Research map / Image generation / Text-to-image diffusion

HEART: Hyperspherical Embedding Alignment via Kent-Representation Traversal in Diffusion Models

NeurIPS 2026 · Image generation › Text-to-image diffusion · approach: Compositional text-to-image guidance · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HEART, a training-free method that edits subjects and attributes in diffusion text-embedding space by geodesic traversal on the hypersphere guided by Kent-distribution concept geometry.

Problem

Controlling subject and attribute edits in text-to-image diffusion models through text conditioning often causes background drift and entangled changes because text embeddings are treated as Euclidean with linear edits.

Approach

Text-encoder token embeddings lie approximately on a hypersphere where concepts are anisotropic Kent distributions; edits are performed as geodesic (slerp/exp-map) traversals with position-aware propagation to EOT and downstream tokens, requiring no finetuning or inversion.

Key results

• Subject swap: HEART reaches Acc 0.620 with LPIPS 0.292, versus LEDITS++ Acc 0.947 / LPIPS 0.284 and SEGA Acc 0.460; HEART-inv reaches Acc 0.958.
• Attribute control on SDXL: HEART-A achieves ΔCLIP Age 2.76 and Hair 0.94 in about 32 s with near-zero memory and no training, versus Concept Sliders at 1263 s and 3.87 GB.
• Kent distribution is selected by BIC for all concept-encoder pairs (100% selection rate) over vMF and moVMF.

Surprising findings

• Plain HEART subject swap has much lower accuracy (0.620) than the Direct Swap regeneration baseline (0.993) and LEDITS++ (0.947), so the training-free version trades target accuracy for structure preservation.
• Linear-distance nearest neighbors in token embedding space return incoherent tokens (characters, random strings), whereas angular neighbors are semantically coherent.

Limitations

• Operates only in text embedding space, so fine-grained spatial or object-localized edits are hard, especially in compositional scenes.
• Inversion-based pipelines can give better reconstruction fidelity for real-image editing.
• Requires more than about 5 denoising steps, so FLUX-Schnell (4 steps) shows weak control.

Experimental setup

Datasets: LAION-style prompts template-based prompt pools

Benchmarks: subject swap prompt pairs (internal) attribute control prompts (age, smile, hair, chubby)

Metrics: Acc CLIP-T LPIPS ΔCLIP time memory

Baselines: Direct Swap LEDITS++ SEGA Concept Sliders Attribute Control Text Sliders

Builds on: LEDITS++ SEGA Concept Sliders Text Sliders

Closest papers by meaning

• Rare Text Semantics Were Always There in Your Diffusion Transformer (NeurIPS 2025 · similarity 0.88)
• A Cat Is A Cat (Not A Dog!): Unraveling Information Mix-ups in Text-to-Image Encoders through Causal Analysis and Embedding Optimization (NeurIPS 2024 · similarity 0.87)
• Localizing and Editing Knowledge In Text-to-Image Generative Models (ICLR 2024 · similarity 0.87)
• Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.87)
• SweetDreamer: Aligning Geometric Priors in 2D diffusion for Consistent Text-to-3D (ICLR 2024 · similarity 0.86)
• Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.86)