PaperlayerView exact revision on arXiv

Source-linked AI summary

Vibe Spaces for Creatively Connecting and Expressing Visual Concepts

Huzheng Yang, Katherine Xu, Andrew Lu, Michael D. Grossberg, Yutong Bai, Jianbo Shi

arXiv:2512.14884v1cs.CV

TL;DR

Creative blending must identify which shared visual attributes matter and connect distant concepts along paths that preserve meaning. Vibe Space learns a compact graph-manifold representation for these paths, and human evaluations favor its blends over Gemini and GPT, especially for difficult pairs.

  • Problem

    Creative blending requires identifying the relevant shared attributes between images and finding nonlinear paths through feature spaces where naive interpolation can produce incoherent results.

    Source: §1 · pp. 1–2, §3.1 · p. 3

  • Method

    Vibe Space learns a low-dimensional representation aligned with multiscale graph-manifold geometry, where linear paths approximate geodesics in image feature space.

    Source: §3.1 · p. 3, §4.1 · p. 4

  • Results

    Humans preferred Vibe Space blends over Gemini and GPT, especially on difficult image pairs; on high-difficulty Totally Looks Like pairs, they preferred the method 3× more often as the second-ranked GPT.

    Source: §5.3 · p. 7, §6 · p. 8

  • Takeaways & Limitations

    A path-nonlinearity score offers a principled approach to identifying challenging image pairs for future dataset expansion.

    Source: §6 · p. 8

    Curating compelling and challenging input image pairs remains an open problem.

    Source: §6 · p. 8

Abstract

from arXiv · show

Creating new visual concepts often requires connecting distinct ideas through their most relevant shared attributes -- their vibe. We introduce Vibe Blending, a novel task for generating coherent and meaningful hybrids that reveals these shared attributes between images. Achieving such blends is challenging for current methods, which struggle to identify and traverse nonlinear paths linking distant concepts in latent space. We propose Vibe Space, a hierarchical graph manifold that learns low-dimensional geodesics in feature spaces like CLIP, enabling smooth and semantically consistent transitions between concepts. To evaluate creative quality, we design a cognitively inspired framework combining human judgments, LLM reasoning, and a geometric path-based difficulty score. We find that Vibe Space produces blends that humans consistently rate as more creative and coherent than current methods.

1. Introduction

Vibe Blending creates coherent hybrids by identifying and merging the most relevant shared attributes between images. Vibe Space finds nonlinear paths between distant concepts, while the paper also develops an evaluation framework for creative potential and blend difficulty.

1. Introduction

  • Vibe Blending merges the most relevant shared attributes between images into coherent hybrids, illustrated by combining a violin’s size with a guitar’s playing style to suggest a lute.The paper frames this as a way to connect distinct concepts through their “vibe.”Source: §1 · p. 1
  • LLMs may miss which visual attributes matter and how to merge them coherently, such as the instrument and body pose in a violin–guitar example.The paper links this challenge to attention and control, alongside LLMs’ expansive memory.Source: §1 · p. 2
  • Vibe Space learns a compact, low-dimensional manifold from context images to guide nonlinear paths between distant concepts in pretrained feature spaces such as CLIP.Steps in this “small world” correspond to meaningful transitions in the original feature space.Source: §1 · p. 2
  • The evaluation framework uses human judgments, LLMs, and a path-based blend-difficulty metric, with results indicating that LLMs are a useful starting point for quantifying creativity.Datasets include conceptually interesting pairs from Totally Looks Like and architectural design images.Source: §1 · p. 2
  • The paper introduces Vibe Blending and Vibe Space, conducts human-subject experiments, and reports that its method generates more creative blends than Gemini and GPT.The Vibe Space method learns a hierarchical graph manifold for geodesics in feature spaces with under 1M parameters.Source: §1 · p. 2

2. Related Work

The paper situates creative blending in theories of associative paths between concepts and distinguishes its attribute-focused approach from prior latent-space blending and language-based methods.

2. Related Work

  • Cognitive accounts describe creativity as traversing nonlinear paths through intermediate associations between distant concepts, motivating blends based on relevant shared attributes.The paper relates this view to semantic memory networks and creative cognition.Source: §2 · p. 2
  • Prior image-blending work includes pixel-level blending in GAN and diffusion latents, as well as methods for finding semantic directions in diffusion-model latent spaces.The cited directions include noise, weight, and text-embedding spaces.Source: §2 · p. 2
  • AID [21] fuses attention inside diffusion models but treats attributes uniformly and relies on attention correspondences rather than selecting the most relevant blend attributes.The paper contrasts this with identifying which shared attributes matter for a blend.Source: §2 · p. 2
  • Language-based creative image blending can miss precise visual attributes; the paper instead operates in image feature space and reports outperforming custom-trained diffusion morphing models and recent multimodal LLMs.Figure 2 describes diffusion-based morphing methods as struggling with distant concepts and AID [21] as missing hairstyle as the relevant attribute.Source: §2 · p. 2, Figure 2 · p. 2

3. Creative Path Finding

The method uses graph diffusion coordinates to make manifold paths easier to compute, then extends path finding across nested scales to account for both coarse and fine geometry.

3.1. Finding Paths to Connect Concepts Creatively

  • Naive latent-space interpolation can produce inconsistent or ghosted images because high-dimensional feature spaces contain implausible regions, and existing morphing methods struggle with distant or misaligned concepts.The paper hypothesizes that these “holes” reflect a mismatch between intrinsic data-manifold dimension and the higher-dimensional latent space.Source: §3.1 · p. 3
  • Graph Laplacian eigenvectors provide coordinates that capture manifold geometry, making curved paths approximately linear in graph diffusion space.The graph is based on feature similarity, and the smallest nonzero-eigenvalue eigenvectors serve as coordinates.Source: §3.1 · p. 3
  • Graph diffusion distance equals Euclidean distance in the diffusion embedding, allowing path construction through intermediate points with smoothly varying distance to the endpoints.The method interpolates diffusion coordinates and then recovers the corresponding path on the original manifold through inverse mapping.Source: §3.1 · pp. 3–4
  • Inverse mapping uses a closed-form Jacobian from eigenvalue perturbation theory and Nyström updates to align a candidate point’s diffusion coordinates with the interpolated target.The resulting path stays close to the data manifold.Source: §3.1 · p. 4

3.2. Flag Space for Multiscale Paths

  • A fixed number of eigenvectors selects one scale and can focus a path on too many or too few attributes.Leading eigenvectors describe global structure, while higher-order ones encode local variation.Source: §3.2 · p. 4
  • Flag space uses nested embeddings to represent coarse and fine manifold structure, reducing the impact of choosing the wrong number of eigenvectors.The multiscale path is recovered by minimizing average reconstruction error across scales, enforcing consistency across global and local geometry.Source: §3.2 · p. 4

4. Vibe Space and Image Generation

Vibe Space maps image features into a compact representation aligned with multiscale manifold geometry, where linear paths approximate geodesics. The pipeline matches image regions, interpolates and decodes paths, and can suppress unwanted attributes with negative examples.

4. Vibe Space and Image Generation

  • Two lightweight MLPs are trained on the fly to approximate the flag-space diffusion map and inverse mapping, with training taking under 30 seconds.Feature extraction and graph-eigenvector computation run in milliseconds.Source: §4 · p. 4

4.1. Learning the Vibe Space

  • The encoder maps dense DINO patch features into a typically 6-dimensional Vibe Space, while the decoder maps that representation to CLIP space.Vibe Space geometry aligns with multiscale flag-space diffusion maps so its Euclidean distances and linear paths correspond to manifold geodesic distances and semantic paths.Source: §4.1 · p. 4
  • Matching latent Gram similarities to flag-space kernels aligns the learned representation with nested eigen-embeddings, while random-sample regularization targets generalization beyond observed data.The encoder and decoder connect DINO features to CLIP-compatible generation.Source: §4.1 · p. 4
  • Vibe Blending trains the space, matches image-region correspondences, interpolates a path, and decodes it into generated intermediate images.The method can optionally use related exemplars, although two images suffice to train the space and identify dominant attributes.Source: §4.1 · pp. 4–5
  • A transition learned between one image pair can be reused as a displacement field on a related image to produce an analogous transition.Figure 4 illustrates this with morphing Leonardo DiCaprio’s face into a playing card.Source: §4.1 · p. 5, Figure 4 · p. 5

4.2. Creative Control

  • Negative exemplars suppress unwanted attributes by projecting the positive attribute basis away from directions spanned by the negative examples.The filtered basis is used to train Vibe Space to avoid the undesired attributes.Source: §4.2 · p. 5
  • In the rotation-and-style example, subtracting a negative style vibe leaves rotation as the blended attribute, whereas blending without negative examples transfers both.The figure describes the positive pair as specifying rotation plus style and the negative pair as specifying style to suppress.Source: Figure 5 · p. 5

5. How to Measure the Creativity of a Blend?

The section evaluates blend creativity through human judgments, path geometry, and LLM preferences, using 44 Totally Looks Like image pairs and 300 architectural pairs.

5.1. How Do Humans Rate Blend Creativity?

  • Human ratings assess Creative Potential and Blend Difficulty separately; higher difficulty tends to accompany higher potential and greater conceptual difference.Ten raters made repeated pairwise comparisons, which were aggregated into low, medium, and high bins.Source: Figure 6 · p. 6, §5.1 · p. 6

5.2. How to Estimate Blend Difficulty?

  • PNS compares a Vibe Space path with straight-line interpolation in CLIP space, combining excess path length and directional changes.PNS agrees with human-rated Blend Difficulty 80.0% of the time in high-consensus comparisons; higher PNS indicates traversal through multiple intermediate feature-space regions.Source: Figure 7 · p. 6, §5.2 · p. 7

5.3. How Do Human Preferences Reflect Creativity?

  • On high- and medium-difficulty Totally Looks Like pairs, humans prefer our method 3× and 2.4× more often, respectively, than GPT.On easier pairs, preference for GPT and CLIP Avg increases; the authors associate the smaller Architecture margin with its more conceptually related image pairs.Source: Table 1 · p. 7, §5.3 · p. 7

5.4. How Well Do LLMs Gauge Blend Creativity?

  • The LLM judge most often prefers our method, and GPT, on high-difficulty Totally Looks Like pairs, and consistently favors our method on Architecture.Raters first identify shared attributes and then rank how coherently each output blends them.Source: Table 1 · p. 7, §5.3 · p. 7, §5.4 · p. 8

5.3. How Do Human Preferences Reflect Creativity?

  • Human preference judgments show 63–77% agreement on Totally Looks Like and 66–75% on Architecture, with greater variation on Totally Looks Like.The authors report greater human consistency when evaluating blends of conceptually related images.Source: Table 2 · p. 8, §5.3 · p. 8

5.4. How Well Do LLMs Gauge Blend Creativity?

  • Human–LLM agreement rises from 35.7% and 31.3% for human top-1 blends to 55.1% and 51.8% for top-2 blends on Totally Looks Like and Architecture, respectively.The authors interpret the overlap as partial but meaningful: the LLM tends to select some blends humans rate highly.Source: Table 2 · p. 8, §5.4 · p. 8
  • LLM judges can misidentify shared attributes or overemphasize irrelevant ones, leading them to prefer a different blend than humans.Examples include overlooking the map–chicken pair’s shared shape and favoring color or texture over a shared hairstyle.Source: §5.4 · p. 8

6. Conclusion

The conclusion presents Vibe Blending as combining relevant shared visual attributes into coherent hybrids, and Vibe Space as a graph manifold guiding feature-space geodesics.

6. Conclusion

  • Vibe Space guides geodesic paths in feature spaces such as CLIP to produce blends that humans prefer over strong baselines, especially for difficult image pairs.The evaluation framework combines human and LLM preferences with a path nonlinearity score for blend difficulty.Source: §6 · p. 8
  • Curating compelling and challenging input image pairs remains an open problem.The path nonlinearity score is presented as one principled direction for identifying such examples and expanding datasets.Source: §6 · p. 8
  • More reliable and scalable evaluation is identified as a future direction, alongside improving LLM judges’ attention to key visual attributes.Source: §6 · p. 8

A. Comparisons to Current Methods

The comparison evaluates blends against user-identified shared attributes and finds Vibe Space strongest among concept-level methods, while CLIP Avg leads pixel-level methods.

A. Comparisons to Current Methods

B. Ablation Studies

Ablations on Totally-Looks-Like use Attribute-Masked DreamSim across difficulty levels to test flag loss, correspondence, and DINO–CLIP feature mixing.

B. Ablation Studies

  • Multiscale flag-loss supervision is most robust across difficulty levels; removing flag loss significantly degrades performance.Fine-only works best on low-difficulty pairs, while coarse-only works best on high-difficulty pairs.Source: B. Ablation Studies · p. 16, Table 4 · p. 17
  • Correspondence outperforms its removal on medium- and low-difficulty cases; fewer clusters help on high-difficulty pairs with less spatial alignment.Without correspondence, performance is competitive only for high-difficulty pairs and can produce incoherent blends when inputs are spatially misaligned.Source: B. Ablation Studies · p. 16, Table 5 · p. 17
  • DINO–CLIP mixing performs best on medium- and low-difficulty pairs, while CLIP-only encoding sometimes performs slightly better on the most challenging pairs.Mixing remains the more stable choice across difficulty levels.Source: B. Ablation Studies · p. 16, Table 6 · p. 18

C.1. Proof of Vibe Space Analytical Minimum

This appendix formalizes Vibe Space interpolation as an approximate inverse-diffusion geodesic and describes its graph construction, semantic correspondence, vibe transfer and suppression, gradient visualization, and adaptive blend-weight selection. It explains how the method identifies a blend weight by finding a minimum in CLIP consistency between ideal and generated trajectories compared through semantically matched image regions.

C.1. Proof of Vibe Space Analytical Minimum

  • Linear interpolation uniquely minimizes the latent surrogate and approximates the multiscale inverse-diffusion geodesic in the original manifold.The derivation reduces inverse diffusion to matching multiscale flag-space kernels, then substitutes Vibe Space's learned alignment constraint.Source: §C.1 · p. 19

C.2. Graph Laplacian and Diffusion Map

  • Nyström approximation reduces eigen-decomposition cost from O(N^3) to O(S^3) and kernel construction cost from O(N^2) to O(NS), using S=500 in practice.The method samples anchor tokens, computes anchor and cross-affinities, and approximates the full kernel.Source: §C.2 · p. 20

D.1. Correspondence Matching

  • NCut segments each image into semantic regions, then Hungarian assignment matches segment centroids by feature distance to establish corresponding structures.The resulting mapping supports computing the vibe displacement between corresponding regions.Source: §D.1 · p. 20, Figure 16 · p. 20

D.2. Vibe Analogy

  • Vibe Analogy transfers a reference pair’s displacement field to a new subject through semantic correspondence, then decodes and renders the interpolated path with IP-Adapter [45].Each new-subject segment receives the displacement of its matched reference segment.Source: §D.2 · p. 21

D.3. Negative Vibe Control

  • Negative vibe control orthogonalizes positive attributes against a negative subspace; β=1 fully orthogonalizes, while β>1 can push the vibe away from negative attributes.The filtered target kernel uses both positive and negative eigenvectors, although the encoder is trained only on positive exemplars.Source: §D.3 · pp. 21–22

D.4. Vibe Visualization by Gradient

  • Gradient visualization backpropagates through the eigendecomposition to rank DINO feature channels for a semantic cluster and display the top channels as heatmaps.The cluster indicator is obtained from the eigenvectors via k-way Ncut.Source: §D.4 · pp. 22–23, Figure 19 · p. 23

D.5. Vibe Blending: Selecting the Optimal Blend Weight α via Dip in CLIP Consistency

  • The optimal blend weight is the point of minimum CLIP consistency between the decoded ideal path and the re-encoded generated path.A midpoint can miss the best hybrid because correspondence-based interpolation is source-anchored and not perfectly linear in α; the example selects α*=0.6.Source: §D.5 · pp. 23–24, Figure 20 · p. 23
  • For each candidate α, the ideal path decodes the Vibe Space interpolation into CLIP space, while the realized path re-encodes an IP-Adapter-generated image with CLIP.Source: §D.5 · p. 24
  • Consistency is the average cosine similarity between ideal-path and realized-path CLIP features for semantically matched source-image regions.DINO features and spectral clustering segment the source and generated images; the Hungarian algorithm matches cluster centers before region-level scoring.Source: §D.5 · p. 24
  • The optimal α is the one minimizing consistency; this dip marks greatest semantic tension and frequently the most creative, compelling synthesis.The method illustrates selecting α* across a blend trajectory in Fig. 20 and Fig. 21.Source: §D.5 · p. 24

D.6. Vibe Blending: Training with Many Images

Vibe Space can use extra related images during training to identify attributes, while the blending process still uses two images. The framework also supports path extrapolation and blending more than two images.

D.6. Vibe Blending: Training with Many Images

  • Training with extra related images captures the shared “glass window” attribute, producing a glass-texture pyramid in the middle blend.The blending process itself still uses two images.Source: §D.6 · pp. 24–25

D.7. Extrapolating Vibe Blending Paths

  • Setting α > 1 continues the geodesic path beyond the input images and can amplify visual characteristics beyond their presence in the target.The framework’s locally linear manifold allows extrapolation outside the input images’ convex hull.Source: §D.7 · p. 25

D.8. N-Image Blending

  • N-image blending anchors correspondences to one base image and combines weighted semantic displacements from the other images.When one target’s weight is 1 and all others are 0, the result approximates a full morph from the base to that target.Source: §D.8 · p. 25

E.1. Training Procedure

The paper specifies a lightweight training procedure, balances geometric and reconstruction losses, and evaluates blends with LLMs, human studies, and output diversity. Its method produces the most varied blends and the best creative blends according to humans and the LLM.

E.1. Training Procedure

  • Training the 0.72M-parameter encoder–decoder for 1,000 steps takes 15 seconds, and full blending runs in under 30 seconds on an RTX4090 GPU.Graph diffusion-map construction and correspondence matching take milliseconds; image generation takes 2 seconds per image.Source: §E.1 · p. 25

E.2. Loss Balancing

  • The flag-decoder and sample losses are down-weighted by 0.01 because they can overwhelm CLIP reconstruction; the sample loss begins after 500 steps to avoid warm-up instability.The reported weights are λflag dec = 0.01 and λsample = 0.01.Source: §E.2 · p. 26

F.1. LLMs for Generating and Evaluating Blends

  • Gemini is prompted to identify objects and shared attributes before generating a blend, while GPT-5 judges outputs by attribute blending and selects the best.GPT’s generation steps are combined in a single prompt because the OpenAI API lacks a public model for multi-turn image-and-text input and output.Source: §F.1 · pp. 26–27

F.2. User Study Details

  • Ten participants compared 44 image pairs on Creative Potential and Blend Difficulty, while a separate study ranked outputs for coherent and creative blending.The first study used 110 comparisons, with each image pair included in five; the output rankings used first-place votes, then second-place votes.Source: §F.2 · pp. 28–29

F.3. How Does Creativity Relate to Diversity?

  • Our method produces the most varied blends and the best creative blends according to humans and the LLM.Output diversity is measured as mean pairwise DreamSim distance across three outputs; Table 7 also reports CLIP-based diversity.Source: §F.3 · p. 30, Table 7 · p. 30

F.4. Curated Datasets

The evaluation uses curated sets of semantically distinct image pairs and architectural design pairs.

F.4. Curated Datasets

  • The Totally Looks Like subset contains 44 high-quality, semantically distinct image pairs after similarity-based filtering and human review.Pairs with CLIP similarity above 0.65 and DreamSim distance below 0.65 were automatically excluded; a human annotator then removed pairs with text or low-quality images.Source: §F.4 · p. 31
  • The architecture set contains 300 pairs drawn across styles and architects, with a human annotator verifying pair quality.Each pair contains images from different architectural styles and different architects.Source: §F.4 · p. 31

G. Failure Cases

The failure cases expose limits in negative-vibe separation, attribute extrapolation, region correspondence, and decoding out-of-distribution inputs.

Entangled Vibes

  • When desired and undesired attributes are entangled, negative-vibe orthogonalization cannot filter out the entangled attribute.The example entangles style change with color change.Source: G. Failure Cases · p. 31, Figure 31 · p. 32

Extrapolation Uncertainty

  • Extrapolation does not always meaningfully exaggerate relevant attributes; beyond α > 1, it does not further rotate the input object.This failure is illustrated for Vibe Blending extrapolation.Source: G. Failure Cases · p. 31, Figure 32 · p. 32

Correspondence Failure

  • Unsupervised correspondence can vary across random seeds; incorrect region matches produce incoherent blends with mismatched object parts.Good correspondence yields cleaner, semantically meaningful blends.Source: G. Failure Cases · p. 32, Figure 33 · p. 33

Reconstruction Failure

  • Vibe Space depends on IP-Adapter, which reconstructs in-distribution inputs reliably but fails on out-of-distribution inputs.The failures produce unstable and inconsistent reconstructions.Source: G. Failure Cases · pp. 32–33, Figure 34 · p. 33
Loading 2512.14884v1…