Attention, May I Have Your Decision? Localizing Generative Choices in Diffusion Models

Katarzyna Zaleska; Łukasz Popek; Monika Wysoczańska; Kamil Deja

cs.CVMar 29, 2026

Attention, May I Have Your Decision? Localizing Generative Choices in Diffusion Models

Authors: Katarzyna Zaleska, Łukasz Popek, Monika Wysoczańska, Kamil Deja

Organizations: Warsaw University of Technology · IDEAS Research Institute

Abstract

Text-to-image diffusion models exhibit remarkable generative capabilities, yet their internal operations remain opaque, particularly when handling prompts that are not fully descriptive. In such scenarios, models must make implicit decisions to generate details not explicitly specified in the text. This work investigates the hypothesis that this decision-making process is not diffuse but is computationally localized within the model's architecture. While existing localization techniques focus on prompt-related interventions, we notice that such explicit conditioning may differ from implicit decisions. Therefore, we introduce a probing-based localization technique to identify the layers with the highest attribute separability for concepts. Our findings indicate that the resolution of ambiguous concepts is governed principally by self-attention layers, identifying them as the most effective point for intervention. Based on this discovery, we propose ICM (Implicit Choice-Modification) - a precise steering method that applies targeted interventions to a small subset of layers. Extensive experiments confirm that intervening on these specific self-attention layers yields superior debiasing performance compared to existing state-of-the-art methods, minimizing artifacts common to less precise approaches. The code is available at https://github.com/kzaleskaa/icm.

Figures & tables

Figure 1: We use linear probes to localize layers with the highest attribute separability for concepts not present in the prompt (e.g., a color when prompting for “ a photo of flowers ”). We show that we can use the learned probes to steer towards less probable outcomes.

Figure 2: Overview of ICM. We identify optimal layers for steering by measuring their discriminability using an external classifier. Layers are ranked by the classification accuracy of a linear probing (denoted here as LP) on their activations, and the top-performing layers (here in purple ) are selected for targeted intervention. The selected layers can then be used for two applications: generation control via finetuning (top right) or activation steering (bottom right), enabling precise attribute manipulation while preserving image quality.

Figure 3: Comparison of mean accuracy across all timesteps for the examined concepts, evaluated on the test set.

Method	Gender (2)				Age (3)				Race (4)
Method	FD $\downarrow$	FID $\downarrow$	CLIP-I $\uparrow$	CLIP-T $\uparrow$	FD $\downarrow$	FID $\downarrow$	CLIP-I $\uparrow$	CLIP-T $\uparrow$	FD $\downarrow$	FID $\downarrow$	CLIP-I $\uparrow$	CLIP-T $\uparrow$
Original	0.564	120.06	-	0.6155	0.752	120.06	-	0.6155	0.558	120.06	-	0.6155
Finetuning [ 34 ]	0.050	161.47	0.8779	0.6095	0.746	161.47	0.8779	0.6095	0.198	161.47	0.8779	0.6095
ICM (Finetuning)	0.535	143.98	0.9187	0.6189	0.681	122.56	0.9075	0.6198	0.449	123.47	0.9020	0.6188
Latent Editing [ 17 ]	0.408	166.11	0.8253	0.6005	0.682	200.90	0.8527	0.6122	0.524	153.05	0.8804	0.6086
H-Distribution [ 26 ]	0.222	151.68	0.8475	0.6087	0.506	147.71	0.8345	0.6098	0.544	126.90	0.8255	0.6100

Table 1: Results for gender, age, and race debiasing with SD v1.5. Our ICMmethods achieve a competitive Fairness Discrepancy (FD

\downarrow

) while best preserving image quality (FID

\downarrow

) and prompt-text alignment (CLIP-T

\uparrow

). Results of competing methods from [ 35 ] .

Panel of Figure 4 — Figure 4: Accuracies of linear probes trained to predict gender from activations collected after self- and cross-attention layers of the SD 1.5 UNet. Self-attention layers exhibit generally higher discriminative power than cross-attention layers, with a clear peak towards the middle block. Results are compared for the general prompts and explicit prompts .

	Gender (2)				Age (3)				Race (4)
Method	FD ↓	FID ↓	CLIP-I ↑	CLIP-T ↑	FD ↓	FID ↓	CLIP-I ↑	CLIP-T ↑	FD ↓	FID ↓	CLIP-I ↑	CLIP-T ↑
Steer cross-attn	0.365	118.39	0.8631	0.6171	0.612	114.26	0.8991	0.6167	0.428	120.28	0.8561	0.6179
Steer self-attn	0.085	118.31	0.8556	0.6136	0.273	112.57	0.8954	0.6153	0.298	129.51	0.8403	0.6172
Finetuning cross-attn	0.535	143.98	0.9187	0.6189	0.681	122.56	0.9075	0.6198	0.449	123.47	0.9020	0.6188
Finetuning self-attn	0.463	139.04	0.9408	0.6175	0.770	131.25	0.9280	0.6186	0.524	124.36	0.9148	0.6186

Table 2: Debiasing performance comparison when applying steering after Self-Attention vs. Cross-Attention layers. Self-Attention yields a dramatically better Fairness Discrepancy (FD), while maintaining comparable image quality (FID) and prompt alignment (CLIP-T).

Figure 5: Visual ablation of layer selection for activation steering. ICMselects the 3 best layers to preserve image quality while achieving desired modifications, compared to steering: 1) the 3 worst-performing layers, or 2) all layers without selection.

Figure 6: Example generations from the SANA model with applied gender steering. We compare steering using only 6 best performing probes on self-attention layers with steering using all of the layers.

Specific Prompt	Train Acc.	Test Acc.
“a photo of a man”	70.37	62.08
“a photo of a woman”	73.37	70.54
“a photo of an older person”	84.24	68.63
“a photo of a younger person”	82.85	88.86
“a photo of a dark hair person”	92.84	88.63
“a photo of a light hair person”	88.79	80.40

Table 3: Train and test accuracies for linear probes distinguishing between activations from generic prompt ( “a photo of a person” ) and explicit (specific prompt) generation.

	FD $\downarrow$	FID $\downarrow$	CLIP-I $\uparrow$	CLIP-T $\uparrow$
General	0.089	119.45	0.883	0.613
Specific	0.087	122.08	0.850	0.614

Table 4: Comparison of steering performance when linear probes are trained on activations from general vs. specific prompts (Gender task). The same settings (the

\alpha

value and start timestep) and layers were used for each run.

Figure 7: Example generations from SDXL model steered using probes trained on prompts describing a USA president, where positive class is ‘Barack Obama’, and negative is all of the other generated presidents.

Layer Selection	Steering Success (%)	Avg. Confidence	CLIP-I $\uparrow$
Top 20 (Ours)	83.0%	0.917	0.893
Random 20	51.0%	0.838	0.932
All Layers	91.0%	0.879	0.779

Table 5: Comparison of steering 100 images towards Barack Obama using different layer selection strategies. Our method (Top 20) achieves effective steering while maintaining the highest image similarity to the original generations, avoiding the severe image degradation seen when steering all layers.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Effect of increasing $\alpha$ values along the young–old direction. Larger $\alpha$ produces stronger age-related changes.

Direction	Layers	$\alpha$
woman	up_blocks.1.attn.2.t_blocks.0.attn1	-10
	up_blocks.1.attn.1.t_blocks.0.attn1
	up_blocks.1.attn.0.t_blocks.0.attn1
	mid_block.attn.0.t_blocks.0.attn1
man	up_blocks.1.attn.2.t_blocks.0.attn1	10
	up_blocks.1.attn.1.t_blocks.0.attn1

Appendix

Table 6: Parameters used for steering across different directions.

Figure 9: Example generations showing that applying the steering vector at later timesteps preserves the overall image structure while still shifting the predicted age direction. Here, $t$ denotes the timestep at which the modification begins.

Figure 10: Test accuracy across the selected layers and timesteps for gender dataset.

Method	Gender (2)				Age (3)				Race (4)
Method	FD $\downarrow$	FID $\downarrow$	CLIP-I $\uparrow$	CLIP-T $\uparrow$	FD $\downarrow$	FID $\downarrow$	CLIP-I $\uparrow$	CLIP-T $\uparrow$	FD $\downarrow$	FID $\downarrow$	CLIP-I $\uparrow$	CLIP-T $\uparrow$
Original	0.564	120.06	-	0.6155	0.752	120.06	-	0.6155	0.558	120.06	-	0.6155
Finetuning (rank=32, selected)	0.515	128.27	0.9028	0.6172	0.699	114.34	0.8943	0.6173	0.485	127.90	0.9062	0.6190
Finetuning (rank=32, random)	0.542	133.62	0.9230	0.6177	0.733	117.61	0.8916	0.6191	0.537	123.64	0.9412	0.6159

Appendix

Table 7: Finetuning comparison for cross-attention layers selected via prompt injection versus randomly chosen layers.

Figure 11: Comparison of generations from the original model (top), steering trained on general prompts (middle), and steering trained on specific prompts (bottom).

Figure 12: Example generations showing how increasing the steering strength $\alpha$ can progressively degrade the output.

Figure 15: Example SANA images generated after injecting a specific prompt into a chosen cross-attention layer. The general prompt is “A realistic photo of a doctor sitting and writing notes” , and the specific prompt is “a photo of a woman” .

T2I Diffusion Models

Text-to-image (T2I) diffusion models generate images from prompts by learning to reverse noise under text conditioning.

See all 402 papers

Concept Guidance: Precise, Training-Free Latent Control for Text-to-Image Generation
Aug 14, 2026Nikolai Röhrich, Isabell Hans, Felix Krause +1Diffusion Model Guidance T2I Generation
LMU Munich, Germany · Konrad Zuse School of Excellence in Reliable AI (relAI), Germany · Munich Center for Machine Learning (MCML), Germany
Training-Free Debiasing of Diffusion Models via CLIP-Guided Denoising Optimization
Jul 1, 2026Dain Kim, Jinseo Kim, Sungyong BaikDiffusion Model Guidance Text Embeddings
Dept. of Artificial Intelligence, Hanyang University, South Korea · Dept. of Data Science, Hanyang University, South Korea
DebFilter: Eradicating Biases Stashed in Value
May 27, 2026Seung Hyuk Lee, Songkuk KimDiffusion Model Guidance Cross-Attention
School of Integrated Technology, BK21 Graduate Program in Intelligent Semiconductor Technology Yonsei University

Aug 14, 2026J/K move · Enter open · S save

Concept Guidance: Precise, Training-Free Latent Control for Text-to-Image Generation

Nikolai Röhrich, Isabell Hans, Felix Krause, Björn Ommer

LMU Munich, Germany · Konrad Zuse School of Excellence in Reliable AI (relAI), Germany · Munich Center for Machine Learning (MCML), Germany

Diffusion Model Guidance T2I Generation Controllable Image Generation Diffusion Guidance Diffusion Models

Open paper AskPDF arXiv

Text-to-image diffusion models have two major drawbacks that severely limit their practical utility: (1) standard models lack an intrinsic mechanism for continuous, concept-specific guidance (e.g., for precisely controlling how aesthetically pleasing an image looks), and (2) they lack reliability for tasks requiring high local coherence (e.g., generating text or human hands). To tackle these issues, we introduce a novel notion of concept-wise mutual information and find large, concept-dependent differences between individual layers, demonstrating that the generation of specific structures is localized in distinct parts of the network. We exploit this insight by reinforcing the impact of concept-relevant layers in Concept Guidance (CoG), a precise, target-specific guidance method that works for models out-of-the-box without additional training, external models, gradients, or prompt engineering. CoG first quantifies each layer's concept-specific impact and then guides denoising using a weighted combination of predictions generated with concept-relevant layers skipped. We demonstrate performance increases across various targets and popular models like PixArt-alpha, SD3, SD3.5, and FLUX.1-dev. Code is available at https://github.com/CompVis/concept_guidance

Abstract

T2I Diffusion Models

Explore similar work

Concept Guidance: Precise, Training-Free Latent Control for Text-to-Image Generation

Training-Free Debiasing of Diffusion Models via CLIP-Guided Denoising Optimization

DebFilter: Eradicating Biases Stashed in Value