Epipolar Geometry Improves Video Generation Models

arXiv:2510.21615 · cs.CV · Submitted 2025-10-24 · Read on arXiv

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: "Epipolar Geometry Improves Video Generation Models".

Jane: Here's the gist of the paper. Video generation models still struggle with geometric inconsistencies, unstable motion, and visual artifacts that break the illusion of realistic 3D scenes The paper explores how epipolar geometry constraints improve modern video diffusion models by aligning them using preference-based optimization.

Tom: Pruning redundant reasoning tokens The core finding is that classical geometric constraints provide more stable optimization signals than modern learned metrics The authors propose a simple approach that bridges modern video diffusion models with classical computer vision algorithms by using well-established non-differentiable geometric constraints as reward signals in a preference-based finetuning framework This method leverages epipolar geometry to assess 3D consistency between frames, allowing the model to rank generations based on adherence to projective geometry principles.

Jane: The approach bypasses the difficulties of directly using non-differentiable computer vision algorithms in the training loop by implementing Direct Preference Optimization (DPO) which requires only relative rankings.

Tom: How the training signal is built The objective function is formulated as Flow-DPO, which requires only pairwise preferences from a dataset where one sample has better consistency than another The loss is defined as L = -E h log σ - βt2 V wθ − Vw ref − Vlθ + V l ref i, where x∗ t = (1−t)x∗ 0 + tϵ∗, v∗ = ϵ∗ − x∗ 0, βt = β(1−t2).

Title and authors: Tom: So we're talking about this paper, "Epipolar Geometry Improves Video Generation Models." Basically, these current video models are still messy; they have geometric inconsistencies and unstable motion that really break the illusion of a real three dee scene <ref:2510.21615#pg1>.

Jane: Right. And this paper is proposing that using epipolar geometry constraints can help fix those issues by aligning the diffusion models using preference-based optimization. It's trying to bridge the gap between modern AI video generation and classical computer vision methods.

Lu: What they found is that classical geometric constraints actually provide more stable optimization signals than the learned metrics these models use right now, which is really interesting for how we train these things.

Meng: Stability is key for deployment. If the motion is unstable, it won't work in any real application. So what exactly are they using to enforce this geometry?

Tom: They’re using epipolar geometry to assess three dee consistency between frames by ranking generations based on how well they follow projective geometry principles <ref:2510.21615#pg1>. It lets the model learn from objective, mathematically grounded signals instead of just subjective human scores.

Jane: That sounds like a big step because those human judgments can be really noisy and expensive to collect, so getting a signal that is mathematically grounded in physics seems smart.

Lu: The methodology involves this Flow-DPO objective function, which requires only pairwise preferences from a dataset where one sample has better consistency than another. The loss function is defined as L equals negative E h log sigma minus beta t squared V w theta minus Vw ref minus Vl theta plus V l ref i.

Meng: That’s a lot of math to track. So, how does this translate into actual training for a model that needs to run fast at scale? What's the cost here?

Lalam: The paper suggests a temporal variation penalty, Ltemp equals negative lambda E Vart(xzero), which is used to prevent the model from degenerating by just reducing motion to achieve three dee consistency <ref:2510.21615#pg1>.

Tom: That penalty keeps things dynamic while still aiming for that geometric consistency, and they also use a low-rank adapter, or LoRA, for finetuning instead of retraining the whole thing.

Jane: And they also mention using implicit KL-divergence regularization within the Flow-DPO objective to help with generalization to different dynamic content later on.

Lu: The evaluation shows that this epipolar alignment actually improves three dee consistency, motion stability, and generalization to dynamic scenes across the board <ref:2510.21615#pg1>. They found a seventy-one point eight percent consistency rate compared to the baseline's fifty-four point one percent when evaluated on human-rated geometric consistency.

Meng: Seventy-one point eight percent is a big jump from fifty-four point one, but what about actual motion quality when we look at it in terms of preference?

Tom: They achieved a sixty nine point five percent win rate in human preference comparisons against the baseline's eighteen point five percent for motion quality. Plus, they reduced epipolar error by thirty one percent, going from zero point one nine zero to zero point one three one.

Jane: That reduction in error is significant because it confirms that using classical geometric alignment principles actually works well for this kind of video generation problem.

Lu: The paper also addresses the challenge of motion-consistency tradeoff by introducing that temporal variation penalty, which stops the model from just making everything look static when it tries to enforce geometry.

Meng: But they also have limitations, right? What's the catch for deploying this on actual dynamic scenes where objects move independently?

Tom: The authors admit the approach primarily focuses on static scenes with dynamic camera movements, which means its effectiveness might be limited when you have independent object movement or non-rigid deformations.

Jane: And they also noted that there’s a potential for false positives or false negatives when the scenes have repetitive textures or significant motion blur, so it’s not a perfect solution everywhere.

Lu: But they still see the bigger picture, suggesting that geometry-aware video generation could actually facilitate other three dee vision tasks like scene reconstruction and SLAM by producing more realistic synthetic data <ref:2510.21615#pg1>.

Meng: If we can use this to generate better training data for reconstruction or SLAM, that’s valuable for our production pipeline. It moves us away from just generating pretty videos toward generating usable geometric information.

Tom: Exactly. So, summarizing the "Epipolar Geometry Improves Video Generation Models" paper, they align diffusion models using preference-based optimization by using epipolar geometry to create more stable training signals than learned metrics alone provide.

Jane: And the results show concrete improvements: seventy-one point eight percent three dee consistency and a sixty nine point five percent win rate in motion quality compared to the baseline.

Lu: The methodology relies on Flow-DPO training with a static penalty, plus temporal variation penalties and LoRA for finetuning dynamic content generalization.

Meng: The real question is scalability. If we use this framework, what’s the overhead compared to just running a standard diffusion model? Is the training time prohibitive?

Lalam: The authors mention that they released a dataset of one hundred sixty-two thousand generated videos with geometric consistency annotations, which is available for further research in geometry-aware video generation.

Tom: So, while it's not perfect for every dynamic scenario, this work provides a clear direction by grounding the optimization process in classical computer vision principles.

Jane: It’s a promising direction because it shows how we can integrate those mathematically rigorous constraints with modern generative models to get better physical understanding.

Lu: This paper is important because it proves that objective geometric metrics can serve as reliable signals for identifying generations that adhere to projective geometry, which guides the model toward better consistency.

Meng: For us at the startup, this means we have a new way to rank and fine-tune our models if we want to push the quality of our output beyond what simple visual fidelity achieves.

Tom: So, that's the gist of "Epipolar Geometry Improves Video Generation Models," showing that using epipolar constraints in a preference-based framework yields measurable gains in three dee consistency and motion stability <ref:2510.21615#pg1,Epipolar Geometry Improves Video Generation Models>.

Jane: It’s a solid piece of work because it moves beyond just subjective quality scores and anchors the generation process to actual geometric rules.

Lu: This paper is important for anyone working on geometric understanding in generative models, because it demonstrates that classical computer vision algorithms can be integrated into the training loop effectively.

Meng: It’s a practical method for improving output quality without needing massive amounts of human labeling for every single frame comparison.

Tom: We'll keep an eye on this, because if we can refine these geometric constraints further, it could really help make synthetic data more useful for things like scene reconstruction.

Jane: It certainly points toward a future where AI systems have a stronger physical world understanding when it comes to generating realistic three dee scenes <ref:2510.21615#pg1>.

The paper's summary: Tom: So, to recap, this paper is about taking these messy video generation models and fixing their geometric inconsistencies by using epipolar geometry constraints in a preference-based optimization setup. It’s basically using classical computer vision rules to guide the modern AI training process.

Jane: Right, so instead of just letting the model guess what looks consistent based on pixels, they are giving it a mathematical rule about how frames should relate to each other in three dee space. It’s like giving the model a strict set of physics constraints to follow.

Lu: What they found is that these classical geometric constraints give way more stable optimization signals than the metrics these models have learned on their own, which is what really matters for training. They are using something called Direct Preference Optimization, or DPO, which just needs us to rank one video as better than another in terms of three dee consistency.

Meng: So they’re bypassing the whole headache of trying to plug a non-differentiable computer vision algorithm directly into the training loop by just using those relative rankings for preference optimization. That sounds like a clever workaround for a problem we always have with incorporating real-world constraints.

Tom: Exactly, it makes things much more tractable because we only need those pairwise preferences, not some complex differentiable loss function that’s hard to implement reliably. The objective function they use is called Flow-DPO, and it’s designed to make the predicted velocity field align with videos that actually look consistent in three dee space.

Jane: And they didn't just stop there; they added a temporal variation penalty, which is important because if you only focus on geometry, the model can get lazy and just reduce all motion to satisfy the rules. This penalty keeps things moving and dynamic while still aiming for that three dee consistency.

Lu: They also made it smarter about different kinds of scenes by finetuning only a low-rank adapter, which is a LoRA setup, and they used implicit KL-divergence regularization within the Flow-DPO objective to help with generalizing to new dynamic content later on.

Meng: So they’re tackling both the immediate stability issue and the long-term generalization problem at once, which is a tough balancing act when you’re shipping something to production. That adaptive timestamp weighting thing sounds like it helps manage that complexity.

Tom: The results are pretty telling for us; they got seventy one point eight percent three dee consistency compared to fifty four point one percent for the baseline, when judged by humans on geometric consistency. That is a significant jump in fidelity.

Jane: And for motion quality, they hit sixty nine point five percent in human preference comparisons against eighteen point five percent for the baseline. That shows a real improvement in how smooth and realistic the camera movement looks to people watching it.

Lu: Plus, they reduced the epipolar error by thirty one percent; that goes from zero point one nine zero down to zero point one three one, which confirms that using these classical geometric alignment principles actually works well for this type of video generation problem.

Meng: So what does this mean for deployment? It means we get output that’s not only visually pleasing but also has a better underlying three dee structure, which could be useful if we ever want to use these generated videos for things like scene reconstruction later on.

Tom: The authors are optimistic, suggesting this geometry-aware video generation could facilitate other three dee vision tasks, like scene reconstruction and SLAM, by producing more realistic synthetic data. It’s a promising direction because it integrates classical geometry principles with modern generative models to get better physical world understanding.

Jane: It really does sound like a step toward AI systems having a stronger physical world understanding when it comes to generating three dee scenes. But there are still limitations, though.

Lu: Yeah, the paper itself is honest about it; this approach primarily focuses on static scenes with dynamic camera movements, so its effectiveness might be limited when you have independent object movement or non-rigid deformations in the scene.

Meng: And they flagged that there’s a potential for false positives or false negatives when the scenes have repetitive textures or significant motion blur, which means it’s not a perfect solution everywhere. It stops working where the input data is too messy.

Tom: But overall, this work proves that objective geometric metrics can serve as reliable signals for identifying generations that adhere to projective geometry, which guides the model toward better consistency. It’s a solid piece of work because it moves beyond just subjective quality scores and anchors the generation process to actual geometric rules.

Jane: It’s important because it shows how we can integrate those mathematically rigorous constraints with modern generative models to get better physical understanding, which is exactly what we need for truly realistic three dee scenes. So, next time you see a video AI generating something weirdly warped or jittery, remember this paper is trying to fix that using geometry.

The paper's improvements: Tom: So, we’re looking at what they actually improved here; they found that using epipolar alignment significantly boosts three dee consistency, motion stability, and generalization to dynamic scenes compared to their baseline models. It’s not just a small tweak; it's a real uplift in how realistic the video looks.

Jane: Right, so imagine the difference between two videos generated by these models. One might look okay at first glance, but the other will have much smoother camera paths and fewer strange artifacts that break the illusion of reality. It’s about making the movement feel more physically correct.

Lu: The numbers back it up; they saw a seventy one point eight percent consistency rate compared to fifty four point one percent on human-rated geometric consistency metrics, which is a big jump. And in motion quality, the aligned model achieved a sixty nine point five percent win rate against eighteen point five percent for the baseline.

Meng: Sixty nine point five percent in human preference comparisons is substantial; that means people are actually finding the generated motion much better than before. That’s something I care about because stability is what we need if we want to deploy these models in anything real-time.

Tom: And they quantified the error reduction too; they reduced epipolar error by thirty one percent, dropping it from zero point one nine zero down to zero point one three one, which shows the classical geometric alignment is working exactly as intended. It’s a measurable improvement on a fundamental geometric property of three dee space.

Jane: What this means for us on the show is that we’re moving past just looking at pretty pictures to actually training models that respect the underlying physics of how things are positioned in three dee space. It changes how we think about what makes a video *real*.

Lu: The methodology also gave them tools for better long-term performance. They introduced that temporal variation penalty to stop the model from just making everything look static when it tries to enforce geometry, and they used the low-rank adapter, or LoRA, along with implicit KL-divergence regularization for generalizing to different dynamic content.

Meng: That combination of penalties sounds like a smart way to keep things stable without sacrificing the ability to make interesting shots. I wonder if that adaptive timestamp weighting helps manage the cost at scale?

Tom: It does, in a sense; it allows the model to handle varying scene complexity better while still sticking to those geometric rules. The authors suggest this geometry-aware generation can actually help with other three dee vision tasks like scene reconstruction and SLAM by creating more realistic synthetic data.

Jane: That’s huge; if we can generate synthetic data that has strong three dee structure, it becomes much more useful for training other AI systems that need to understand physical space. It moves the whole process toward better physical world understanding.

Lu: The authors are really pointing toward a future where we integrate these classical geometry principles directly into the generative pipeline, making the models inherently better at representing three dee relationships from the start.

Meng: For us, that means if we can use this to generate training data for SLAM or reconstruction, our synthetic datasets become much higher quality without needing as much expensive real-world collection. It's a win for production efficiency and data quality simultaneously.

Tom: So, in short, the improvements show that grounding the optimization process in classical computer vision algorithms provides objective signals that lead to more consistent three dee scenes with demonstrably better motion stability. It’s a solid step forward for making AI output geometrically sound.

Conclusion: Tom: So, to wrap up, we’re talking about the paper "Epipolar Geometry Improves Video Generation Models." The main point is that using classical computer vision rules based on epipolar geometry in a preference-based optimization framework makes video generation models much more consistent and stable.

Jane: Right, so it’s about moving away from just getting pretty pictures to training AI that respects the actual three dee structure of space, which is a big deal for realism.

Lu: The implication is that we can anchor these modern generative models to mathematical rules rather than just relying on learned metrics that might be unstable. It suggests a path where classical computer vision algorithms can be integrated into the training loop effectively.

Meng: For deployment, it means we get outputs with better motion quality and stability, which is something I need when shipping models to production so they don't break in unexpected ways.

Lalam: I’m confident in this direction because the paper shows that objective geometric metrics can provide reliable signals for identifying generations that adhere to projective geometry. That’s a very solid foundation for training any generative model.

Tom: It’s a practical method because it lets us use objective signals instead of just subjective human scores, which saves us time and makes the training process more predictable. The results show tangible improvements in consistency and motion quality metrics compared to the baseline.

Jane: It really shows how much physical world understanding we can bake into these AI systems, even when they are generating complex scenes with dynamic camera movements. It’s a step toward more reliable synthetic data for things like scene reconstruction.

Lu: The authors are optimistic that this geometry-aware video generation could facilitate other three dee vision tasks, suggesting it’s a promising direction for enhancing AI systems with stronger physical world understanding.

Meng: I’m just waiting to see how much overhead this adds to the training pipeline before we commit to using it on our actual traffic data. We need to make sure this improvement scales efficiently enough.

Tom: Yeah, we need those numbers on scaling, but the mechanical improvements in consistency and error reduction are undeniable when you look at the paper "Epipolar Geometry Improves Video Generation Models."

Jane: It’s a solid piece of work because it moves beyond just subjective quality scores and anchors the generation process to actual geometric rules. It’s a step toward AI systems having a stronger physical world understanding.

Lu: This paper is important for anyone working on geometric understanding in generative models because it demonstrates that classical computer vision algorithms can be integrated into the training loop effectively.

Meng: For me, it’s about getting better quality output without needing massive amounts of human labeling for every single frame comparison, which cuts down on our manual effort.

Lalam: I'm sure this is the right direction because it proves that objective geometric metrics can serve as reliable signals for identifying generations that adhere to projective geometry.

Tom: So, we’re leaving with the understanding that using epipolar constraints in a preference-based framework yields measurable gains in three dee consistency and motion stability.

Jane: It’s a promising direction because it shows how we can integrate those mathematically rigorous constraints with modern generative models to get better physical understanding.

Lu: The next thing we look at is whether this method holds up when we introduce more complex scene dynamics, like independent object movement in addition to dynamic camera motion.

Meng: That’s the next hurdle; if it doesn't generalize well beyond just camera motion, then we might need to revisit those temporal variation penalties they introduced.

University of Oxford · Google Research

cs.CV

Submitted: 2025-10-24

Updated: 2026-11-10

Project page: https://epipolar-dpo.github.io/Abstract

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: Video generation models still struggle with geometric inconsistencies, unstable motion, and visual artifacts that break the illusion of realistic 3D scenes The paper explores how epipolar geometry…

Key concepts

Epipolar Geometry
This is a set of mathematical constraints from computer vision that describes the geometric relationship between two images taken from different viewpoints. It ensures that points in one frame correspond to the same physical location in another frame, which is crucial for maintaining 3D consistency in generated videos.
Flow-DPO
This is a specific training objective function used to fine-tune video diffusion models. It relies on pairwise preferences—ranking one generated video as better than another—to guide the model's predicted motion field toward outputs that exhibit superior 3D consistency while maintaining high motion quality.
Direct Preference Optimization (DPO)
DPO is a training technique that uses relative rankings between samples instead of needing complex, differentiable loss functions. In this paper, it allows the model to learn from human preferences by simply comparing two video generations and optimizing the model to favor the more geometrically consistent one.
Temporal Variation Penalty
This is a specific loss term introduced during training designed to stop the model from sacrificing motion quality just to satisfy geometric consistency. It penalizes excessive changes over time, ensuring that the generated motion remains smooth and realistic rather than becoming degenerate.

Terminology

Summary

Video generation models still struggle with geometric inconsistencies, unstable motion, and visual artifacts that break the illusion of realistic 3D scenes The paper explores how epipolar geometry constraints improve modern video diffusion models by aligning them using preference-based optimization. p. 0

Where the paper says it

io/ Abstract Video generation models have advanced significantly through the latent diffusion transformers trained with rectified flow techniques. Yet these models still struggle with geometric inconsistencies, unstable motion, and visual artifacts that break the illusion of realistic 3D scenes.

Pruning redundant reasoning tokens

The core finding is that classical geometric constraints provide more stable optimization signals than modern learned metrics The authors propose a simple approach that bridges modern video diffusion models with classical computer vision algorithms by using well-established non-differentiable geometric constraints as reward signals in a preference-based finetuning framework This method leverages epipolar geometry to assess 3D consistency between frames, allowing the model to rank generations based on adherence to projective geometry principles. The approach bypasses the difficulties of directly using non-differentiable computer vision algorithms in the training loop by implementing Direct Preference Optimization (DPO) which requires only relative rankings. p. 1

Where the paper says it

, 2023), requiring only relative rankings rather than absolute reward values. This bypasses the difficulties of directly using non-differentiable computer vision algorithms in the training loop.

How the training signal is built

The objective function is formulated as Flow-DPO, which requires only pairwise preferences from a dataset where one sample has better consistency than another The loss is defined as L = -E h log σ - βt2 V wθ − Vw ref − Vlθ + V l ref i, where x∗ t = (1−t)x∗ 0 + tϵ∗, v∗ = ϵ∗ − x∗ 0, βt = β(1−t2). This loss guides the predicted velocity field vθ to align with videos exhibiting better 3D consistency while preserving motion quality. p. 5

Where the paper says it

Our final objective is Ltotal = L + Ltemp. Minimizing this loss encourages the model to improve denoising performance on preferred samples x w t relative to less preferred samples x l t , guiding the predicted velocity field vθ to align with videos exhibiting better 3D consistency while preserving motion quality.

Where the gains come from

The gains stem from several key design choices that address challenges like motion-consistency tradeoff and generalization to dynamic scenes The authors introduce a temporal variation penalty, Ltemp = -λ E Vart(ˆx0), to prevent the degenerate solution of reducing motion to achieve 3D consistency. Furthermore, generalization to dynamic content is achieved by finetuning only a low-rank adapter (LoRA) and using implicit KL-divergence regularization in the Flow-DPO objective. p. 5

Where the paper says it

(2022) preserves the base model’s capacity for diverse dynamic content. Second, implicit KL-divergence regularization in the Flow-DPO objective keeps the finetuned model close to the reference distribution, preventing overfitting to static scene characteristics.

Evaluation and results

Evaluation across multiple dimensions demonstrates that epipolar alignment improves 3D consistency, motion stability, and generalization to dynamic scenes The method achieves a 71.8% consistency rate compared to the baseline's 54.1% when evaluated on human-rated geometric consistency. p. 7

Where the paper says it

Human Evaluation: While numerical metrics capture geometric aspects, 3D inconsistencies often manifest as artifacts, jitter, or unnatural changes that humans excel at detecting (Figure 4a). Annotators evaluated videos for scene consistency, realism, and artifact-free content.

In terms of motion, the aligned model achieves a 69. p. 5
Where the paper says it

Motion-Consistency Tradeoff: Static bias is one of the key challenges in optimizing model consistency, significantly increasing the complexity of the task. In theory, a perfectly static video (single image repeated) achieves perfect consistency yet fails as a generative video task.

5% win rate in human preference comparisons against the baseline's 18.5%. The study also shows that epipolar error is reduced by 31%, from 0. p. 7
Where the paper says it

Geometric Consistency Metrics: We directly measure adherence to projective geometry principles using classical computer vision algorithms. The Sampson epipolar error shows a dramatic 31% reduction from 0.

190 to 0.131, verifying the effectiveness of classical geometric alignment. p. 7
Where the paper says it

The Sampson epipolar error shows a dramatic 31% reduction from 0. 131, verifying that our alignment successfully optimizes the metric used for preference selection and confirming that classical epipolar geometry provides clean optimization signals.

Limitations and impact

The approach primarily focuses on static scenes with dynamic camera movements, which limits its effectiveness for scenes with independent object movement or non-rigid deformations Limitations include the potential for false positives or false negatives when scenes have repetitive textures or significant motion blur. However, the impact statement suggests that geometry-aware video generation can facilitate various 3D vision tasks, including scene reconstruction and SLAM, by producing more realistic synthetic data. p. 0

Where the paper says it

Yet these models still struggle with geometric inconsistencies, unstable motion, and visual artifacts that break the illusion of realistic 3D scenes. 3D-consistent video generation could significantly impact numerous downstream applications in generation and reconstruction tasks.

The integration of classical geometry principles with modern generative models represents a promising direction for enhancing AI systems with stronger physical world understanding. p. 3
Where the paper says it

, 2025a) improves the physical stability of video generators using multi-component reward functions. Our method leverages classical computer vision algorithms to provide objective, mathematically grounded preference signals based on epipolar geometry, resulting in more reliable alignment with 3D consistency principles.

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

  1. Improved 3D Scene Reconstruction: The model can now generate videos with enhanced 3D consistency, leading to improved 3D scene reconstruction fidelity, as demonstrated by a PSNR increase from 22.32 to 23.13 on the DL3DV test set.

  2. Enhanced Motion Stability and Quality: Generated videos exhibit smoother camera trajectories and reduced artifacts, resulting in improved motion quality scores like VBench Motion Quality (69.5% win rate) compared to the baseline's 50%.

  3. Better Generalization to Dynamic Scenes: The system can generalize to dynamic content even when trained only on static scenes with dynamic cameras by using a combination of implicit KL-divergence, compact trainable adapter, and adaptive timestamp-weighting mechanisms.

  4. More Reliable Alignment Signals: By leveraging epipolar geometry constraints via preference-based optimization, the model is aligned using classical computer vision algorithms to provide objective, mathematically grounded preference signals based on epipolar geometry.

  5. Superior Preference Ranking: The alignment method uses "epipolar geometry metrics provide reliable signals for identifying which generations better adhere to the principles of projective geometry, enabling us to rank videos and to create training pairs that guide the model toward improved geometric consistency."

Abstract

Video generation models have advanced significantly through the latent diffusion transformers trained with rectified flow techniques. Yet these models still struggle with geometric inconsistencies, unstable motion, and visual artifacts that break the illusion of realistic 3D scenes. 3D-consistent video generation could significantly impact numerous downstream applications in generation and reconstruction tasks. We explore how epipolar geometry constraints improve modern video diffusion models. Despite using massive training data, these models fail to capture fundamental geometric principles. We align diffusion models using pairwise epipolar geometry constraints via preference-based optimization, directly addressing unstable trajectories and geometric artifacts through mathematically principled geometric enforcement. Our approach efficiently enforces geometric principles without requiring end-to-end differentiability. Evaluation demonstrates that classical geometric constraints provide more reliable preference signals than the video-quality and geometry metrics based on foundational models. We construct preferences from static scenes with camera motion, where the epipolar metric is most reliable, while the aligned model generalizes to dynamic scenes. By bridging data-driven learning with classical computer vision, we reduce epipolar error by 31% and improve human-rated consistency from 54% to 72% without compromising visual quality.

Sources

Related papers