PaperlayerView exact revision on arXiv

Source-linked AI summary

The Attention Triangle in Audio-Video Models

Sagi Polaczek, Noa Kraicer, Gal Metzer, Zhuo Ning, Ali Mahdavi-Amiri, Daniel Cohen-Or, Raja Giryes

arXiv:2609.03586v1cs.AI

TL;DR

Audio-video diffusion models can leak semantics because cross-modal attention routes information beyond intended bindings, particularly along a bias-shaped audio-video pathway. The paper analyzes the attention triangle, induces controlled leakage, and uses attention-derived signals for inference-time steering; experiments show improved grounding while preserving generation quality. The authors also delimit the method’s reliance on learned representations, initial-generation anchors, and reliable visual segmentation.

  • Problem

    Cross-modal attention in audio-video diffusion can produce systematic semantic leakage and incorrect source attribution as text, audio, and video interact through coupled pathways.

  • Method

    The paper models three cross-attention edges as an attention triangle, extracts attention-derived attribution signals, induces controlled leakage, and jointly steers the edges during inference.

  • Results

    Joint steering of all triangle edges resolves both appearance leakage and incorrect source attribution, while experiments across diverse prompts support probing and mitigation.

  • Takeaways & Limitations

    Leakage reflects structured, bias-driven routing along specific cross-modal pathways, so analyzing and steering those pathways can improve sound-to-source grounding.

  • Takeaways & Limitations

    The method cannot reliably create bindings outside the model’s learned distribution and depends on initial-generation anchors and accurate visual segmentation.

Abstract

from arXiv · show

Audio-video diffusion models rely on cross-modal attention to coordinate text, sound, and visual content, yet this same mechanism can introduce subtle and systematic semantic leakage. We study these models by probing and analyzing the ``attention triangle,'' comprising the three cross-attention edges connecting the text, audio, and video streams, and examine how semantic information is routed across modalities during generation. Our analysis reveals that routing along the audio-video edge is bidirectional: audio can influence video generation, while video can influence audio generation. This edge is shaped by biases encoded in the model's parameters and emerges as a major contributor to leakage: when prompts are in tension with learned priors, cross-modal interactions may override the intended conditioning and reroute semantics toward visually canonical but incorrect outcomes. These effects suggest that semantic artifacts arise not merely from attention spreading beyond its intended target, but from structured, bias-driven interactions along specific pathways. Building on this perspective, we extract attention-derived signals that expose how semantics are distributed and grounded across modalities, and use them as a diagnostic tool to both analyze and deliberately incur leakage under controlled conditions. This enables us to probe the internal dynamics of cross-modal routing and isolate the role of individual interactions. We further leverage these signals to guide inference-time interventions that encourage more consistent cross-modal alignment. Extensive experiments support our analysis and demonstrate improved semantic grounding while preserving generation quality.

1 Introduction

Cross-modal attention in audio-video diffusion can spread semantics beyond intended targets, especially through bidirectional audio-video interactions shaped by learned biases. The paper analyzes these pathways, induces controlled leakage, and uses attention-derived signals for inference-time grounding interventions.

  • Soft, global attention mixing can produce attribute leakage, semantic entanglement, and unstable bindings by spreading associations beyond intended targets.
  • In the attention triangle, text, audio, and video jointly constrain one another, allowing competing interpretations to propagate and reduce global consistency.The paper identifies the audio-video pathway as a major contributor to semantic leakage and hypothesizes that temporal-only audio position leaves source localization underconstrained.
  • Semantic routing is bidirectional: audio can influence visual interpretation, while visual priors can bias sound generation and source attribution.
  • When prompts conflict with learned cross-modal biases, the audio-video edge may override text-induced bindings and reroute sound toward visually canonical but incorrect sources.This frames leakage as bias-driven interaction along specific pathways, not only association spread to unintended entities.
  • Attention-derived signals expose source attribution, support controlled leakage induction and interaction-level analysis, and guide inference-time interventions toward more consistent grounding.
  • Joint steering of all triangle edges resolves both appearance leakage and incorrect source attribution, whereas steering either text edges or the audio-video edge corrects only one.The example illustrates that no single interaction suffices to control cross-modal routing.
  • Across diverse prompts and settings, the framework supports both probing and mitigation of consistent leakage patterns.

2 Background and Related Work

Prior work establishes cross-attention as a steerable control surface but also identifies leakage from soft, global mixing. Audio-video systems extend this problem across modalities, where coupled streams and mismatched spatial-temporal representations create additional source-attribution challenges.

  • Self-attention aggregates context within one modality, whereas cross-attention couples token sequences so latent tokens selectively use conditioning information.
  • Existing work treats cross-attention as a semantically meaningful, inference-time control surface for layout, attribute binding, and leakage correction without retraining.
  • Soft, global attention mixing allows features to bleed between distinct entities, producing failures such as swapped attributes and contextual contradiction.
  • The attention-triangle perspective extends leakage analysis beyond visual entities to cross-modal routing among text, audio, and video.
  • Audio-video diffusion evolved from silent-video systems toward joint audio-video generators using paired subnetworks, shared semantic latents, or dual transformer branches.
  • Recent systems structure cross-modal interaction through modality-specific alignment, learned routing, bidirectional attention, spatial-audio estimation, or reference-conditioned control.
  • Figure 2 represents the triangle with edge-specific biases that amplify intended pairings and attenuate pairings involving competing sources.The figure defines intended pairings across source, sound, and text-related signals, alongside agreement and text-edge masks.
  • Audio-video attention inherits image-domain leakage while adding a dimensional mismatch between 1D temporal audio embeddings and 3D spatiotemporal video embeddings.

3 Identifying the Triangle Problem

Joint text-to-video and text-to-audio generation forms an attention triangle in which semantic information can travel directly and indirectly across text, audio, and video. Evidence from visualizations, attention rollouts, and audio ablation identifies the audio-video pathway as a major source of leakage, routing speech toward visually canonical but incorrect entities.

  • 3.1 The Attention Triangle: Speech assigned to a parrot can instead be routed to a pirate because learned audio-video biases favor human-shaped visual patches.The audio-video edge can override the binding induced by text and drag sound toward a statistically familiar source.
  • 3.1 The Attention Triangle: The attention triangle couples text tokens, audio tokens, and video patches through pairwise cross-attention, allowing attributes to reach video directly or through audio.These paths also create audio-mediated video-to-video interactions through composed attention matrices.
  • 3.2 Audio Leakage Analysis: The first-order probe sums attention from video queries to speech-active audio keys, revealing higher scores on the wrong subject than on the prompt-designated speaker.In the example, pirate queries attend more strongly to speech-active audio keys than parrot queries.
  • 3.2 Audio Leakage Analysis: Composed audio-video attention exposes spatial transport in which mass migrates from intended source patches, such as the parrot, to the visually canonical pirate.This video-to-video coupling is absent from any single attention layer and makes audio-mediated misbinding spatially visible.
  • 3.2 Audio Leakage Analysis: Attention-triangle steering redirects high active-audio scores from the pirate toward the prompt-designated parrot in the same-prompt, same-seed comparison.The post-steering visualization concentrates scores sharply on the parrot rather than the visually canonical human source.
  • 3.2 Audio Leakage Analysis: Audio ablation removes leakage cases found in joint T2AV generation while preserving the shared visual backbone, implicating the audio-video pathway as a major contributor.A text-edge-only intervention leaves source misattribution unresolved, providing complementary evidence for this conclusion.

4 Steering: Grounding Audio and Video Regions

The method builds fixed anchors from text annotations, baseline visual masks, and attention-derived audio masks, then applies agreement- and mask-based biases across all three attention edges during inference. Joint steering is designed to suppress mismatched source-sound interactions and reduce residual leakage left by partial interventions.

  • Steering framework: The training-free steering algorithm jointly modifies text-audio, text-video, and audio-video attention using anchors extracted from an unsteered baseline pass.The anchors remain fixed across denoising steps, and the method operates at inference time without auxiliary loss or fine-tuning.
  • Text-conditioned anchors: User annotations identify the intended source, sound or action, and optional competing source, whose token indices define intended and conflicting text-conditioned cells.These annotations are cleaned before encoding and mapped to token-index sets used to construct the masks.
  • Visual and audio anchors: Visual anchors come from SAM3 masks on decoded baseline frames, while audio anchors come from aggregated audio-query/text-key attention and thresholding.The visual source mask is hard, whereas the audio mask is soft before thresholding at θ_A=0.3.
  • Audio-video edge: The audio-video agreement matrix assigns high values to matched source-sound and non-source-non-sound pairs and low values to mismatched pairs.The resulting disagreement penalty is added in both directions: video-query/audio-key and audio-query/video-key.
  • Text-conditioned edges: The text-conditioned edges use intended, conflicting, and neutral masks, with fixed β=0.5 and γ=2.0 placing greater weight on suppressing conflicting cells.When no competing source is supplied, the conflicting mask vanishes and the bias becomes an intended-cell boost.
  • Joint intervention: Ablations show that steering all three edges jointly is necessary because partial interventions leave residual leakage.The method incurs roughly 2.5× the cost of unsteered generation because it requires an additional full denoising trajectory plus visual-anchor processing.

5 Experiments

The experiments analyze leakage through attention visualizations and evaluate the proposed steering across diverse audio-video scenarios using automatic metrics, ablations, and human preferences. The results show that Ours-Full redirects source attribution, addresses multiple leakage modes, and preserves visual and audio-text fidelity.

  • Attention-leakage analysis: First-order attention scores concentrate on the visually canonical but incorrect pirate in the baseline and on the prompt-designated parrot after steering.The visualization derives active-audio scores from the video-query/audio-key matrix P_VA.
  • Attention-leakage analysis: Second-order V→A→V rollout reveals audio-mediated misrouting by moving seeded attention mass from the parrot to the pirate before steering and back after steering.The intermediate audio-token distribution is diffuse, exposing an indirect pathway not represented by either individual cross-attention matrix alone.
  • Comparisons and ablations: The evaluation spans realistic humans, animals, animated characters, and multi-entity scenes, identifying source-attribution and appearance leakage among recurring failure modes.The comparison includes native LTX-2, adapted Bounded Attention, and partial variants that isolate text-edge and audio-video-edge contributions.
  • Quantitative evaluation: Ours-Full obtains the highest Qwen source-attribution score, 0.1349 versus 0.1216 for native LTX-2, and is preferred over every alternative by VA-Judger.VA-Judger mean preference scores range from 57.6% to 73.3%, with all pair-bootstrap 95% confidence intervals above 50%.
  • Quantitative evaluation: Ours-Full achieves best VBench subject consistency 0.990, background consistency 0.986, and aesthetic quality 0.604 while remaining competitive on CLAP, 0.383 versus 0.384.These results indicate no large degradation in the measured fidelity metrics.
  • User study: Human annotators prefer Ours-Full in 79.2% of attribution, 82.9% of leakage, and 80.1% of overall-quality comparisons.All three aggregate preferences are statistically significant after Bonferroni correction, with p<10^-28.
  • Comparisons and ablations: Ablations show that Ours-Text corrects appearance but not attribution, Ours-AV corrects attribution but reintroduces appearance leakage, and only Ours-Full addresses both.The partial-edge variants therefore isolate complementary contributions of the attention-triangle edge families.

6 Discussion

The analysis identifies cross-modal coupling, especially the audio-video pathway, as a major contributor to bias-driven leakage and shows that attention steering can expose and mitigate these failures across diverse scenarios.

  • 6 Discussion: The Attention Lesson states that stronger cross-modal coupling can amplify bias-driven routing rather than resolve it.The authors frame this as a broader consequence of learned priors overriding prompt-specified bindings.
  • 6 Discussion: The evidence establishes the audio-video pathway’s contribution, but not a unique underlying mechanism.The authors hypothesize that spatial video tokens and temporally organized audio tokens may leave grounding underconstrained.
  • 6 Discussion: The study evaluates steering across realistic humans, animals, animated characters, stylized cartoons, and multi-entity scenes with varied layouts and interactions.Prompts intentionally create tension between requested sounds and canonical visual identities.
  • 6 Discussion: Recurring leakage includes wrong source attribution, appearance changes, canonical voice characteristics, and generation suppression.These failures affect both semantic binding and the resulting visual or audio realization.

B Additional Qualitative Comparisons

Additional qualitative comparisons show that leakage recurs when requested sounds conflict with canonical visual priors, while jointly manipulating attention edges preserves both source attribution and visual identity.

  • B Additional Qualitative Comparisons: Baselines and partial-edge variants either assign the growl to the furry monster or give the teddy monster-like visual features.Ours-Full preserves the teddy’s identity while binding the growl to it.
  • B Additional Qualitative Comparisons: The furry-monster example asks a teddy bear to produce a deep growl despite the monster’s canonical association with that sound.The setup creates tension between the requested source and the competing visual prior.
  • B Additional Qualitative Comparisons: Ours-Text cannot overcome the audio-video prior, whereas Ours-AV transfers the growl to the teddy but introduces monster-like appearance leakage.Only Ours-Full jointly manipulates the text and audio-video edges to achieve both correct binding and preserved identity.
  • B Additional Qualitative Comparisons: The horse-neigh and furry-monster examples show recurring failure modes across different visual domains and sound categories.They also indicate that mitigating these failures requires joint manipulation of all attention-triangle edges.

C Quantitative Evaluation Protocols

The quantitative protocols combine controlled prompt-seed evaluation, counterbalanced model judging, a targeted source-attribution score, and a broader pairwise audiovisual preference check.

  • C Quantitative Evaluation Protocols: The main comparison evaluates Native LTX-2, video-only and external baselines, Bounded Attention, and three attention-triangle intervention variants.Reported targets include source attribution, CLAP audio-text alignment, and VBench subject consistency.
  • C Quantitative Evaluation Protocols: The source-attribution evaluation uses 100 prompts, four matched seeds, 400 videos per audio-bearing method, and 4,800 Qwen evaluations.The video-only diagnostic is excluded because it contains no audio.
  • C Quantitative Evaluation Protocols: Counterbalancing swaps candidate order and maps responses to intended, competing, both, absent, and unclear canonical labels before averaging.This removes fixed Candidate-A or option-order preference from the reported score.
  • C Quantitative Evaluation Protocols: The attribution score S_attr equals intended-only probability divided by visible-source probability, while competing-only and both-source assignments count as leakage.Event-absent and source-unclear mass remain separate diagnostics.
  • C Quantitative Evaluation Protocols: The VA-Judger compares Ours-Full with references using the original prompt and synchronized clips scored across five dimensions.Comparisons use deterministic decoding, reversed presentation order, and 400 pairs with 800 decisions per reference.
  • C Quantitative Evaluation Protocols: Across 2,000 pairs and 4,000 decisions, Ours-Full is preferred to every alternative, with all confidence intervals above 0.5 and positive mean dimension differences.The audit retained equal-total answers as ties and found no inference errors or duplicate keys.

D User Study

The user study compares Ours-Full with four alternatives on source attribution, visual leakage, and overall quality using randomized paired videos and explicit annotations of intended sources and silent entities.

  • D User Study: Human preferences compare Ours-Full with Native LTX-2, Bounded Attention, Ours-AV, and Ours-Text on attribution, visual leakage, and overall generation quality.Each criterion is assessed from paired videos generated with the same prompt.
  • D User Study: Each trial randomizes which side contains Ours-Full, while participants receive the intended source, target sound, and silent entity annotations.The active pool contains 37 unique prompt-seed examples and 118 pairwise comparisons.
  • D User Study: Annotators answer which video better attributes the sound, prevents incorrect visual behavior, and performs better overall.Same responses are treated as ties and excluded from decided-vote win rates and significance tests.
  • D User Study: The study collects 30 complete submissions covering 480 pairwise trial records, with percentages computed over all votes including Same responses.All submissions contain the full 16-trial set and parse successfully.
  • D User Study: The leakage challenge set is curated through iterative human review that selects interpretable binding failures and mutates successful prompt families.Mutations vary entities, sounds, dialogue, semantic associations, visual cues, and wording.
  • D User Study: Static frames understate synchronized audio-video failures, especially voice-characteristic leakage and generation suppression.The authors therefore encourage readers to inspect the accompanying videos directly.

F Limitations

The training-free approach inherits limitations from the base generator and its grounding signals, including representational, anchoring, segmentation, cost, and model-family constraints.

  • The method cannot reliably create audio-video bindings far outside the base generator’s learned distribution.Attention steering can redirect associations the model already represents but cannot reliably invent distant bindings.
  • Initial unsteered-generation anchors may provide too little signal when audio is completely ungrounded or the intended source is visually static.
  • Visual-anchor interventions are sensitive to SAM3 segmentation errors for small, occluded, or ambiguous entities.
  • The extra anchor-extraction pass increases inference cost, and experiments center on LTX-2-style audio-video generation.Evaluating the intervention across additional model families remains future work.

G Analysis Details

The analysis visualizes how attention-derived distributions transport semantic mass across an audio-mediated round trip and how steering changes that routing toward intended sources.

  • Leakage visualization: The propagated visual distribution identifies which video patches receive attention mass after one audio-mediated round trip.On leakage prompts, mass concentrates on competing-source patches, exposing audio as a hidden routing mechanism.
  • Routing comparisons: Ours-Full concentrates returned visual mass on the intended sound source while suppressing the competing source in two LTX-2 examples.
  • Experimental setup: The analysis uses LTX-2’s text-to-audio-video variant to study the audio stream and full attention triangle against a shared visual backbone.The text-to-video variant lacks an audio branch, while the joint variant adds audio tokens and the triangle.
  • Anchor design: Hard SAM3 masks define intended and competing visual sources, while the audio side uses a model-derived soft mask because no analogous external segmenter is available.
  • Audio localization: The audio soft mask is thresholded at θ_A=0.3 and sufficiently localizes speaking regions to drive text-to-audio bias without over-constraining generation.
  • Quantitative routing effects: For the golden retriever example, Ours-Full raises intended-source mass from 46.1% to 90.3% and lowers competing-source mass from 39.5% to 6.6%.

H.2 Hyperparameter Analysis

The method fixes steering strengths across experiments, uses distinct controls for text-edge gaps, and incurs substantially higher per-clip runtime than the baseline.

  • Audio-video edge: The audio-video penalty is fixed at λ=10 across experiments and is not tuned per prompt.The penalty form interprets λ as the maximum logit penalty applied to a conflicting cell.
  • Text edges: Text-edge controls remain distinct: β governs the intended-versus-neutral gap, while γ governs the neutral-versus-conflicting gap.The asymmetric defaults are β=0.5 and γ=2.0, fixed across LTX-2 experiments.
  • Ovi comparisons: Figure 12 compares baseline and Ours-Full outputs for two source-attribution prompts using matched visual frames across timestamps.Sound-source attribution must be assessed from the corresponding videos rather than still frames alone.
  • Audio mask: The audio-mask threshold is fixed at θ_A=0.3 across all LTX-2 experiments and is not tuned per prompt.
  • Runtime: Ours-Full takes approximately 13 to 14 min/clip versus approximately 5.5 min/clip for baseline LTX-2 at the stated configuration.Both use approximately 29 GB peak memory.
Loading 2609.03586v1…