GASS: Geometry-Aware Spherical Sampling for Disentangled Diversity Enhancement in Text-to-Image Generation

Ye Zhu; Kaleb S. Newman; Johannes F. Lutzeyer; Adriana Romero-Soriano; Michal Drozdzal; Olga Russakovsky

cs.CVFeb 19, 2026

GASS: Geometry-Aware Spherical Sampling for Disentangled Diversity Enhancement in Text-to-Image Generation

Authors: Ye Zhu, Kaleb S. Newman, Johannes F. Lutzeyer, Adriana Romero-Soriano, Michal Drozdzal, Olga Russakovsky

Organizations: Laboratoire d’Informatique (LIX), CNRS, École Polytechnique, IPP, France · Department of Computer Science, Princeton University, USA · FAIR at Meta - Montreal, Canada · McGill University, Canada ·

Abstract

Despite high semantic alignment, modern text-to-image (T2I) generative models still struggle to synthesize diverse images from a given prompt. In this work, we enhance the T2I diversity through a geometric lens. Unlike most existing methods that rely primarily on entropy-based guidance to increase sample dissimilarity, we introduce Geometry-Aware Spherical Sampling (GASS) to enhance diversity by explicitly controlling both prompt-dependent and prompt-independent sources of variation. Specifically, we decompose the diversity measure in CLIP embeddings using two orthogonal directions: the text embedding, which captures semantic variation related to the prompt, and an identified orthogonal direction that captures prompt-independent variation (e.g., backgrounds). Based on this decomposition, GASS increases the geometric projection spread of generated image embeddings along both axes and guides the T2I sampling process via expanded predictions along the generation trajectory. Our experiments on different frozen T2I backbones (U-Net and DiT, diffusion and flow) and benchmarks demonstrate the effectiveness of disentangled diversity enhancement with minimal impact on image fidelity and semantic alignment.

Figures & tables

Figure 1: Illustration of our geometric decomposition of sample diversity and GASS enhancement method in CLIP space. We decompose the diversity of generated image batches from T2I models in the CLIP hypersphere along two orthogonal axes: text embedding $\mathbf{e}_{t}$ (i.e., prompt-dependent) and our identified direction $\mathbf{u}_{\text{ind}}$ (i.e., prompt-independent). Our GASS method explicitly expands the geometric spread along both axes, thus enhancing the diversity of generated images across prompt-dependent content (e.g., object viewing angles) and prompt-independent visual attributes (e.g., backgrounds).

Figure 2: Illustration of our proposed Geometry-Aware Spherical Sampling (GASS) method. At the generation inference step $t$ , original T2I sampling first estimates the predicted clean image $\hat{\mathbf{x}}_{0|t}$ based on the intermediate noisy samples $\mathbf{x}_{t}$ , and then predict the noise we should remove from $\mathbf{x}_{t}$ to get $\mathbf{x}_{t-1}$ . Our GASS alters the predicted clean image from $\hat{\mathbf{x}}_{0|t}$ to $\hat{\mathbf{x}}^{*}_{0|t}$ through the geometric expansion (see Sec. 4.1 ) and the gradient-based optimization (see Sec. 4.2 ), thus guiding the iterative sampling process with frozen generative backbones.

Img Source	ClipScore	$\mathcal{D}_{\text{dep}}$	$\mathcal{D}_{\text{ind}}$	$SPP$
SD2.1	0.303 $\pm$ 0.03	0.071 $\pm$ 0.02	0.075 $\pm$ 0.02	0.146 $\pm$ 0.04
SD3-M	0.302 $\pm$ 0.02	0.060 $\pm$ 0.03	0.065 $\pm$ 0.03	0.126 $\pm$ 0.05
Real	0.293 $\pm$ 0.02	0.110 $\pm$ 0.03	0.110 $\pm$ 0.02	0.220 $\pm$ 0.05

Table 1: Quantitative comparison of spherical spread scores (SPP) between generated and real images. Real images show greater coverage spread than generated ones in both prompt-dependent (

\mathcal{D}_{\text{dep}}

) and prompt-independent (

\mathcal{D}_{\text{ind}}

) measures. Mean and std reported over 1000 classes from ImageNet.

Figure 3: Non-cherry-picked qualitative comparisons with other diversity enhancement methods on ImageNet ( Russakovsky et al., 2015 ) and Drawbench ( Saharia et al., 2022 ) . Compared to other methods (i.e., PG ( Corso et al., 2024 ) , CADS ( Sadat et al., 2024 ) , IG ( Kynkäänniemi et al., 2024 ) , and SPELL ( Kirchhof et al., 2025 ) ), our proposed GASS generates images with both richer semantic variation (e.g., object poses and layout) and more detailed and diverse backgrounds.

Base Models	SD2.1					SD3-M
Methods	Density $\uparrow$	Coverage $\uparrow$	VS $\uparrow$	ClipScore $\uparrow$	SPP $\uparrow$	Density $\uparrow$	Coverage $\uparrow$	VS $\uparrow$	ClipScore $\uparrow$	SPP $\uparrow$
Vanilla Guidance (CFG)	1.130	0.623	31.826	0.308	0.143	1.105	0.588	28.119	0.308	0.137
PG-ICLR’24	1.039	0.614	32.082	0.309	0.146	1.103	0.586	28.119	0.308	0.129
CADS-ICLR’24	0.897	0.567	28.935	0.306	0.142	1.374	0.636	28.456	0.309	0.133
IG-NeurIPS’24	0.901	0.574	28.935	0.306	0.140	1.389	0.627	27.415	0.310	0.129
SPELL-ICML’25	0.912	0.578	32.601	0.305	0.144	1.105	0.585	28.433	0.302	0.128

Table 2: Quantitative evaluations on ImageNet with SD2.1 and SD3-M as base T2I models.

Base Models	SD2.1				SD3-M
Methods	VS $\uparrow$	ImageReward $\uparrow$	ClipScore $\uparrow$	SPP $\uparrow$	VS $\uparrow$	ImageReward $\uparrow$	ClipScore $\uparrow$	SPP $\uparrow$
Vanilla Guidance (CFG)	8.599	0.217	0.306	0.122	8.115	0.779	0.318	0.113
PG-ICLR’24	8.637	0.231	0.306	0.121	8.002	0.799	0.318	0.113
CADS-ICLR’24	8.784	0.173	0.305	0.125	8.118	0.726	0.316	0.112
IG-NeurIPS’24	8.761	0.215	0.306	0.125	8.002	0.798	0.318	0.112
SPELL-ICML’25	8.726	0.205	0.305	0.131	8.166	0.671	0.315	0.112

Table 3: Quantitative evaluations on DrawBench with SD2.1 and SD3-M as base T2I models.

Prompt	VS $\uparrow$	ImageReward $\uparrow$	ClipScore $\uparrow$	SPP $\uparrow$
Short	8.301 $\rightarrow$ 8.535	0.748 $\rightarrow$ 0.698	0.322 $\rightarrow$ 0.321	0.113 $\rightarrow$ 0.121
Medium	7.663 $\rightarrow$ 7.918	1.115 $\rightarrow$ 1.043	0.318 $\rightarrow$ 0.322	0.103 $\rightarrow$ 0.107
Long	7.549 $\rightarrow$ 7.935	0.572 $\rightarrow$ 0.622	0.310 $\rightarrow$ 0.310	0.092 $\rightarrow$ 0.099

Table 4: Quantitative analysis w.r.t. different levels of prompt complexities from Drawbench. We show the performance change compared to the CFG baseline. For instance, 8.301

\rightarrow

8.535 means that the metric changes from 8.301 (CFG baseline) to 8.535 (with GASS).

Figure 4: GASS controls the source of diversity by expanding the geometric spread along specified directions. Specially, GASS on prompt-dependent axis $\mathbf{e}_{t}$ diversifies images through variations via poses and layout, while expansion along prompt-independent direction $\mathbf{u}_{\text{ind}}$ changes attributes like background and styles.

Method	VS $\uparrow$	ImageReward $\uparrow$	ClipScore $\uparrow$	SPP $\uparrow$
IP	8.203	0.774	0.308	0.113
RD	8.206	0.778	0.313	0.113
GASS	8.212	0.778	0.320	0.114

Table 5: Ablation results on different strategies for constructing the decomposition basis. IP stands for Isotropic Perturbation, and RD represents Random Direction.

GASS Variants	VS $\uparrow$	ImageReward $\uparrow$	ClipScore $\uparrow$	SPP $\uparrow$
w/o Norm.	8.876	0.732	0.313	0.123
$r_{\text{dep}}=0$ , $r_{\text{ind}}=0.02$	8.207	0.787	0.319	0.111
$r_{\text{dep}}=0.02$ , $r_{\text{ind}}=0$	8.206	0.780	0.320	0.112
$r_{\text{dep}}=r_{\text{ind}}=0.02$	8.212	0.778	0.320	0.114
$r_{\text{dep}}=r_{\text{ind}}=0.05$	8.205	0.778	0.320	0.112
t=10 (consecutive)	8.215	0.808	0.318	0.114

Table 6: Additional ablation results of GASS variants. We show the impact of the re-normalization technique, expansion range, and perturbation steps on Drawbench.

Figure 5: GASS still introduces generated image diversity, even when provided with more complex text prompts.

Figure 6: GASS can be applied sparsely on early consecutive or uniform perturbation steps along the T2I generation trajectories. The early consecutive scheduler tends to generate images with lower color saturation compared to the uniform counterpart.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Backbone	Method	Hyperparameters
SD3-M	IG-NeurIPS’24	$\sigma\in(0.2,0.8]$
	PG-ICLR’24	power coeff = 30.0
	CADS-ICLR’24	$\tau_{1}=0.9,\tau_{2}=1.20,s=0.10,\psi=1.0$
	SPELL-ICML’25	$\lambda=1.6,r=250$
SD2.1	IG-NeurIPS’24	$\sigma\in(0.1,0.9]$
	PG-ICLR’24	power coeff = 30.0

Appendix

Table 7: Hyperparameter settings for different baseline methods across SD2.1 and SD3-M.

Figure 7: Additional non-cherry-picked qualitative results comparison with other methods on ImageNet with the example class ”goldfish” .

Figure 8: Additional non-cherry-picked qualitative results comparison with other methods on Drawbench.

Figure 9: Failure case analysis. (Left): A small number of images still resemble the original ones after GASS. (Right:) When the base models can’t generate accurate counts specified by the prompt, despite GASS introducing extra diversity, it is less likely to correct these inconsistencies by itself.

T2I Generation

Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.

See all 294 papers

Breaking the Lock-in: Diversifying Text-to-Image Generation via Representation Modulation
Jun 5, 2026Dahee Kwon, Haeun Lee, Jaesik ChoiDiffusion Transformer T2I Generation
Korea Advanced Institute of Science and Technology (KAIST) · INEEJI
T-DIG: Text-Conditioned Diversity for Image Generation
Oct 21, 2026Simran Khanuja, Rohit Gandikota, Gaurav Parmar +1T2I Generation Controllable Image Generation
Carnegie Mellon University · Northeastern University
TPSO: Training-Free Diverse Image Generation via Semantic Prompt Embedding Optimization
Nov 25, 2025Debin Meng, Chen Jin, Zheng Gao +3T2I Generation T2I Diffusion Models
Queen Mary University of London · Centre for AI, AstraZeneca, UK · University of Bedfordshire +1

Jun 5, 2026J/K move · Enter open · S save

Breaking the Lock-in: Diversifying Text-to-Image Generation via Representation Modulation

Dahee Kwon, Haeun Lee, Jaesik Choi

Korea Advanced Institute of Science and Technology (KAIST) · INEEJI

Diffusion Transformer T2I Generation

Open paper AskPDF arXiv

Recent text-to-image models built on large-scale Transformer backbones and flow-based objectives deliver strong text-image alignment and high visual quality, yet often produce overly similar samples under a fixed prompt. Existing diversity-enhancement methods alleviate this issue, but typically require expensive sampling or auxiliary optimization, incurring non-trivial overhead. To investigate the root cause of this homogeneity, we examine intermediate Transformer features and observe that the zero-frequency spatial average (DC) component rapidly converges across seeds early in generation, causing early trajectory lock-in that limits downstream variation. Building on this observation, we propose DC Attenuation for diVersity Enhancement (DAVE), a training-free representation-level intervention that selectively attenuates this component in the early regime. DAVE preserves the sampling pipeline with negligible overhead, improving prompt-consistent diversity while maintaining competitive image quality.

Abstract

T2I Generation

Explore similar work

Breaking the Lock-in: Diversifying Text-to-Image Generation via Representation Modulation

T-DIG: Text-Conditioned Diversity for Image Generation

TPSO: Training-Free Diverse Image Generation via Semantic Prompt Embedding Optimization