Controlling Embedding Spaces with Text-Conditioned Transformations

ECCV 2026
1 Institute of Artificial Intelligence, University of Central Florida
2 Adobe Research

Abstract

Multimodal embedding spaces in models like CLIP enable powerful capabilities such as semantic similarity retrieval and cross-modal zero-shot classification. These embeddings compress high-level semantics into a single vector, which comes at the cost of primarily expressing a dominant semantics like main object while suppressing other important attributes such as camera angle or color tone. We propose a text-conditioned transformation of visual embeddings that makes such attributes explicitly accessible. Given a natural language description of an attribute category (e.g., “color” or “art style”), a network generates an affine transformation that emphasizes the specified attribute. Conditioning on text enables it to learn many attributes simultaneously, accessing them at inference time through an intuitive interface. The network is trained to align transformed embeddings with the frozen latent space, enabling retrieval using existing large-scale embeddings without any re-encoding. When applied to a full set, the same mechanism transforms the latent space for attribute disentanglement tasks such as multi-clustering. By operating directly in latent space, our method provides a unified and efficient framework for controlling embedding spaces, demonstrating state-of-the-art performance across both attribute-based retrieval and multi-attribute organization tasks with near-zero inference cost.

Overview

ControlEmbed maps a natural-language attribute description such as “color,” “camera angle,” or “mood” to an affine transformation of frozen visual embeddings. Applying the transform only to a query enables attribute-aware retrieval over an unchanged gallery while applying it globally reorganizes the same embedding set for attribute-specific clustering.

Diagram of a text-conditioned hypernetwork generating query-only and global embedding transformations for retrieval and clustering.
Figure 1: Overview of text-conditioned embedding-space transformations. (Left) We propose a hypernetwork that maps natural-language attribute descriptions to a modification function. Applying this transformation to a query embedding produces an attribute-aware representation that emphasizes the requested property. (Right) By controlling the embedding space with attribute transforms, the same mechanism enables both attribute-aware retrieval and attribute-specific clustering.

Method

Multi-attribute training workflow using a text-conditioned affine hypernetwork, multi-positive contrastive learning, and multimodal alignment.
Figure 2: Multi-attribute training workflow. An attribute category is sampled (e.g., “action” or “location”) and passed through the proposed hypernetwork to generate a transformation function, which is applied to a query embedding to produce attribute-focused representations aligned with base embeddings of images sharing the same category attribute label via a multi-positive contrastive loss. An additional multimodal guidance loss further aligns transformed features with the text embedding of the corresponding attribute label. The process iterates over the full set of valid attributes, where the set of positive/negative samples varies based on selected category.

Text-conditioned transformation. ControlEmbed leaves the pretrained vision-language model frozen and learns one small generator of feature-space transformations. An attribute phrase such as “color” or “camera angle” is encoded as \(\mathbf{t}\), and the hypernetwork \(H_{\phi}\) converts that representation into an affine matrix \(\mathbf{W}\) and bias \(\mathbf{b}\). The resulting operator acts directly on a frozen image embedding \(\mathbf{v}\):

\[ \mathbf{t}=f_t(T),\quad \mathbf{v}=f_v(I),\qquad (\mathbf{W},\mathbf{b})=H_{\phi}(\mathbf{t}),\quad \mathbf{v}'=\mathbf{v}\mathbf{W}^{\top}+\mathbf{b}. \]

Changing the text input selects a different comparison axis without producing a new encoder or attribute-specific checkpoint. Because the vision and text encoders remain frozen, the method operates entirely on stored features and does not require image re-encoding.

Multi-attribute training. At each optimization step, one attribute taxonomy is chosen and its labels determine the positive groups in the batch. A transformed query is compared against unmodified base embeddings: samples sharing the selected attribute label are treated as positives, while the remaining samples provide negatives. Rotating through taxonomies lets the same hypernetwork learn many attribute controls without forcing unrelated label spaces into a single fixed grouping.

\[ \mathcal{L}_{SR}=-\frac{1}{N} \sum_{i,j=1}^N \frac{\mathds{1}_{i,j}}{\sum_{k=1}^N \mathds{1}_{i,k}} \log \frac{\exp(\vv_{i}^{'\top} \vv_{j} / \tau)}{\sum_{k=1}^N \exp(\vv_{i}^{'\top} \vv_{k} / \tau)} \]

Here, the feature vectors are \(\ell_2\)-normalized, \(N\) is the batch size, \(\tau\) controls the contrastive temperature, and \(\mathds{1}_{i,j}\) identifies pairs with the same label under the active taxonomy. The loss pulls each transformed query toward matching frozen features while preserving the original space as the retrieval target. A multimodal alignment term additionally relates transformed image features to text embeddings of the attribute labels, grounding the learned transformations in language.

Two inference modes. Query mode changes only the query vector, leaving gallery features untouched so existing indexes can be reused. Global mode applies the generated operator to every stored vector, yielding an attribute-specific space for clustering. Both modes use the same trained hypernetwork and differ only in where the transformation is applied.

Results

Attribute-aware retrieval. ControlEmbed transforms only the query while searching the original frozen gallery, preserving compatibility with precomputed large-scale indexes. Base features reach 33.8 mAP on Clevr-4, 43.7 on Stanford40, and 25.2 on ShotBench, while our query mode reaches 77.0, 86.3, and 41.0 mAP, respectively. A single hypernetwork supports multiple categories instead of storing one adapter per attribute.

Multi-clustering. The same learned transform can be applied to the full embedding set, followed by Faiss k-means and Hungarian matching. Our global mode reaches 73.0 clustering accuracy on Clevr-4, 80.0 on Stanford40, and 51.6 on ShotBench with near-instant inference time. Our proposed query-focused training method naturally handles this attribute disentanglement setting and achieves the highest average clustering accuracy across datasets while using a single transformation network trained once for all attributes.

Table comparing frozen-gallery retrieval mAP for ControlEmbed and baseline methods on Clevr-4, Stanford40, and ShotBench.
Table 1: Frozen-gallery attribute-aware retrieval results reported as mAP averaged across dataset taxonomies. ControlEmbed simultaneously handles multiple categories with one set of weights while substantially improving performance. † denotes an MLP adapter trained with the multi-positive loss.
Table comparing multi-clustering accuracy and inference time for ControlEmbed and baseline methods on Clevr-4, Stanford40, and ShotBench.
Table 2: Multi-clustering results averaged over 14 attribute categories. Inference time assumes an evaluation set of 5,000 images. ControlEmbed improves performance with minimal compute overhead and near-instant inference. * denotes results reported on the full action set.
COCO retrieval examples comparing base CLIP similarity with color- and location-conditioned transformations.
Figure 3: Qualitative attribute-aware retrieval results under different attribute transformation functions. Our text-conditioned query transformations selectively retrieve images that match the specified attribute without losing the scene semantics at first, then prioritize the attribute over general semantics. From top to bottom, we show retrievals for "action", "count", then fine-grained perceptual property "camera angle".

Ablations and Qualitative Analysis

Data efficiency and capacity

Figure 4 shows near-peak Stanford40 performance with six training samples per class. Figure 5 shows less than 1.5% variation while scaling one hypernetwork to 19 attributes and 187 subclasses.

Line plot of Stanford40 clustering accuracy versus training samples per class for action, location, and mood.
Figure 4: Multi-clustering accuracy on Stanford40. As the number of training images per class varies, the transformation approaches peak action, location, and mood performance with six images per class.
Line plot of mean retrieval mAP as one hypernetwork scales to more jointly trained attribute categories.
Figure 5: Capacity across jointly trained attribute categories. Mean category mAP remains stable as training scales from one category per model to a unified hypernetwork supporting more attributes.

Backbone generalization

Figure 6 shows gains across a diverse set of both uni- and multi-modal backbones: CLIP, SigLIP, DINOv2, and Qwen3-VL-Embed.

Comparison of base, query-transformed, and globally transformed retrieval mAP across CLIP, SigLIP, DINOv2, and Qwen3-VL-Embed backbones.
Figure 6: Retrieval mAP across backbone architectures. Query and global transformations improve attribute-aware retrieval for CLIP, SigLIP, DINOv2, and Qwen3-VL-Embed representations.

Embedding geometry

Figure 7 shows shot-size labels changing from entangled clusters to a continuous, interpretable axis after transformation.

Side-by-side t-SNE plots of shot-size embeddings before and after transformation, colored by shot-size label.
Figure 7: Shot-size embedding geometry before and after transformation. The transformed space changes entangled baseline clusters into a continuous progression through the shot-size taxonomy.

Cross-domain qualitative behavior

Figure 3 shows color- and location-conditioned similarity on MS-COCO images, despite training the corresponding transforms on smaller source datasets.

Conclusion

We presented a text-conditioned transformation framework that enables fine-grained control over pretrained embedding spaces without modifying encoder weights or recomputing gallery features. A single hypernetwork learns to generate attribute-specific affine transformations from natural-language descriptions, supporting both attribute-aware retrieval and multi-attribute organization with near-zero inference overhead. While the method generalizes across diverse attributes, its expressiveness is currently bounded by the range of categories seen during training. Future work may explore scaling up training to find emergent transformation properties across attributes completely unseen during training.

BibTeX

@inproceedings{fioresi2026controlling,
  title={Controlling Embedding Spaces with Text-Conditioned Transformations},
  author={Fioresi, Joseph and Caba Heilbron, Fabian and Nathani, Pankaj and Shah, Mubarak and Kafle, Kushal},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}