Slot-RAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders

arXiv:2607.11196 · cs.CV · Submitted 2026-07-13 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Slot-RAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders".

Tom: Deploying object-centric models for real-world scene understanding typically requires complex pipelines to achieve both robust scene decomposition and high-fidelity generation.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: Well, we've just finished listening to a paper called "SlotRAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders," and I gotta say, it’s something that really makes you think about how much complexity we usually throw into these systems. It seems they’ve tackled the typical problem of needing heavy external tools like VAEs or massive generative priors just to get object understanding and high-quality generation in real scenes.

Jane: That's right, Tom; it sounds like they're trying to simplify the pipeline significantly by operating directly inside the continuous semantic feature space of a vision foundation model, like DINOv3. It’s a really clever way to bypass those bottlenecks we usually have to deal with when we try to connect scene decomposition and generation.

Lu: I think what stands out is their core idea: they replace those external generative priors with a feature-space diffusion process using a Diffusion Transformer, which allows for both unsupervised object discovery and high-fidelity generation at the same time, which feels really ambitious.

Meng: From an engineering standpoint, I’m curious about how they manage that integration. If we're working with frozen foundation models like DINOv3 for the features, how do you ensure that the training process doesn't just get stuck optimizing something superficial?

Lalam: Lalam thinks it’s significant because it moves us away from needing huge text-to-image datasets to train these generative components, which really opens up avenues for more diverse and nuanced cultural representations in AI.

Tom: Exactly, Meng; that move toward training from scratch within the semantic feature space is a big deal for generalization. So, let's talk about what they actually proposed in this paper: "SlotRAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders." It’s essentially introducing a fully integrated framework that uses a frozen DINO encoder to get dense features, then compresses those into discrete object tokens called slots using Slot Attention.

Jane: And then they condition a Diffusion Transformer, or DiT, which is the generative engine, directly on these object slots by concatenating them into the self-attention sequence during the denoising process. It’s like telling the generator exactly what objects are present in the scene without needing any external image generation knowledge.

Title and authors: Lu: That feature space diffusion process is key because it allows them to model the generative aspect probabilistically, while still using deterministic object tokens for structure, which seems to unlock simultaneous unsupervised discovery and high-fidelity synthesis.

Meng: So, they’re not just doing reconstruction; they are modeling the features themselves during denoising based on these semantic inputs? That requires a very specific loss function setup to keep everything aligned.

Lalam: Lalam finds the use of the Representation Alignment head particularly interesting because it forces the predicted features to stay close to the target DINO space, which seems like a strong constraint for maintaining semantic consistency across different generation steps.

Tom: It is, and that leads us into their suggested improvements. The paper points out several ways they streamlined this approach, focusing on making the learning process more robust and efficient than previous object-centric methods. They suggest replacing the standard VAE bottlenecks with this direct feature-space diffusion formulation entirely.

Jane: They also highlighted that by training these core components from scratch, without relying on subsidized text-to-image priors, you get a system that is less dependent on those external biases when it comes to understanding scenes and generating new ones.

Lu: One of the improvements they mention is how they manage the conditional diffusion; instead of just using slots as simple inputs, they use slot-conditioned DiT where the slots are projected into the feature dimension before being concatenated as prefix tokens in the self-attention sequence.

Meng: That’s a specific technical detail that addresses how to best integrate those discrete object tokens into a continuous feature flow for the transformer architecture, which is crucial for practical implementation.

Lalam: For Lalam, the improvement regarding efficiency is what really resonates; if this framework can achieve significant reductions in training time and inference speed compared to traditional latent diffusion baselines, it means these kinds of complex visual intelligence engines could run much more practically on consumer hardware.

Tom: It’s a big deal, Meng; those efficiency gains suggest that high-fidelity scene understanding isn't just for massive data centers anymore; it could actually be deployed where resources are tighter. So, let's look at the final conclusions of "SlotRAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders." They summarize their findings on unsupervised object discovery and image reconstruction quality on the COCO dataset.

Title and authors: Jane: They conclude that SlotRAE outperforms early pixel-reconstruction approaches like Slot Attention and DINOSAUR in terms of object discovery, even achieving performance comparable to stronger deterministic methods like DINOSAUR (Transf.) while surpassing standard diffusion decoders on cluttered scenes.

Lu: Furthermore, they show that the reconstructions produced by SlotRAE are visually closer to the original images than methods like Stable-LSD or GLASS when looking at PSNR and LPIPS scores on COCO.

Meng: The paper does state a limitation, though, which is that while it excels in feature space generation and reconstruction quality, they rely on the quality of the underlying VFM encoder; if you use a weaker backbone instead of DINOv3, the performance drops significantly.

Lalam: That makes sense because it highlights that this approach scales well with strong semantic features, but it also shows that you still need a solid visual foundation model to get the best results from SlotRAE.

Tom: So, to wrap up: "SlotRAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders" introduces a novel object-centric diffusion framework that operates entirely within the semantic feature space of vision foundation models, allowing for simultaneous prior-free object discovery and high-fidelity generation.

Jane: It’s a neat way to combine unsupervised scene decomposition with generative capabilities without needing those heavy external VAE latent spaces or pre-trained image generation priors.

Lu: The implication is that we can get robust, semantically structured understanding and synthesis directly from the features learned by powerful vision models, which opens up a lot of creative possibilities for novel AI applications.

Meng: For practical application, it means we might see real-time systems for complex scene manipulation or analysis running on devices that aren't designed for massive generative models anymore.

Lalam: Lalam feels this paper is important because it shows that cultural understanding and image synthesis can be driven by inherent semantic structures in foundation models rather than just sheer data volume, which could lead to more context-aware AI systems.

Tom: That’s our take on "SlotRAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders." It’s a framework that simplifies the path from raw visual data to object understanding and generation by leveraging the inherent structure of modern vision models. We've got some big ideas coming up next, so stick with us.

The paper's summary: Tom: So, we're diving deeper into "SlotRAE," which essentially summarizes how they built this whole system to get object understanding and image generation without needing those heavy VAE bottlenecks that usually slow things down.

Jane: That’s right, Tom; the core summary is that SlotRAE creates a single framework where you take dense features from a vision foundation model and immediately use them to condition a diffusion process directly in that same feature space, cutting out the need for separate latent representations.

Lu: What I find really compelling in their summary is how they managed to unify unsupervised scene decomposition—finding the objects—with high-fidelity generation all through this single, continuous feature space approach. It’s very elegant from a theoretical standpoint.

Meng: From my side, I'm interested in how they handled the training objective; if you're doing it entirely from scratch without pre-trained priors, the stability of that learning process is going to be a big practical hurdle for deployment.

Lalam: For me, what really stands out in their summary is how this technique could fundamentally change cultural AI development because it suggests we can build systems that understand complex visual scenes purely from inherent semantic structures, not just memorizing huge datasets.

Tom: Exactly, Lalam; and as an engineer, I’m thinking about the practical implications of that unification—if it works as summarized, we could have a system that doesn't just label objects but can manipulate entire scenes coherently.

Jane: And from a concept standpoint, imagine being able to swap out an object in a generated scene and have the lighting and shadows update perfectly because the underlying feature representation is consistent across both tasks. It makes sense how they framed it as streamlining the pipeline.

Lu: The paper emphasizes that this approach doesn't just give you better reconstruction; it fundamentally alters how we think about what a vision model learns, suggesting that its internal semantic structure is inherently suited for both discovery and synthesis simultaneously.

Meng: That brings us to the real-world impact; if the summary holds up, we could see a massive reduction in the computational overhead for high-fidelity image generation because you're not running two separate, heavy models back-to-back.

Lalam: And that efficiency gain is huge because it means these advanced visual intelligence capabilities can actually be deployed on much more accessible hardware than we currently think is possible.

Tom: Right, so we’ve seen the summary—it boils down to a unified, scratch-trained object-centric diffusion framework that skips the traditional VAE steps entirely. What do you guys see as the next major area this technique could impact?

The paper's improvements: Tom: So, we’re looking at what the authors suggest as improvements for SlotRAE, which focuses on making this feature-space approach even more robust and versatile than what they first presented.

Jane: They propose several tweaks that aim to make the system less fragile when it comes to handling different types of scenes or when you want to prioritize either object discovery or visual synthesis in the output.

Lu: I’m particularly interested in how they suggest tuning the REPA objective strength; adjusting that hyperparameter should give users more granular control over whether they want a very strict decomposition into discrete slots or a more holistic, blended feature representation during generation.

Meng: From an engineering standpoint, I see the suggestion to use different conditioning mechanisms, like switching between slot-based cross-attention and the adaLN approach, as something that lets us tailor the system for different computational constraints or specific performance targets.

Lalam: For me, what’s exciting is how these improvements directly address the limitations of relying too heavily on one single type of feature space; it suggests a modular design where we can swap components to optimize for either pure object identification or high-fidelity visual output.

Tom: That modularity is key, Jane; it means the system isn't locked into just one mode and can adapt its behavior based on what the user needs most at that moment.

Jane: And I think the paper’s direction on adaptability really speaks to how flexible these models can become for different visual tasks beyond just standard image generation.

Lu: The authors also suggest ways to scale this framework by using stronger vision backbones, implying that the performance gains aren't just about changing the diffusion process, but fundamentally about feeding it a richer set of semantic information from a better foundation model.

Meng: That means for us building production systems, we should probably focus on selecting the right feature extractor early on to ensure we get those high-quality semantic tokens they’re aiming for.

Lalam: And from a cultural perspective, this adaptability means that AI systems won't be rigid; they can evolve their understanding of visual information as our societal needs change without needing a complete re-architecture.

Tom: So, the authors are pushing for a system that’s not just one thing, but something configurable and scalable across different feature quality levels.

Jane: Indeed, Tom; it shows they aren't just aiming for one perfect answer but creating a flexible toolset that can handle the messy reality of real-world visual data.

Lu: And looking ahead, I think the biggest potential is in combining this with other modalities, like incorporating temporal information from video data into these slots to create truly dynamic scene understanding.

Meng: That would be a significant jump in complexity for the inference pipeline, but if they can manage it smoothly with this structure, it opens up a whole new class of applications.

Lalam: I think the most impactful area is how this could help us build more nuanced digital representations of human culture that are both accurate and capable of creative synthesis based on abstract concepts.

Tom: That’s a big thought, Lalam; so while the paper focuses on COCO results, the direction points toward making these models versatile enough to tackle much more complex visual problems down the road.

Conclusion: Tom: So we’re wrapping up our discussion on "SlotRAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders," which is essentially this new framework that bypasses heavy VAEs to integrate object discovery and generation right inside a vision foundation model's feature space.

Jane: That’s right, Tom; the main point is how they managed to tie scene decomposition and generation together seamlessly using only the semantic features already present in models like DINOv3. It’s a very integrated way of thinking about visual AI.

Lu: What really stands out is how they handled the training objective; it's structured to force the predicted noise to align directly with the target feature space, which is a clever way to constrain the generative process without external supervision.

Meng: I agree with Lu; that alignment mechanism seems crucial for keeping things stable when you’re training from scratch without relying on massive pre-trained generation priors.

Lalam: For me, this work shows us that we can build AI systems that deeply understand visual scenes purely from their inherent structure, which could lead to more context-aware tools in the future.

Tom: And that's huge, Lalam; it means these systems aren't just pattern matchers but true scene interpreters.

Jane: Exactly, and when you look at the results on COCO, they show that this method delivers reconstruction quality that is quite high for an unsupervised approach.

Lu: The authors also highlighted that the system’s performance scales with the underlying foundation model; using a stronger encoder definitely leads to better object discovery metrics.

Meng: That gives us a clear direction for implementation, Tom; we need to think about how we select the right backbone to maximize our potential gains.

Lalam: And I really see this impacting culture because it opens up the possibility of creating visual tools that are inherently more representative of diverse, complex human environments.

Tom: So, in short, "SlotRAE: Streamlining Object-Centric Learning via Direct Representation Auto-Encoders" provides a powerful method for achieving high-fidelity synthesis and object discovery within a single feature space.

Jane: It’s a very clean way to approach the problem of connecting structure and generation without needing those complicated latent space intermediaries.

Lu: The theoretical implications are interesting because it suggests that feature spaces learned by large models inherently possess the organization needed for both tasks simultaneously.

Meng: I just hope that in practice, this efficiency translates into a system that can actually run on consumer hardware reliably, which is the biggest hurdle for me right now.

Lalam: If we get this kind of integrated understanding working well, I think it will help us move toward creating AI that truly understands and synthesizes the world around us in a way that respects its underlying visual logic.

Tom: Well, we’ve covered a lot today on SlotRAE; it’s definitely a paper worth keeping an eye on for how we can simplify complex tasks in vision.

Jane: It was a really insightful look at how feature space diffusion can handle both structure and generation without the usual baggage.

Lu: I think the future work they suggest, especially around multi-modal integration, is where the real creative potential lies for this framework.

Alexandre Chapin, Emmanuel Dellandrea, Liming Chen

Ecole Centrale de Lyon, LIRIS

cs.CV

Submitted: 2026-07-13

Updated: 2026-09-28

Importance score: 83/100

The gist: Deploying object-centric models for real-world scene understanding typically requires complex pipelines to achieve both robust scene decomposition and high-fidelity generation.

Key concepts

Semantic Feature Space: This is the continuous, high-dimensional feature space learned by large vision foundation models (like DINOv3). Instead of working with pixel data, SlotRAE operates in this abstract space where visual concepts and objects are naturally represented. This allows the model to understand scene structure without needing complex intermediate representations.
Slot Attention: This module takes dense semantic features from the frozen backbone and compresses them into discrete 'slots' or tokens representing specific objects in a scene. It refines these slots through iterative cross-attention, focusing on object relationships rather than raw spatial locations, making the scene structure explicit for the next generative step.
Diffusion Transformer (DiT) Decoder: This is the generative engine of SlotRAE. It uses a standard diffusion process to create new images by predicting noise directly within the semantic feature space. The object slots condition this decoder, allowing it to generate high-fidelity images that are strictly guided by the discovered scene structure.
Representation Alignment (REPA) Head: This is a component integrated into the DiT that stabilizes generation. It enforces similarity between the features predicted by the noise prediction and the target DINO feature space. This alignment ensures that generated features adhere closely to what a powerful vision model considers 'real' or semantically correct.

Terminology

Fidelity — 33× in this paper · explained in 2 episodes
Transformer — 31× in this paper · explained in 2 episodes
Foundation Model — 27× in this paper · explained in 3 episodes
Alignment — 23× in this paper · explained in 1 episode
Foundation Models — 15× in this paper · explained in 7 episodes
Diffusion Model — 12× in this paper · explained in 8 episodes
Vision Foundation Model — 11× in this paper
Core — 9× in this paper · explained in 2 episodes
Diffusion Models — 9× in this paper · explained in 6 episodes
Ablation — 7× in this paper · explained in 1 episode
Latent Space — 7× in this paper · explained in 7 episodes
Generative Modeling — 6× in this paper · explained in 3 episodes
Ablation Studies — 4× in this paper · explained in 1 episode
Latent Representation — 4× in this paper · explained in 2 episodes
Benchmark — 3× in this paper · explained in 2 episodes
Consistency — 3× in this paper · explained in 3 episodes
Latent Diffusion Model — 3× in this paper · explained in 2 episodes
Latent Diffusion Models — 3× in this paper · explained in 2 episodes
Latent Representations — 3× in this paper · explained in 3 episodes
Accuracy — 2× in this paper · explained in 1 episode
Adaptability — 2× in this paper · explained in 1 episode
Generative Model — 2× in this paper · explained in 2 episodes
Generative Models — 2× in this paper · explained in 2 episodes
Transformer Architecture — 2× in this paper · explained in 2 episodes
U-net architecture — 2× in this paper · explained in 3 episodes
Adam — 1× in this paper · explained in 2 episodes
Adaptation — 1× in this paper · explained in 1 episode
Aggregation — 1× in this paper · explained in 1 episode
Attention Mechanism — 1× in this paper · explained in 2 episodes
Classifier-Free Guidance — 1× in this paper
Computational overhead — 1× in this paper · explained in 2 episodes
Cross-Attention Mechanism — 1× in this paper · explained in 2 episodes
Decoupling — 1× in this paper · explained in 2 episodes
Diffusion Processes — 1× in this paper · explained in 2 episodes
Embeddings — 1× in this paper · explained in 2 episodes
Feature Extraction — 1× in this paper · explained in 2 episodes
Fine-Tuning — 1× in this paper · explained in 5 episodes
Generalization — 1× in this paper · explained in 13 episodes
Learned Perceptual Image Patch Similarity — 1× in this paper
Loss Function — 1× in this paper · explained in 2 episodes
Mean Best Overlap — 1× in this paper
Modularity — 1× in this paper · explained in 9 episodes
Multi-Layer Perceptron — 1× in this paper
Nash Equilibrium — 1× in this paper · explained in 2 episodes
Peak Signal-to-Noise Ratio — 1× in this paper
Representation Learning — 1× in this paper · explained in 2 episodes
Segmentation Masks — 1× in this paper · explained in 2 episodes
Temporal consistency — 1× in this paper · explained in 2 episodes
Trace — 1× in this paper · explained in 2 episodes
Variational Bayes — 1× in this paper · explained in 2 episodes

Summary

Deploying object-centric models for real-world scene understanding typically requires complex pipelines to achieve both robust scene decomposition and high-fidelity generation. SlotRAE proposes a simpler, fully integrated framework that operates directly within the continuous semantic feature space of visual foundation models (e.g., DINOv3), eliminating the need for heavy, pretrained generative priors and VAE bottlenecks while achieving state-of-the-art results on COCO.

The gist: SlotRAE introduces a novel object-centric generative framework that unifies slot-based scene decomposition with feature-space diffusion, training its core components from scratch within the semantic feature space of a vision foundation model.

Framework Overview and Core Components

SlotRAE is a fully integrated framework that processes inputs via several key stages. The process begins by extracting dense semantic features using a frozen DINOv3 encoder, which are then compressed into discrete object tokens (slots) via Slot Attention. The generative process itself is handled by a Diffusion Transformer (DiT) decoder operating directly within the continuous DINO feature space, where these slots condition the diffusion through concatenation in the self-attention sequence.

The framework incorporates several crucial elements:

A frozen visual backbone defining the semantic feature space (e.g., DINOv3).
A Slot Attention module that maps dense features into discrete object tokens (slots), refined through iterative cross-attention where normalization is applied over the slots rather than spatial features.
A Diffusion Transformer (DiT) decoder that models the generative process using a standard forward diffusion formulation, conditioned entirely on the unsupervised object slots.
A Representation Alignment (REPA) head integrated into an early layer of the DiT to stabilize feature-space generation by enforcing similarity between un-noised predicted features and target DINO space.
An optional, frozen RAE decoder used strictly for pixel-space visualization and latent composition in the final stage.

Training Objective and Methodology

The training objective is formulated to minimize noise prediction directly within the continuous semantic feature space, bypassing traditional VAE bottlenecks. The primary diffusion loss is defined as:

(Equation 4)

The total training objective combines the diffusion loss with a representation alignment term:

(Equation 5)

Key aspects of the training methodology include:

The VFM feature extractor remains strictly frozen, and only the Slot Attention module and the conditional DiT are trained end-to-end from scratch.
The slot-conditioned DiT predicts noise based on the augmented sequence [S; Ft], where slots are projected to the feature dimension before concatenation as prefix tokens.
The REPA objective is incorporated with a balancing hyperparameter λ to stabilize generation and improve convergence, forcing alignment of predicted features with the target DINO space.

Key Contributions and Advantages

SlotRAE makes several significant contributions by operating directly in the semantic feature space:

It introduces the first object-centric diffusion framework that operates entirely within the semantic feature spaces of vision foundation models, eliminating the need for VAE-based latent representations.
It proposes a slot-conditioned Diffusion Transformer with representation alignment that is trained entirely from scratch, providing a rigorous baseline isolated from pre-trained image-generation priors.
It demonstrates that foundation-model feature spaces naturally support both unsupervised scene decomposition and high-fidelity generation, enabling strong performance on real-world scenes and competitive zero-shot objectlevel composition.

The advantages of this approach include:

(First)

It completely eliminates the structural artifacts and compression bottlenecks imposed by VAEs.

(Second)

It aligns the object discovery mechanism with emergent semantic structures rather than raw pixel statistics, extending the benefits of feature-reconstruction methods to the generative domain.

(Third)

Because reconstructed features can be mapped back to pixel space via a frozen, pre-trained RAE decoder, the synthesis quality remains fully observable while strictly preserving object-centric compositionality.

Experimental Validation and Results

Experiments on the COCO dataset demonstrate that SlotRAE achieves state-of-the-art results across four tasks:

Unsupervised Object Discovery: SlotRAE significantly outperforms early pixel-reconstruction approaches like Slot Attention and DINOSAUR, achieving highly comparable performance to stronger deterministic methods like DINOSAUR (Transf.) and surpassing standard diffusion decoders (e.g., SlotDiffusion) on cluttered scenes.
Image Reconstruction Quality: Qualitatively, SlotRAE produces reconstructions that are visually closer to the original images, preserving object boundaries and textures better than Stable-LSD or GLASS, achieving superior PSNR (13.93) and LPIPS (0.46) scores on COCO compared to baselines like GLASS (0.59 LPIPS).

Improvements for AI systems

Based on the Slot-RAE paper, here are specific improvements that can be made to AI systems, categorized by the capabilities they would gain:

) Improvements for AI Systems Using Slot-RAE:

A fully integrated framework operating directly within the continuous semantic feature space of a Vision Foundation Model (VFM), bypassing the need for heavy, pre-trained VAEs or external generative priors (like Stable Diffusion).
Ability to perform simultaneous unsupervised object discovery and high-fidelity image generation/reconstruction without requiring large-scale text-to-image training data or massive generative prior biases.
Robust zero-shot compositionality: The system can seamlessly manipulate, swap, add, and subtract object slots extracted from multiple independent scenes and generate a novel scene that preserves object identities while maintaining a coherent global structure.
Superior scene decomposition fidelity: The AI system can decompose complex, cluttered real-world scenes into discrete semantic entities with sharper boundaries and better preservation of fine-grained textural details compared to existing methods (outperforming baselines like Stable-LSD and GLASS in PSNR/SSIM).
Enhanced efficiency for generative tasks: The system achieves a two-to-fivefold reduction in training time and a significant increase (up to 25x) in inference speed compared to traditional latent diffusion baselines, making high-fidelity generation more practical on resource-constrained hardware (e.g., consumer GPUs).
Adaptability based on feature space quality: The system's performance scales directly with the quality of the underlying VFM encoder; using stronger backbones (like DINOv3-B) leads to better object discovery metrics.
Tunable trade-off between decomposition and synthesis: Users can tune the system's behavior by adjusting hyperparameters (e.g., slot dimension, REPA objective strength, or conditioning mechanism—adaLN vs. cross-attention) to prioritize either strict object factorization (discovery) or holistic generative fidelity (reconstruction).

) What the Improved AI System Can Do:

The improved AI system can function as a highly efficient and self-contained visual intelligence engine capable of:

Meticulously analyzing an input image to automatically identify every distinct object and structural component, providing precise, semantically labeled masks for each entity without external supervision or text prompts.
Synthesizing photorealistic images from scratch based purely on the learned semantic structure of those objects—it can generate novel scenes or modify existing ones by inserting new objects or removing existing ones while ensuring the resulting image is structurally sound and visually coherent.
Performing complex scene manipulation in a zero-shot manner: Given three different photographs of a room, it could take the sofa from Photo A and place it into the kitchen shown in Photo B, producing a new, realistic image where the sofa seamlessly integrates with the kitchen's lighting and geometry, rather than just pasting an object onto a background.
Operating rapidly on edge devices: Due to its streamlined DiT decoder and lack of heavy VAE bottlenecks, this system can perform real-time (sub-second latency) high-fidelity image synthesis or scene analysis directly on mobile or consumer hardware.
Creating highly accurate 3D/Scene understanding representations: By leveraging the rich, dense feature space of VFMs, the system can generate compact, semantically meaningful representations that are superior for downstream tasks like robotic scene navigation and object interaction planning.