k = 32 \text{ (SD 1.x/2.x)},\ 64 \text{ (SDXL)}$$

• Decoder: Linear mapping back to CLIP feature space:

$$\hat{z} = W_{\mathrm{dec}} h + b_{\mathrm{pre}}, \quad W_{\mathrm{dec}} \in \mathbb{R}^{d \times m}$$

with $b_{\mathrm{pre}}$ initialized to the geometric median of training $z$ vectors.

• Loss: The SAE objective:

$$\mathcal{L}_{\mathrm{SAE}} = \frac{1}{N} \sum_{i=1}^N \|z_i - \hat{z}_i\|_2^2 + \lambda \cdot \mathcal{L}_{\mathrm{aux}}$$

with $\mathcal{L}_{\mathrm{aux}}$ encouraging uniform unit activation.

This SAE is trained on the “Bias in Bios” dataset (≈257,000 biographies, 28 professions, binary gender labels) (Wu et al., 28 Jul 2025).

2. Identification of Gender-Relevant Directions

Post-training, the SAE provides a latent space where protected-attribute association is localized and interpretable. For each profession $P$, the gender-bias subspace is computed as follows:

• Sparse mean codes: For prompts describing profession $P$ and each gender,
  • Mean sparse code for male: $$\mu_{\text{male}}^{(P)} = \frac{1}{N_m} \sum_{i=1}^{N_m} f(z_{\text{male},i}^{(P)})$$
  • Mean sparse code for female: $$\mu_{\text{female}}^{(P)} = \frac{1}{N_f} \sum_{j=1}^{N_f} f(z_{\text{female},j}^{(P)})$$
• Bias direction: The difference vector in sparse code space,

$$\Delta h^{(P)} = \mu_{\text{male}}^{(P)} - \mu_{\text{female}}^{(P)}$$

This $\Delta h^{(P)} \in \mathbb{R}^m$ represents the primary direction along which gender-related stereotypical attributes shift for profession $P$.

In practice, $\Delta h^{(P)}$ aligns with interpretable feature axes such as those corresponding to “face/hair” or “clothing” latent units, revealing the localization of gender stereotypes (Wu et al., 28 Jul 2025).

3. Inference-Time Debiasing Mechanism

To intervene in downstream T2I diffusion pipelines, SAE Debias operates via inference-time residual steering.

Debias Procedure:

1. Encode: Map the prompt embedding $z$ to sparse code $h = f(z)$. For unbiased interpolation, the non-top-$k$ activations can be preserved.
2. Bias direction selection:
   • If $P$ is among known professions, set $\Delta h^{\mathrm{final}} = \Delta h^{(P)}$.
   • Otherwise, interpolate using:

   $$w_i = \frac{\exp(\cos(h, \Delta h^{(P_i)}) / T)}{\sum_j \exp(\cos(h, \Delta h^{(P_j)}) / T)}, \qquad \Delta h^{\mathrm{final}} = \sum_i w_i \Delta h^{(P_i)}$$

3. Bias suppression:

   • In the encoded (latent) space, subtract the projection of $z$ onto the normalized debias direction:

   $$z_{\mathrm{debiased}} = z - \alpha \langle z, d_P \rangle d_P,\quad d_P = \Delta h^{\mathrm{final}} / \|\Delta h^{\mathrm{final}}\|_2$$

• Alternatively, via “decoder injection”:

  $$z_{\mathrm{debiased}} = z + W_{\mathrm{dec}}(\gamma \Delta h^{\mathrm{final}})$$

  where $\gamma$ is chosen such that this is equivalent to the latent projection above.

During image generation, the diffusion model’s cross-attention to the prompt’s embedding is replaced by $z_{\mathrm{debiased}}$, which suppresses gender stereotyping in the generated output.

4. Quantitative Evaluation and Tradeoffs

The effectiveness of SAE Debias was established on multiple T2I models (Stable Diffusion 1.4, 1.5, 2.1, SDXL), using the following metrics (Wu et al., 28 Jul 2025):

• Bias metrics:

  • Mismatch Rate: Proportion of generated images whose gender (automatic BLIP-2 classifier) mismatches the gendered prompt (“a photo of a man/woman who works as a P”).
  • Composite Misclassification Rate ($MR_C$): $MR_C = \sqrt{MR_O^2 + (MR_F - MR_M)^2}$, capturing both overall error and gender disparity.
  • Skew (neutral prompt): For “a photo of a person who works as a P”, proportion of max-group images across genders.
• Image quality:
  • Inception Score (IS)
  • CLIP Score (text–image alignment)

Representative experimental results:

Model	Mismatch Rate (Base)	Mismatch Rate (SAE Debias)	Skew (Base)	Skew (SAE Debias)	IS (Base)	IS (Debias)	CLIP (Base)	CLIP (Debias)
SD 1.4	0.84%	0.06%	85.2%	83.5%	16.10	15.72	21.13	21.13
SD 2.1	0.78%	0.60%	83.0%	82.5%	–	–	–	–
SDXL	0.00%	0.96%	94.0%	87.9%	–	–	–	–

Ablations show that bias drops steeply for $\alpha \approx 0.5$, with negligible IS/CLIP degradation until higher suppression (Wu et al., 28 Jul 2025).

5. Interpretability, Control, and Reusability

SAE Debias derives its interpretability and control from the k-sparse structure of the SAE:

• The bias directions $\Delta h^{(P)}$ align with compact, attribute-specific latent subspaces.
• Cross-attention attribution maps (DAAM) demonstrate that, after debiasing, attention is distributed away from gender-salient features (e.g., facial cues) and more toward profession-indicative context.
• The same SAE and $\Delta h^{(P)}$ library are reusable across all CLIP-based diffusion pipelines without additional retraining.
• Hyperparameters $k$ (sparsity) and $\alpha$ (suppression strength) trade off between reconstruction fidelity and bias-removal efficacy; best practice is to calibrate $k$ to the model scale and set $\alpha$ between 0.4–0.6 (Wu et al., 28 Jul 2025).

6. Extensions, Limitations, and Comparisons

SAE Debias presents several notable strengths and limitations:

• Model agnosticism: Applicable to any CLIP-backed diffusion model.
• Attribute generalization: Multi-attribute debiasing (e.g., race, age) is feasible by training on datasets labeled for those features; multi-gender spectra can be handled by relabeling corpora.
• Efficiency: SAE is trained once per model family; the bias direction library is simply looked up or interpolated.
• Interpretability: Latent units correspond to human-interpretable subspaces (such as “hair length” for occupational gender stereotypes).
• Failure modes: Overly sparse ($k$ too small) or under-sparse ($k$ too large) codes degrade either reconstruction or debiasing efficacy.
• Limitations: Current method is tailored to binary gender, with suggested extensions for non-binary cases; generalization to non-occupational stereotyped attributes requires appropriate labeled data (Wu et al., 28 Jul 2025).

A plausible implication is that, as this method is fundamentally post-hoc and operates in latent space, it may not eliminate all forms of stereotype bias, particularly if such biases are distributed across highly entangled features or arise in image regions beyond text-conditioned space.

7. Relationship to Other SAE Debiasing Approaches

SAE Debias complements and differs from other SAE-based debiasing techniques:

• SP TopK: The Select-and-Project Top K (SP TopK) method (Bărbălau et al., 13 Sep 2025) employs encoder-side feature selection and geometric orthogonalization for debiasing VL embeddings, demonstrating up to 3.2× improvement in fairness metrics on retrieval tasks, but is not specifically targeted at generative pipelines.
• Tokenized SAE: In contrast to SAE Debias and SP TopK, Tokenized SAE debiasing (Dooms et al., 24 Feb 2025) directly addresses feature redundancy in LM-residual reconstruction by learning per-token offset lookups, thus enhancing feature interpretability in NLP domains.

While all three approaches leverage SAE interpretability and sparse representations, SAE Debias uniquely focuses on generative fairness in T2I models, offering a plug-and-play, cross-architecture solution with robust empirical validation (Wu et al., 28 Jul 2025).