DCR: Counterfactual Attractor Guidance for Rare Compositional Generation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "DCR: Counterfactual Attractor Guidance for Rare Compositional Generation".
Paper summary: Tom: To recap, the paper DCR: Counterfactual Attractor Guidance for Rare Compositional Generation identifies default completion bias in diffusion models where they collapse toward high-frequency patterns when asked for rare combinations >
Jane: And their solution is Default Completion Repulsion, or DCR. They claim this training-free framework constructs a counterfactual attractor by relaxing the rare factor to find an alternative trajectory that reflects the model’s actual preference >
Lu: The key mechanism involves defining a counterfactual drift between the target and this attractor, and then applying a projection-based repulsion mechanism to remove guidance components aligned with that drift >
Meng: So they are essentially calculating a difference between what you want and what the model defaults to, and then subtracting that unwanted part from the actual guidance update >
Tom: Exactly. The paper shows they use this technique across eight different compositional categories, including temporal misalignment, attribute rebinding, and environment recomposition >
Jane: And they show that when you look at specific categories like material state conflicts, the violation rate drops from zero point seven two in the Mochi baseline to zero point one eight with DCR >
Lu: The authors built a controlled evaluation suite with four hundred prompts designed to isolate these specific failure modes, and they use a multimodal judge called GPT-4o to assess if both compositional factors are present and coherent > <ref:2605.06512#pg2>
Meng: I wonder about that evaluation setup. If we're deploying this, do we have real traffic data to ensure the bias isn't something else entirely that breaks our assumptions >
Tom: The paper says the formulation is backbone-invariant by construction, meaning it works across different diffusion model architectures without needing any modifications to them >
Jane: That’s important because if it’s backbone-invariant, it means we don't have to rewrite the core parts of our generative models just to use DCR >
Conclusion: Tom: So looking at DCR: Counterfactual Attractor Guidance for Rare Compositional Generation, the main thing is that they identified this default completion bias as a specific generative failure mode >
Jane: And they introduced a training-free method to counteract it by modeling and suppressing this tendency through counterfactual attractor guidance >
Lu: The implication is that we can reliably generate rare but semantically valid compositions without relying on prompt engineering tricks or extensive retraining efforts >
Meng: It means we can get better results without the massive computational cost of training new models, which is something my team would find very appealing for production deployment >
Tom: Exactly. They show that by treating default completion bias as a principled semantic operator rather than just a prompt engineering problem, we can get faithful adherence to complex prompts >
Jane: So it’s about making the generative model more honest about what it understands versus what it just defaults to when things get tricky and rare >
University of Maryland at College Park
cs.CV
Submitted: 2026-05-07
Updated: 2026-11-17
Comments: 53 pages, 34 figures, v2: corrected typos and more results added
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: Diffusion models often fail to produce rare but plausible compositions when prompted with combinations that are valid but underrepresented in training data This failure mode is identified as default…
Key concepts
- Default Completion Bias
- This is the failure mode where diffusion models generate outputs that reflect statistically dominant visual patterns instead of rare but semantically valid compositions. The model's denoising process is implicitly attracted toward these high-frequency configurations, causing it to ignore specific, less common prompt instructions.
- Counterfactual Attractor
- DCR constructs a hypothetical prompt (the attractor) that represents a relaxed version of the original request. This allows the framework to model the model's preferred completion tendency by defining an alternative trajectory that reflects what the model would likely generate if it weren't biased toward common patterns.
- Projection-Based Repulsion
- This is the core mechanism where DCR calculates a 'counterfactual drift'—the difference between the target and attractor trajectories. It then removes any guidance components aligned with this drift direction, effectively pushing the denoising trajectory away from frequent alternatives and toward the desired rare composition.
Terminology
Summary
Diffusion models often fail to produce rare but plausible compositions when prompted with combinations that are valid but underrepresented in training data This failure mode is identified as default completion bias, where denoising trajectories are implicitly attracted toward high-frequency semantic configurations The authors introduce Default Completion Repulsion (DCR), a training-free framework that explicitly models and suppresses this default completion behavior DCR constructs a counterfactual attractor by relaxing the rare compositional factor while preserving surrounding semantics, inducing an alternative denoising trajectory reflecting the model’s preferred completion The framework defines a discrepancy between target and attractor trajectories as a counterfactual drift, and proposes a projection-based repulsion mechanism that removes guidance components aligned with this drift direction
Identifying the Failure Mode
The paper identifies compositional collapse as a key failure mode in diffusion models for rare compositional generation This occurs when prompts combine semantically valid elements that rarely co-occur in training data, causing generated outputs to reflect more common visual patterns The failure arises from the model’s intrinsic preference for statistically dominant patterns during denoising, which causes the trajectory to drift away from the specified rare composition Existing guidance mechanisms fail because they do not explicitly model this competing tendency
How the Training Signal is Built
DCR models default completion bias through a counterfactual attractor constructed by relaxing the original prompt This involves defining an attractor prompt that represents a relaxed version of the original prompt, such as replacing “a snowy beach with waves” with “a tropical beach” A probe branch is constructed using a reduced guidance scale to expose this competing tendency, defined by the equation ϵprobe = ϵuncond + wattr (ϵattr − ϵuncond) The counterfactual attractor drift is defined as at = ϵprobe − ϵtarget, which distinguishes the update from simple guidance-scale rescaling
Where the Gains Come From
The core mechanism of DCR is a projection-based repulsion that removes the component of the CFG update aligned with this attractor drift, yielding the corrected update ∆⋆t = ∆ref − λtat This prevents collapse toward frequent alternatives while preserving other semantic components Quantitative evaluation shows DCR significantly improves compositional fidelity compared to standard classifier-free guidance and other baselines, while maintaining high visual quality The method achieves the highest Compositional Compliance Score (CCS) of 4.1300 and the lowest Compositional Violation Rate (CVR) of 0. p. 63100 across all evaluation metrics
Evaluation and Robustness
The authors constructed a controlled evaluation suite consisting of eight complementary categories (ENV, TEMP, OBJ, ATTR, SCALE, CTX, MAT, DENS), with 400 prompts designed to isolate specific compositional failure modes The evaluation uses a multimodal judge (GPT-4o) to assess whether both compositional factors are simultaneously present and coherently realized in the generated video Ablation studies confirm that removing the attractor prompt leads to the largest performance drop, while removing repulsion degrades CCS to 3.9500 and raises CVR to 0. p. 6 p. 63475 The formulation is backbone-invariant by construction, meaning it works across different diffusion model architectures without modification DCR requires three denoiser forward passes per step, adding approximately 78 sec to the runtime compared to Mochi (standard two-branch CFG) The method is robust to hyperparameter variation across a wide range of values, demonstrating that DCR does not require careful tuning for effective attractor suppression The overall contribution is the identification of default completion bias as a generative failure mode and the introduction of a projection-based repulsion mechanism that suppresses it at inference time
Limitations and Future Work
A limitation is that the attractor prompt requires identifying a semantically meaningful frequent counterpart for each target composition, which may be ambiguous The projection-based repulsion operates at the level of global latent directions and may not fully capture spatially localized or temporally varying compositional constraints Future work may extend this formulation to broader classes of distributional constraints beyond rare co-occurrence, including negation and quantification Integrating learned attractor estimation with neural trajectory control could further reduce dependence on explicit attractor specification The work positions implicit attractor suppression as a central challenge for next-generation controllable diffusion systems The formulation is backbone-invariant in the same sense that classifier-free guidance itself is, being fully specified by three operations that do not reference the architecture The target failure mode is empirically backbone-agnostic, as all tested backbones exhibit consistent collapse toward statistically dominant completions The computational overhead of DCR is modest and represents a practical trade-off for improved compositional fidelity The paper concludes that DCR enables reliable generation of rare but semantically valid compositions by treating default completion bias as a principled semantic operator rather than a prompt engineering problem
Ethics Statement
DCR is designed to improve compositional fidelity in generative video models by suppressing distributional biases that cause models to ignore rare but valid semantic combinations The method is intended for research purposes and for applications where faithful prompt adherence is required, such as creative content generation, simulation, and visual prototyping DCR does not introduce new generative capabilities beyond those of the underlying backbone model; it modifies guidance at inference time to better reflect the intended prompt The DCR-Bench evaluation suite is constructed solely for research evaluation purposes, with prompts designed to be physically plausible and free of harmful content Computational resources are limited to standard research GPU hardware, as DCR operates entirely at inference time without requiring additional training By improving the ability of generative models to faithfully realize rare but valid compositions, DCR may contribute to more reliable and controllable AI-generated content The ethical considerations are addressed by ensuring the benchmark is free of harmful or offensive content and that DCR does not introduce new generative capabilities beyond the backbone model
Detailed Benchmarking Dataset
The authors argue that standard vision-language benchmarks are insufficient because they evaluate representation-level similarity rather than the distributional behavior of generated samples under compositional constraints The benchmark consists of eight categories, each containing 50 prompts, resulting in a total of 400 evaluation samples, prioritizing controlled rarity and failure isolation Each prompt implicitly defines a compositional constraint over generated samples by requiring that the generated sample satisfies the compositional constraint specified by prompt p Compositional collapse occurs when gk(x0) ∈ S / k(p), typically drifting toward a high-probability alternative under the model prior The dataset is explicitly designed to induce such conflicts between target composition and prior preference
Detailed Quantitative Evaluation
The evaluation uses five complementary metrics: CLIPScore for global prompt alignment, CLIP-attr to measure attractor suppression, BLIP Caption Alignment for text-level semantic fidelity, CCS (Compositional Compliance Score), and CVR (Compositional Violation Rate) CCS is obtained from GPT-4o by having it directly reason over sampled video frames and assign a 1–5 compliance score based on whether both compositional factors are simultaneously present and coherently realized CVR measures the empirical frequency of attractor collapse, where vi = 1 if collapsedi = True, 0 otherwise The results show that DCR achieves the best performance across all metrics, improving CLIPScore (0.3131) while reducing CLIP-attr (0. p. 62558), and attaining the highest CCS (4. p. 613) and lowest CVR (0. p. 631) The analysis confirms that improvements translate to actual compositional correctness under multimodal reasoning
Category-wise Compositional Analysis
The benefits of DCR are most pronounced in categories requiring suppression of strong distributional priors, such as MAT, where CVR is reduced from 0.72 (Mochi) to 0. p. 118 In ENV, CVR drops from 0. p. 148 to 0.36 with a CCS gain of +0. p. 682 CTX and OBJ show the largest CLIPScore improvements alongside substantial CVR reductions The paper confirms that material-state conflicts are particularly susceptible to attractor collapse, and targeted repulsion is highly effective in these settings
Attractor Trajectory Visualization
Figure 9 compares videos generated by Ours (DCR), Mochi (standard CFG baseline), and pattr only (standard CFG applied to the attractor prompt alone) across three compositional categories In all three cases, the pattr-only output reflects the model’s default completion tendency and closely resembles the Mochi baseline output This validates that pattr serves as an accurate proxy for the implicit attractor, confirming that DCR effectively suppresses this tendency to generate rare but semantically valid compositions The results confirm that DCR enables reliable generation of rare but semantically valid compositions without degrading visual quality or scene realism
Backbone Choice and Method Generality
The target failure mode is empirically backbone-agnostic, as all three backbones exhibit consistent collapse toward statistically dominant completions across all eight categories DCR’s formulation is backbone-invariant by construction, requiring no modification to the transformer, scheduler, or text encoder The only quantity DCR inherits from each backbone is the magnitude scale of ϵtext − ϵuncond, which determines appropriate values for guidance hyperparameters Mochi was selected as the primary testbed because it exposes a clean classifier-free guidance interface allowing for controlled ablation studies The paper suggests that calibration overhead on new backbones is modest rather than load-bearing
Computation Time and Memory Consumption
The full method requires three denoiser forward passes per step, adding approximately 78 sec to the runtime compared to Mochi (standard two-branch CFG) The projection-based repulsion contributes negligible overhead beyond the forward pass, as it only requires a scalar inner-product and norm computation The inference time for the full method on an H100 is 192 sec, which is comparable to other baselines like HunyuanVideo (396 sec) The computational overhead of DCR is modest and represents a practical trade-off for improved compositional fidelity The results confirm that all three components contribute meaningfully to suppressing default completion bias and improving compositional fidelity
Ablation Study
The ablation study confirms that removing the attractor prompt leads to the largest performance drop, indicating that explicit attractor modeling is the most critical component Removing repulsion degrades CCS to 3.9500 and raises CVR to 0.3475, confirming that trajectory-level correction is necessary beyond probe computation alone Removing the schedule further reduces CCS to 3. p. 79675 and increases CVR to 0. p. 6
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Counterfactual Attractor Guidance (DCR) explicitly models and suppresses default completion behavior by constructing a counterfactual attractor by relaxing the rare compositional factor while preserving surrounding semantics, inducing an alternative denoising trajectory that reflects the model’s preferred completion.
-
Projection-based Repulsion removes guidance components aligned with the counterfactual drift direction, preventing the denoising trajectory from collapsing toward frequent alternatives while preserving other semantic components.
-
The improved AI system can generate rare but semantically valid compositions by explicitly identifying and suppressing competing completion tendencies, moving beyond target-only guidance to control generation under distributional constraints.
-
The system can achieve superior compositional fidelity across eight categories (ENV, TEMP, OBJ, ATTR, SCALE, CTX, MAT, DENS) by utilizing a structured evaluation suite that probes multiple dimensions of compositional collapse.
-
The model can perform better in challenging scenarios like material-state conflict (MAT) and environment recomposition (ENV), where baseline models consistently collapse toward frequent semantic alternatives.
-
The system maintains high visual quality while improving reasoning-level metrics, achieving the highest Compositional Compliance Score (CCS) and the lowest Compositional Violation Rate (CVR) across all evaluated categories.
Sources
- Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets
- CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models
- Training-Free Structured Diffusion Guidance for Compositional Text-to-Image Synthesis
- Emu Video: Factorizing Text-to-Video Generation by Explicit Image Conditioning
- Latent Video Diffusion Models for High-Fidelity Long Video Generation
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text
- Imagen Video: High Definition Video Generation with Diffusion Models
- Classifier-Free Diffusion Guidance
- CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers
- NEGATE: Constrained Semantic Guidance for Linguistic Negation in Text-to-Video Diffusion
- HunyuanVideo: A Systematic Framework For Large Video Generative Models
- LLM-grounded Video Diffusion Models
- FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling
- Make-A-Video: Text-to-Video Generation without Text-Video Data
- Learn "No" to Say "Yes" Better: Improving Vision-Language Models via Negations
- Phenaki: Variable Length Video Generation From Open Domain Textual Description
- Gen-L-Video: Multi-Text to Long Video Generation via Temporal Co-Denoising
- ModelScope Text-to-Video Technical Report
- HunyuanVideo 1.5 Technical Report
- CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs