cs.CVMar 6, 2026
SaveNEGATE: Constrained Semantic Guidance for Linguistic Negation in Text-to-Video Diffusion
Organizations: University of Maryland at College Park, United States
Abstract
Negation is a fundamental linguistic operator, yet it remains inadequately modeled in diffusion-based generative systems. In this work, we present a formal treatment of linguistic negation in diffusion-based generative models by modeling it as a structured feasibility constraint on semantic guidance within diffusion dynamics. Rather than introducing heuristics or retraining model parameters, we reinterpret classifier-free guidance as defining a semantic update direction and enforce negation by projecting the update onto a convex constraint set derived from linguistic structure. This novel formulation provides a unified framework for handling diverse negation phenomena, including object absence, graded non-inversion semantics, multi-negation composition, and scope-sensitive disambiguation. Our approach is training-free, compatible with pretrained diffusion backbones, and naturally extends from image generation to temporally evolving video trajectories. In addition, we introduce a structured negation-centric benchmark suite that isolates distinct linguistic failure modes in generative systems, to further research in this area. Experiments demonstrate that our method achieves robust negation compliance while preserving visual fidelity and structural coherence, establishing the first unified formulation of linguistic negation in diffusion-based generative models beyond representation-level evaluation.
Figures & tables
Figure 1: Negation beyond object removal. Structural Functional Negation (SFN), Double Negation Sensitivity (DNS), and Scoped Negation Disambiguation (SND). Baseline diffusion models collapse negation into semantic inversion or scope errors, whereas our model enforces negation as a constraint on diffusion dynamics, yielding correct logical interpretation and stable scene composition. Qualitative results and comparisons with more video diffusion models are provided in Section 4.3 .
Figure 2: Overview of our negation-aware framework via convex feasibility projection. Given a natural-language prompt containing linguistic negation (e.g., “a person holding a phone but not using it”), we first perform semantic decomposition into affirmed concepts , negated concepts , and scope structure . In classifier-free guidance (CFG), the reference guidance increment is defined as , which attracts the diffusion trajectory toward affirmed semantics but does not constrain negated variables. We construct a negation direction , representing the semantic increment that increases alignment with the negated concept. At each reverse-time step ( ), we enforce a half-space constraint in guidance space and project the reference increment onto the feasible region, producing the corrected update . A temporal scheduling strategy progressively tightens the constraint threshold , allowing early structural formation while enforcing strict negation at later stages. The resulting denoising trajectory yields negation-compliant video generation without retraining the text-to-video diffusion model.
Figure 3: Representative qualitative comparison (SFN). We compare our model with state-of-the-art diffusion baselines (Mochi, HunyuanVideo, CogVideoX) on Structural Functional Negation (SFN): “A person holding a phone but not using it.” While baselines often collapse the negation into unintended interaction (e.g., phone-using gestures) or fail to maintain the intended constraint, our model preserves the object presence and suppresses the prohibited behavior, demonstrating negation control beyond object removal.
| Method | CLIPScore | CLIP-neg | BLIP | DINO-conf | NCS | NVR |
|---|---|---|---|---|---|---|
| Mochi | 0.2857 | 0.2411 | 0.8183 | 0.0175 | 3.5789 | 0.3625 |
| HunyuanVideo | 0.2743 | 0.2620 | 0.7776 | 0.0184 | 3.4584 | 0.3675 |
| CogVideoX | 0.2836 | 0.2442 | 0.7869 | 0.0181 | 3.4080 | 0.3800 |
| Ours (Full Model) | 0.2924 | 0.2336 | 0.8201 | 0.0167 | 4.0700 | 0.2300 |
| Ours w/o Repulsive Energy | 0.2924 | 0.2411 | 0.8128 | 0.0181 | 3.9548 | 0.2625 |
| Ours w/o Constraint Scheduling | 0.2800 | 0.2405 | 0.7869 | 0.0174 | 3.5789 | 0.3625 |
Table 1: Quantitative evaluation of negation-aware video generation. CLIPScore measures overall alignment between generated frames and the full textual prompt. CLIP-neg reports similarity between generated frames and the negated concept (lower is better). BLIP evaluates caption-level semantic alignment without modifying the prompt. DINO-conf denotes the average detection confidence of negated objects (lower indicates stronger negation compliance). Refer Appendix D for detailed analysis.
Figure 4: Ablation study on negation control. We compare the our full model with two ablations— w/o Repulsive Energy and w/o Constraint Scheduling —and a strong diffusion baseline (Mochi) on a double-negation prompt (DNS): “A stage that is not unlit.” The full model produces stable, correctly lit stage scenes. Removing repulsive energy weakens negation enforcement and introduces temporal instability (e.g., lighting flicker), while removing constraint scheduling disrupts early-stage structure formation, causing scale drift and unnatural light-source placement.
Figure 5: User study results for negation-aware video generation. 50 participants evaluated four anonymized methods across four criteria: Negation Satisfaction , Constraint Meaning Accuracy , Scene & Action Alignment , and Artifact Avoidance (left). Our method consistently achieved the highest average ratings across all dimensions (4.54–4.84), substantially outperforming Mochi, HunyuanVideo, and CogVideoX. In the overall preference comparison (right), 77.5% of votes favored our method, compared to 13.8% for HunyuanVideo, 8.8% for Mochi, and 0.0% for CogVideoX.
Appendix figures & tables32 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Qualitative comparison across diverse negation scenarios. We evaluate our model against state-of-the-art diffusion models (Mochi, HunyuanVideo, CogVideoX) on eight negation settings: AOC (Absent Object Consistency), LEN (Late Emergence Negation), INA (Implicit Natural-Only Attribute), MNC (Multi-Negation Composition), SFN (Structural Functional Negation), NMI (Non-Inversion Mitigation), DNS (Double Negation Sensitivity), and SND (Scoped Negation Disambiguation). Across simple object removal, functional suppression, logical inversion, and scope resolution, our model consistently preserves scene structure while correctly enforcing negation constraints.
Figure 7: Additional qualitative comparisons across diverse negation scenarios (Set 1). We further evaluate our method against state-of-the-art diffusion-based video generation models (Mochi, HunyuanVideo, CogVideoX) on eight negation categories: AOC (Absent Object Consistency), LEN (Late Emergence Negation), INA (Implicit Natural-Only Attribute), MNC (Multi-Negation Composition), SFN (Structural Functional Negation), NMI (Non-Inversion Mitigation), DNS (Double Negation Sensitivity), and SND (Scoped Negation Disambiguation). Each row corresponds to a distinct negation prompt requiring suppression of objects, attributes, or functional behaviors while preserving the overall scene structure. Compared to competing methods, our approach consistently maintains semantic coherence while correctly enforcing the specified negation constraints across diverse environments and scene compositions.
Figure 8: Additional qualitative comparisons across diverse negation scenarios (Set 2). This figure presents a second set of prompts covering the same eight negation categories as in Figure 6 , including AOC, LEN, INA, MNC, SFN, NMI, DNS, and SND. These examples further highlight the robustness of our method across different scene contexts, ranging from natural landscapes and urban environments to structured indoor scenes and human-object interactions. While baseline models frequently violate negation constraints by introducing forbidden objects or attributes, our method reliably suppresses such elements while preserving scene realism and structural consistency.
| Method | CLIPScore | CLIP-neg | BLIP | DINO-conf | NCS | NVR |
|---|---|---|---|---|---|---|
| Mochi | 0.2857 0.0196 | 0.2411 0.0288 | 0.8183 0.0774 | 0.0175 0.0314 | 3.5789 1.8548 | 0.3625 0.4813 |
| HunyuanVideo | 0.2743 0.0360 | 0.2620 0.0388 | 0.7776 0.1018 | 0.0184 0.0268 | 3.4584 1.8712 | 0.3675 0.4827 |
| CogVideoX | 0.2836 0.0320 | 0.2442 0.0414 | 0.7869 0.0984 | 0.0181 0.0269 | 3.4080 1.8828 | 0.3800 0.4859 |
| Ours (Full Model) | 0.2924 0.0225 | 0.2336 0.0231 | 0.8201 0.0729 | 0.0167 0.0278 | 4.0700 1.6406 | 0.2300 0.4213 |
| Ours w/o Repulsive Energy | 0.2924 0.0221 | 0.2411 0.0281 | 0.8128 0.0756 | 0.0181 0.0301 | 3.9548 1.6724 | 0.2625 0.4405 |
| Ours w/o Constraint Scheduling | 0.2800 0.0305 | 0.2405 0.0477 | 0.7869 0.0881 | 0.0174 0.0245 | 3.5789 1.8439 | 0.3625 0.4813 |
Table 2: Detailed quantitative evaluation with mean and standard deviation. All metrics are reported as mean standard deviation over the full evaluation set. NCS denotes the Negation Compliance Score obtained from a direct vision-language judge, and NVR denotes the Negation Violation Rate computed from the judge’s binary forbidden-presence decision.
| Category | Method | CLIPScore | NCS | NVR |
|---|---|---|---|---|
| AOC | Mochi | 0.2878 | 2.2800 | 0.6800 |
| HunyuanVideo | 0.2735 | 1.7200 | 0.8200 | |
| CogVideoX | 0.2805 | 2.3200 | 0.6800 | |
| Ours | 0.2967 | 2.4600 | 0.6400 | |
| LEN | Mochi | 0.2886 | 3.3200 | 0.5000 |
| HunyuanVideo | 0.2858 | 2.5800 | 0.6200 |
Table 3: Category-wise quantitative evaluation across eight negation types. We report CLIPScore ( ), Negation Compliance Score (NCS, ), and Negation Violation Rate (NVR, ) for each negation category. While simple implicit negation (INA) shows near-saturation performance across models, more complex forms of negation such as MNC, SFN, DNS, and SND reveal substantial performance gaps, where our method consistently achieves higher compliance and lower violation rates.
| Criterion | Ours | w/o Repulsive | w/o Scheduling |
|---|---|---|---|
| Negation Satisfaction | 5.00 0.00 | 2.90 1.94 | 2.90 1.94 |
| Constraint Meaning Accuracy | 5.00 0.00 | 2.70 1.75 | 2.40 1.58 |
| Scene & Action Alignment | 5.00 0.00 | 3.00 2.02 | 2.90 1.94 |
| Artifact Avoidance | 5.00 0.00 | 2.60 1.70 | 1.90 1.59 |
Table 4: Human perceptual ablation study Participants compared the full model with two ablation variants. The full model consistently achieved the highest ratings across all perceptual criteria.
| Criterion | Model | Mean | Std | 95% CI |
|---|---|---|---|---|
| Negation Satisfaction | Ours | 4.76 | 0.88 | [4.68, 4.85] |
| Mochi | 1.96 | 1.56 | [1.81, 2.12] | |
| HunyuanVideo | 1.61 | 1.32 | [1.48, 1.74] | |
| CogVideoX | 1.34 | 0.84 | [1.26, 1.42] | |
| Constraint Meaning Accuracy | Ours | 4.75 | 0.89 | [4.66, 4.84] |
| Mochi | 1.94 | 1.52 | [1.79, 2.09] |
Table 5: Human evaluation results (50 participants, 400 ratings per model per criterion). Our method consistently achieves the highest mean scores across all evaluation criteria with strong inter-participant agreement.
| Model | Step | Memory Consumption | Computation Time |
|---|---|---|---|
| Ours | Inference | 28.2580 GB / 80.0000 GB (26,949 MiB) | 235 sec |
| Ours w/o Replusive Energy | Inference | 25.2843 GB / 80.0000 GB (24,113 MiB) | 117 sec |
| Ours w/o Constraint Scheduling | Inference | 28.2496 GB / 80.0000 GB (26,941 MiB) | 168 sec |
| Mochi | Inference | 26.2406 GB / 80.0000 GB (25,025 MiB) | 114 sec |
| HunyuanVideo | Inference | 41.3589 GB / 80.0000 GB (39,443 MiB) | 396 sec |
| CogVideoX | Inference | 16.7845 GB / 80.0000 GB (16,007 MiB) | 61 sec |
Table 6: Inference-time memory consumption and runtime on NVIDIA H100 (80GB). We report peak GPU memory usage and end-to-end inference time for our full method and its ablations, together with three strong T2V baselines. Our full method ( Ours ) consumes 28.26 GB and requires 235 sec per inference. Removing the repulsive energy term reduces both memory and runtime to 25.28 GB and 117 sec, respectively. Removing constraint scheduling keeps memory nearly unchanged (28.25 GB) while reducing runtime to 168 sec. Among the baselines, HunyuanVideo shows the largest footprint (41.36 GB, 396 sec), whereas CogVideoX is the most lightweight and fastest (16.78 GB, 61 sec). All measurements are conducted under the same hardware setting (single NVIDIA H100 80GB GPU) and identical evaluation protocol.
Figure 9: Comparison of negation compliance under the AOC (Absence of Object) scenario. Prompt: “A highway at sunset, with no vehicles.” Our model removes the forbidden object while preserving global scene structure and lighting consistency, whereas competing diffusion models introduce unintended vehicles.
Figure 10: Comparison of negation compliance under the LEN (Late Emergence Negation) scenario. Prompt: “A concert stage lighting up, with no performers.” Our model enforces entity-level exclusion while maintaining realistic stage illumination, whereas baselines often hallucinate performers.
Figure 11: Comparison of negation compliance under the INA (Implicit Natural-Only Attribute) scenario. Prompt: “A scene featuring only natural elements.” Our model suppresses human presence while preserving ecological coherence, unlike competing methods that introduce artificial objects.
Figure 12: Comparison of negation compliance under the MNC (Multi-Negation Composition) scenario. Prompt: “A classroom scene with no students, no teacher, and no books.” Our model simultaneously enforces multiple exclusion constraints while preserving structural layout, whereas baselines fail to suppress all forbidden entities.
Figure 13: Comparison of negation compliance under the SFN (Structural Functional Negation) scenario. Prompt: “A person holding a phone but not using it.” Our model suppresses the functional interaction while preserving object presence, whereas baselines conflate object existence with usage.
Figure 14: Comparison of negation compliance under the NMI (Non-Inversion Mitigation) scenario. Prompt: “A dog that is not aggressive.” Our model modulates behavioral attributes while maintaining subject identity, whereas competing models fail to disentangle appearance from disposition.
Figure 15: Comparison of negation compliance under the DNS (Double Negation Sensitivity) scenario. Prompt: “A stage that is not unlit.” Our model correctly resolves double negation semantics, producing illuminated scenes, while baselines collapse into dark or ambiguous outputs.
Figure 16: Comparison of negation compliance under the SND (Scoped Negation Disambiguation) scenario. Prompt: “A teacher helping a student who is not paying attention.” Our model correctly localizes the negation scope to the student’s behavior, whereas competing methods misinterpret the scope or alter unrelated elements.
Figure 17: Qualitative comparison under the AOC (Absence of Object) scenario. Prompt: “A city skyline at night, with no lights on.” Our model correctly suppresses artificial light sources across the skyline while preserving realistic nighttime structure. In contrast, Mochi and HunyuanVideo illuminate buildings despite the negation constraint, while CogVideoX produces unstable or partially lit scenes.
Figure 18: Qualitative comparison under the LEN (Logical Exclusion of Entities) scenario. Prompt: “A ski slope after snowfall, with no ski tracks appearing.” Our model maintains a pristine snow surface consistent with the negation constraint. Competing models frequently introduce ski tracks or surface disturbances, indicating failure to enforce logical exclusion during generation.
Figure 19: Qualitative comparison under the INA (Implicit Natural-Only Attribute) scenario. Prompt: “A meadow under the night sky with only grass and stars.” Our model preserves a purely natural scene without introducing artificial structures. Baselines frequently insert buildings or synthetic elements, demonstrating difficulty enforcing implicit natural-only constraints.
Figure 20: Qualitative comparison under the MNC (Multi-Negation Composition) scenario. Prompt: “A mountain cabin with no doors, no windows, and no chimney.” Our model simultaneously enforces multiple negation constraints while preserving global structure. Competing models fail to remove all prohibited components, often leaving doors or windows visible.
Figure 21: Qualitative comparison under the SFN (Structural Functional Negation) scenario. Prompt: “A clock ticking but not moving its hands.” Our model captures the semantic distinction between object activity and functional motion. Notably, HunyuanVideo misinterprets the constraint and shows moving a person’s hands, while other baselines fail to maintain consistent negation behavior.
Figure 22: Qualitative comparison under the NMI (Negation with Moderated Intensity) scenario. Prompt: “A harbor that is not fully occupied.” Our model correctly represents partial occupancy while respecting the negation condition. Baseline methods frequently produce fully crowded scenes, indicating poor control over graded negation semantics.
Figure 23: Qualitative comparison under the DNS (Double Negation Sensitivity) scenario. Prompt: “A beach that is not shell-less.” Our model correctly resolves the double-negation constraint and generates beaches containing shells. Baselines often collapse to trivial empty scenes or fail to capture the intended semantic inversion.
Figure 24: Qualitative comparison under the SND (Scoped Negation Disambiguation) scenario. Prompt: “A boat anchored near an island that is not inhabited.” Our model correctly applies the negation constraint to the island while preserving the boat. Competing models frequently misinterpret the scope, altering the boat or generating populated islands.
Figure 25: Qualitative comparison under the AOC scenario. Prompt: “A snowfield, with no footprints.” Our model generates a clean untouched snowfield consistent with the negation constraint. Baseline methods frequently introduce footprints or surface artifacts.
Figure 26: Qualitative comparison under the LEN scenario. Prompt: “A dock at dusk, with no ropes appearing.” Our model correctly suppresses rope structures while maintaining scene realism. Baselines often introduce ropes or structural elements violating the negation constraint.
Figure 27: Qualitative comparison under the INA scenario. Prompt: “A natural cave interior featuring only stone and shadows.” Our model maintains a purely geological environment, while competing models frequently introduce vegetation or artificial lighting elements inconsistent with the prompt.
Figure 28: Qualitative comparison under the MNC scenario. Prompt: “A snow-covered road with no tire tracks, no signs, and no vehicles.” Our model enforces multiple negation constraints simultaneously, producing a clean road. Baseline models commonly violate at least one constraint.
Figure 29: Qualitative comparison under the SFN scenario. Prompt: “A factory assembly line active but not producing items.” Our model separates mechanical motion from product output, preserving activity without generated items. Baselines frequently produce items or misinterpret the constraint entirely.
Figure 30: Qualitative comparison under the NMI scenario. Prompt: “A mountain peak that is not completely snow-covered.” Our model correctly produces partial snow coverage. Baselines often generate fully snow-covered peaks, ignoring the moderated negation constraint.
Figure 31: Qualitative comparison under the DNS scenario. Prompt: “A rooftop that is not vent-less.” Our model correctly resolves double negation and generates rooftops with vents. Baselines frequently ignore the constraint and produce vent-less rooftops.
Figure 32: Qualitative comparison under the SND scenario. Prompt: “A robot next to a screen that is not displaying content.” Our model correctly applies the negation constraint to the screen while maintaining the robot and surrounding environment. Baseline methods often display content on the screen or misapply the negation to unrelated objects.
More in this topic
Diffusion Model Guidance
Diffusion model guidance steers denoising toward desired conditions, rewards, constraints, or away from unwanted concepts.
See all 352 papers