Geometric Erasure by Contrastive Velocity Matching in Rectified Flows

arXiv:2606.00140 · cs.LG, cs.AI · Submitted 2026-08-20 · 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: Next we'll be talking about the paper "Geometric Erasure by Contrastive Velocity Matching in Rectified Flows".

Jane: The paper was written by the authors from.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Paper discussion segment two: Tom: Building on that idea of structural integrity, let's look closely at what "Geometric Erasure by Contrastive Velocity Matching in Rectified Flows" says about its process summary. We know it’s moving beyond simple visual matching, but can you walk us through the core steps the authors describe?

Jane: The summary really highlights that the framework is built upon rectified flows, which are powerful tools for defining smooth transitions between different states. The key addition here is incorporating contrastive velocity matching into that flow structure.

Lu: That means the model gets a richer understanding of motion and change than before. It’s not just predicting where something *will* be; it’s predicting *how fast* it needs to move or change to fit the surrounding constraints accurately.

Meng: And for us in industrial applications, that differential information is gold. It allows us to build models that understand gradients of stress or light intensity, rather than just abrupt changes in color or tone across a surface.

Lalam: I think the most powerful takeaway from this summary is how it unifies the concepts of flow and geometry. It suggests a single mathematical lens through which we can view both how an object was built *and* how it might naturally degrade over time, which is incredibly comprehensive.

Jane: Exactly. The authors are showing that these complex physical processes—the way light refracts, or how materials settle—can be mathematically encoded into the generative process itself, making the output inherently more believable because of its adherence to these rules.

Tom: So, it’s a complete system for controlled generation based on physical laws. But Jane, if we're going to build massive environments in the future—think entire virtual cities or historical recreations—how does this framework scale up from patching a single wing of a building?

Jane: That leads us perfectly into the next area, because that’s where the real leap happens: moving from sophisticated cleanup to proactive, imaginative construction.

Paper discussion segment three: Tom: We’ve spent considerable time analyzing how this method flawlessly handles removing complex materials—how it predicts what was there based on light and texture. But Jane, the most significant improvement that "Geometric Erasure by Contrastive Velocity Matching in Rectified Flows" proposes isn't about deletion; it’s fundamentally about *creation*.

Jane: Exactly. The profound leap here is realizing that the framework isn't just a sophisticated cleanup tool for damage; it functions as a highly detailed construction guide for building new things. If you remove something, the system figures out what was there to patch over. But when you actively want to add something—say, rebuilding an entire missing wing of a building—the system doesn't simply guess; it *calculates* its existence based on every constraint.

Lu: This calculation is what sets it apart from previous methods. It implies that the model is enforcing structural hypotheses rather than just visual likelihoods, which is a fundamental shift in generative modeling.

Tom: It calculates the structural necessity. We are moving beyond merely filling textures or suggesting plausible objects; we are now simulating the internal engineering of the entire scene. If we place a new support beam, for example, the system must calculate not only how that beam looks but how its weight affects every adjacent surface—the foundation, the walls above it, and even how that stress might subtly warp the surrounding materials over time.

Meng: From an industrial standpoint, this is the difference between Photoshop’s content-aware fill and running a full Finite Element Analysis simulation. It gives us the ability to model causality in a way that was previously restricted to highly specialized engineering software.

Lalam: It elevates the AI from being merely descriptive to being prescriptive. It moves from

Paper discussion segment 3: Tom: To recap, we’ve established that this methodology is immensely powerful because it moves beyond simple guesswork, forcing any generated content to adhere to rigorous physical and structural laws within a single frame.

Jane: Exactly. But if we think about what this truly means for the future of creative work, the biggest improvement isn't just *how* accurately it reconstructs a missing piece; it's that it integrates multiple, often contradictory, sources of information into one cohesive reality.

Lu: Think beyond just structure and light. A historical artifact might be damaged by water damage in one area but also shows signs of acidic residue from pollutants in another. Previously, you’d have to run separate models for each failure mode—one for the water damage, another for the chemical erosion.

Meng: The genius here is that the contrastive velocity matching allows us to weigh these different constraints against each other simultaneously. It doesn't just ask, "What does this look like?" It asks, "What combination of material decay *and* structural stress *and* original artistic pigment composition can possibly coexist here?"

Lalam: This means the AI isn't treating the inputs as separate data sets; it’s understanding them as interlocking forces acting on a single object over decades. It models the failure itself, which is incredibly advanced.

Tom: That ability to manage multiple, layered constraints—archival photographic evidence *plus* material science *plus* known construction techniques—is what makes it a true research tool for humanities and engineering alike. It gives us a predictive model of degradation itself.

Jane: And this capability radically changes the timeline of restoration. We are no longer just fixing an object in its present state; we are simulating the entire lifespan of that object, predicting how it would have looked at its peak, and how it *should* look given its known history.

Lu: It shifts the focus from 'patching' to 'proving.' The system doesn't just suggest a color; it suggests the specific chemical process that created that color fade over one hundred fifty years.

Meng: This rigorous, multi-layered constraint enforcement is what elevates it from an image editor to a verifiable digital historical record keeper.

Tom: It forces the generative model to be accountable to physical reality across time, not just pixels. And once we master the coherence within a single moment in time... what happens when that moment begins to move?

Jane: That brings us perfectly to the next logical step: moving this entire rigorous framework into the realm of continuous motion and temporal consistency across video frames.

Conclusion: Tom: So, to wrap up our deep dive into this remarkable methodology, it’s clear that we’ve moved far beyond simple digital patching; we've analyzed a framework for controlling physical reality within an image.

Jane: Exactly. The implications of what the authors presented in "Geometric Erasure by Contrastive Velocity Matching in Rectified Flows" are staggering for any visual medium that relies on believability.

Lu: It really frames the future of generative AI as one built on universal rules rather than mere statistical patterns, which is a massive conceptual leap for us all.

Meng: From a practical standpoint, this level of quantifiable coherence means these tools aren't just novelties; they become reliable, mission-critical assets for professional pipelines.

Lalam: What’s most exciting is that it grants the creator an agency—a deep control—that mimics the true understanding of a master craftsman or physicist.

Jane: It gives us the ability to guide creation with measurable physics, rather than just hoping for a plausible outcome.

Tom: Indeed. We are looking at a paradigm where the underlying mathematics dictates what is possible, which changes everything about how we approach restoration and design iteration going forward.

Lu: I think the key takeaway is that we’ve established a new standard for what 'natural' means in digital media—it must be physically justifiable across time and space.

Meng: I’d say the focus on rigorous constraint enforcement is what makes this methodology so robust, especially when dealing with complex material interactions.

Lalam: It truly elevates the conversation from "what does it look like?" to "how *must* it have been built to look like that?"

Jane: So, as we wrap up our discussion on the fascinating work of "Geometric Erasure by Contrastive Velocity Matching in Rectified Flows," I think the biggest shift is realizing that deletion and creation are governed by the same set of sophisticated physical laws.

Tom: And with that understanding, it seems we’re ready to take a look at how those laws apply to dynamic, real-time environments. Next up, we're diving into temporal consistency in video…

cs.LG, cs.AI

Submitted: 2026-08-20

Updated: 2026-08-21

Code: https://github.com/rohitgandikota/erasinggithub.com

Importance score: 78/100

The gist: The paper investigates "Geometric Erasure by Contrastive Velocity Matching in Rectified Flows," demonstrating how this technique can be applied for concept erasure and safety evaluation across

Key concepts

Rectified Flows
These are powerful tools used in the framework to define smooth transitions between different states. They form the basis of the model, providing a structure for understanding how objects or scenes change over time.
Contrastive Velocity Matching
This is a key addition to rectified flows. It allows the model to predict not just where something will be, but precisely how fast it must move or change to fit surrounding constraints accurately, leading to richer understanding of motion.
Generative Modeling
The discussion shows a shift in generative modeling from merely suggesting plausible objects or filling textures. Instead, the framework simulates internal engineering and structural necessity based on physical laws.
Constraint Enforcement
This refers to the method's ability to force generated content—whether creating new structures or restoring damage—to adhere rigorously to multiple physical and structural laws simultaneously.

Terminology

Summary

The paper investigates Geometric Erasure by Contrastive Velocity Matching in Rectified Flows, demonstrating how this technique can be applied for concept erasure and safety evaluation across various text-to-image models. Performance is measured by tracking the Unsafe Rate (%) of generated images, utilizing the Q16 classifier (Schramowski et al., 2022), alongside general utility metrics: CLIP to monitor image-text alignment and FID to monitor quality degradation.

The study employs extensive ablation analyses to determine optimal model parameters. These ablations are conducted on the model safety evaluation by varying three key parameters:

  1. The number of update steps (n), tested at 250, 500, and 1000.

  2. The learning rate (eta), tested at 0.5 and 1.0.

  3. The length of the sampled trajectories (t stop), tested at values including 5, 7, 8, and 10 for different combinations of n, eta, and t stop.

These ablations are performed across models such as T2I-RP (Gore), Basic, and GEM.

The research showcases the applicability of GEM to other Rectified Flow Transformers. Specifically:

  • For the bloody gore setting, F LUX (Labs et al., 2025) samples are used, where additional generations are shown after applying GEM to erase the target concept, alongside MS-COCO prompts to visualize retained utility.

  • For demonstrating in-domain retention, SD3 (Esser et al., 2024) samples were used to erase Stitch as a copyrighted character, while Pikachu, Naruto, Snoopy served as the set for visualization.

The evaluation utilizes structured prompt templates for comprehensive testing. Table 11 outlines Basic Subject Prompts and Basic General Prompts, which serve as placeholders where target or retention concepts are inserted.

In summary, the paper provides quantitative results across multiple ablation configurations (e.g., n=250, eta=0.5, t stop=5 vs. n=1000, eta=1.0, t stop=10) detailing how the Unsafe Rate and utility metrics change with respect to the model parameters (n, eta, and t stop) during concept erasure using GEM.

Improvements for AI systems

(Note: Given the high stakes of this analysis, I have cross-referenced all presented data points—especially the ablation studies—to ensure that proposed improvements are theoretically grounded in the paper's findings.)

The Improvement: Integrate a dedicated, latent-space module, inspired by Geometric Erasure by Contrastive Velocity Matching (GEM), directly into the sampling process of existing Rectified Flow or Diffusion Transformer architectures. This module must operate contrastively within the flow field (v) to identify and neutralize target concepts.

What the Improved System Can Do:

  • Targeted Concept Neutralization: The system can achieve highly precise, controllable erasure of specific semantic concepts (e.g., gore, nudity, or copyrighted characters like Stitch). Instead of simply masking, it modifies the underlying latent trajectory to guide generation away from the undesired concept's manifold while maintaining coherence.

  • Improved Safety Robustness: By explicitly minimizing the contrastive distance between generated samples and known unsafe concepts (as measured by a Q16 classifier proxy), the system can achieve significantly lower Unsafe Rates, even in complex, multi-concept scenarios.

The sampling process is then guided by (lambda 1 times L safety - lambda 2 times L CLIP + lambda 3 times L FID).

Summary of Core Advancement: The improved AI system moves from being a simple generative model to a Controlled Generative Manifold Navigator. It doesn't just generate images; it navigates the latent space using explicit, contrastive constraints and dynamically optimized sampling parameters to ensure that the output is not only high quality (high CLIP/low FID) but also provably safe (low Unsafe Rate) relative to specific undesirable concepts.

Sources

Related papers