NEGATE: Mastering Linguistic Negation in Video Diffusion via Convex Projection

TL;DR

Linguistic negation—phrases like "with no cars" or "not using the phone"—has long been the Achilles' heel of diffusion models. NEGATE transforms this linguistic hurdle into a geometric solution. By modeling negation as a convex feasibility constraint in the guidance space, the authors have developed a training-free method to ensure that what you ask not to see, actually stays hidden.

Context: Why "No" is Hard for AI

Vision-language models (VLMs) are excellent at recognizing what is there, but they struggle with what shouldn't be. When you prompt a video model with "a highway with no vehicles," the word "vehicles" often vibrates stronger than the word "no," leading the model to hallucinate cars. This is not just a data problem; it's a fundamental lack of logical operators in the diffusion process. Existing models treat negation as just another keyword rather than a structured constraint.

The Core Insight: Negation as a Geometric Wall

Instead of hoping the model understands the word "no," NEGATE treats the negated concept as a prohibited direction in the model's latent update space.

1. Semantic Decomposition

The framework first parses the prompt into:

• Affirmed concepts ($y^{+}$): The things we want (e.g., "A teacher helping a student").
• Negated spans ($y^{-}$): The things we don't want (e.g., "paying attention").
• Scope ($S$): How these relate.

2. The KKT Projection Mechanism

The standard Classifier-Free Guidance (CFG) provides a "reference increment" (${\delta }_{ref}$) that pushes the image toward the text. NEGATE identifies a "negation direction" ($a_t$) and ensures that the final update (${\delta }_t^{\ast }$) satisfies a half-space constraint: $a_t^{o}p\delta \le b_t$ Using the Karush–Kuhn–Tucker (KKT) conditions, the model computes the minimal modification needed to the update direction to make it "negation-compliant." If the model tries to move toward a forbidden concept, it is projected back to the feasible region.

Figure 1: Overview of the NEGATE framework. The reference guidance is projected onto a feasible half-space, ensuring the denoising trajectory avoids forbidden semantics.

3. Temporal Scheduling

Timing is everything. NEGATE uses a polynomial schedule (${\alpha }_t$) to apply constraints. Early in the diffusion process, the constraint is loose to allow the global structure to form. As the process continues, the constraint tightens to strictly remove forbidden details and "clean up" the final frames.

Experimental Battleground: The 8 Categories of "No"

The authors introduced a sophisticated benchmark covering eight distinct linguistically grounded failure modes, including:

• AOC (Absent Object): "A highway with no vehicles."
• LEN (Late Emergence): Preventing objects from popping up halfway through a video.
• SFN (Structural Functional): "Holding a phone but not using it"—this is the "Holy Grail" of control, where the object must stay, but its action must be suppressed.

Figure 2: Qualitative comparison for Structural Functional Negation. Top: Baseline models often fail and show the person using the phone. Bottom: NEGATE keeps the phone but suppresses the "using" action.

Quantitative Dominance

The results confirm that geometric constraints outperform prompt engineering:

• Negation Compliance Score (NCS): Improved to 4.07 (outperforming Mochi's 3.57).
• Negation Violation Rate (NVR): Slashed to 0.23 (vs. 0.38 for CogVideoX).
• User Preference: A staggering 77.5% of users preferred NEGATE-generated videos over those from state-of-the-art backbones.

Critical Perspective

While NEGATE is a powerful training-free plug-in, it does come with a computational cost. The additional "negation branch" and projection math increase inference time by roughly 100%. However, for applications where logical reliability is non-negotiable (e.g., safety-critical simulations or high-end creative control), this trade-off is well worth it.

Conclusion

NEGATE proves that the path to better AI isn't always through "more data," but through "better logic." By bridging formal linguistics with the geometry of diffusion dynamics, this work provides a framework for the next generation of logically grounded, controllable generative models.