Text-to-video models often generate the visual concepts requested in natural language prompts but fail to satisfy their intended ordering, timing, and persistence requirements. We propose representing these requirements as specifications in a formal Temporal Logic, and introduce Temporal Specification Guidance (TSG), a training-free inference-time framework that turns temporal specifications into executable objectives for guiding video generation during sampling. TSG scores traces of visual concept evidence using differentiable quantitative semantics and uses the score's gradients to construct a specification-guided denoising direction. We instantiate this framework with video-text attention traces, efficient differentiable scores for Metric Temporal Logic, and gradient-derived attention biases. This instantiation requires neither an external concept classifier nor backpropagation through the video transformer. Its attention-steered forward pass yields a denoising direction that augments classifier-free guidance (CFG). On a benchmark of 220 prompts spanning 11 categories of temporal constraints, including ordered object entry and state transitions, this instantiation improves specification satisfaction accuracy over CFG by 9.5-20 percentage points across five models, with minimal memory overhead. These results demonstrate the potential of temporal specifications and differentiable satisfaction objectives as a flexible interface for controlling video generation.
Figures & tables
Figure 1: TSG improves adherence to temporal requirements. Top: CFG produces an amber light for the first two-thirds while TSG follows the expected yellow–red–green progression. Bottom: CFG omits both requested rescue cues; TSG generates a flame signal while retaining the drone. Videos are generated with CogVideoX-5B.
Figure 2: Temporal Specification Guidance (TSG). During an additional conditional pass, TSG scores each head’s attention traces against the specification and applies gradient-derived logit biases. The resulting prediction augments CFG. Transformer weights receive no gradients. The example presents videos generated with Wan2.2 5B.
Figure 3: TSG benchmark with interval-timed temporal requirements from 11 categories. Shaded intervals indicate permitted event timing within the video duration; ∅ denotes absence, and “or” denotes mutually exclusive outcomes. Pictograms are illustrative.
Model
Method
Sat Acc. (%)
Img. quality
Latency (min)
Peak GPU mem.
CogVideoX 2B
Base
10.5% (+0.0%)
59.80 (+0.0)
0.75 ± 0.03 ( 1.00× )
26.30 GiB
TSG
30.5% (+20.0%)
56.85 (-3.0)
2.62 ± 0.38 ( 3.49× )
26.33 GiB
CogVideoX 5B
Base
8.6% (+0.0%)
63.72 (+0.0)
1.82 ± 0.08 ( 1.00× )
31.28 GiB
TSG
26.8% (+18.2%)
62.18 (-1.5)
6.06 ± 0.86 ( 3.33× )
31.31 GiB
Wan 2.1 1.3B
Base
15.9% (+0.0%)
69.25 (+0.0)
1.39 ± 0.02 ( 1.00× )
30.48 GiB
TSG
31.4% (+15.5%)
67.39 (-1.9)
2.35 ± 0.17 ( 1.69× )
30.52 GiB
Table 1: TSG improves satisfaction accuracy on a benchmark of 220 prompts. Latency and Memory measurements are on B200 for all models, except for Wan2.2 A14B which was run on H200.
Method
Temporal Acc ↑
Latency (min/video) ↓
Peak GPU mem (GiB) ↓
One object (N=80)
Base
69.40% (0.0%)
1.29 ± 0.01 (1.00 × )
30.48 (1.00 × )
TempoControl (one round)
72.44% (+3.0%)
2.23 ± 0.01 (1.72 × )
79.41 (2.61 × )
TempoControl (full)
84.05% (+14.6%)
9.87 ± 2.10 (7.63 × )
79.43 (2.61 × )
TSG
87.02% (+17.6%)
2.02 ± 0.01 (1.56 × )
30.52 (1.00 × )
Two object (N=82)
Table 2: TempoControl benchmark results with Wan2.1 1.3B. Parentheses show accuracy changes in percentage points and computational-cost ratios relative to CFG. Latency reports mean ± standard deviation; memory reports peak allocation. We present setup details in Appendix E.3 .
Figure 4: Category-wise results and ablations. Left: mean satisfaction accuracy across five models, with min–max whiskers. Right: guidance-strength (top) and CFG–TSG (bottom) ablations.
Figure 5: Representative examples comparing CFG-only sampling (top row of each panel) with TSG (bottom row) on ordered entry, three-state transition, and overlapping handover respectively. The first two examples are generated using Wan2.1 T2V-1.3B, and the third example is generated using Wan2.2 TI2V-5B in text-to-video mode.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Category
Temporal behavior
Two-state transition
A subject or setting changes from state A to B .
Three-state transition
A subject or setting progresses through A , B , then C .
Ordered entry of two entities
One entity is initially present; a second joins it and both remain.
Ordered entry of three entities
Two further entities enter in sequence, accumulating in the scene.
Entity handover with overlap
A alone is followed by A and B together, then B alone.
Entity exit and re-entry
An entity is present, becomes absent, then returns.
Appendix
Table 3: The 11 temporal constraint categories in the main benchmark, each containing 20 examples. Descriptions summarize the temporal structure; individual specifications also encode timing, absence, and persistence requirements.
Entity
Candidate states or attributes
airplane
configuration: landed, in flight
apple
condition: whole, sliced
balloon
inflation: visibly deflated, visibly inflated
bicycle
orientation: upright, lying on its side
bird
configuration: perched, in flight
boat
orientation: upright, capsized
Appendix
Table 4: Examples of object entities with corresponding states.
Entity
Candidate states or attributes
beach
illumination: bright daylight, orange horizon, dark night
city street
illumination: bright daylight, dark night; precipitation: no visible precipitation, visible falling rain, visible falling snow
farm field
ground cover: green crops, bare soil, snow-covered; precipitation: no visible precipitation, visible falling rain, visible falling snow
forest
dominant foliage: green foliage, autumn-colored foliage, completely bare branches, snow-covered trees; precipitation: no visible precipitation, visible falling rain, visible falling snow
kitchen
artificial lighting: lights visibly illuminated, no lights visibly illuminated
lake
surface: visible open water, completely ice-covered; visibility: far shore clearly visible, far shore obscured by fog; precipitation: no visible precipitation, visible falling rain, visible falling snow
Appendix
Table 5: Examples of scene and setting entities with corresponding states.
Model
Resolution
Frames/FPS
Latent frames
Steps
CFG
Wan 2.1 1.3B
480×832
81/16
21
50
6
Wan 2.2 TI2V 5B
704×1280
121/24
31
50
5
Wan 2.2 T2V A14B
704×1280
81/16
21
40
4 / 3
CogVideoX 2B
480×720
49/8
13
50
6
CogVideoX 5B
480×720
49/8
13
50
dynamic (6)
Appendix
Table 6: Generation settings for the TSG benchmark. Wan 2.2 A14B uses CFG scales 4 and 3 for its high- and low-noise experts, respectively. CogVideoX 5B uses its dynamic CFG schedule with scale parameter 6. All methods compared within a model share these settings. The TempoControl comparison uses the Wan 2.1 row.
Text-to-video diffusion models generate realistic videos, but often fail on prompts requiring fine-grained compositional understanding, such as relations between entities, attributes, actions, and motion directions. We hypothesize that these failures need not be addressed by retraining the generator, but can instead be mitigated by steering the denoising process using the model's own internal grounding signals. We propose CVG, an inference-time guidance method for improving compositional faithfulness in frozen text-to-video models. Our key observation is that cross-attention maps already encode how prompt concepts are grounded across space and time. We train a lightweight compositional classifier on these attention features and use its gradients during early denoising steps to steer the latent trajectory toward the desired composition. Built on a frozen VLM backbone, the classifier transfers across semantically related composition labels rather than relying only on narrow category-specific features. CVG improves compositional generation without modifying the model architecture, fine-tuning the generator, or requiring layouts, boxes, or other user-supplied controls. Experiments on compositional text-to-video benchmarks show improved prompt faithfulness while preserving the visual quality of the underlying generator.
Ariel Shaulov, Eitan Shaar, Amit Edenzon +2
1Tel-Aviv University · 2Independent Researcher · 3Bar Ilan University +1
Text-to-video diffusion models generate temporally coherent content from natural language, yet when a prompt describes an early scene that persists while a new event emerges on top of it---such as "a tall sandcastle standing on a beach where a wave rushes in and washes it away"---generation frequently fails to realize the late-segment event in the corresponding frames. We identify this failure as Temporal Prior Suppression (TPS): the dominant prior of the early segment captures the cross-attention trajectory across the temporal axis and suppresses the guidance signal needed for late-segment realization, a competing tendency existing guidance mechanisms do not model. We introduce Temporal Prior Decoupling (TPD), a training-free framework that restores suppressed late-segment signals during diffusion sampling. TPD constructs a temporal counterfactual by conditioning on the early segment alone, and defines the discrepancy between the full-prompt and counterfactual trajectories as a suppressed signal direction. Rather than removing this direction as in prior subtractive projection methods, TPD restores it through a frame-selective lower-bound constraint resolved jointly over diffusion timestep and video frame, realizing the suppressed event in the late frames without disrupting early-segment coherence: where prior work enforces upper-bound feasibility to remove unwanted semantics, TPD enforces lower-bound feasibility to guarantee suppressed-signal contribution. TPD runs entirely within standard diffusion sampling without retraining, and is defined purely in classifier-free guidance space, making it backbone-agnostic by construction. Experiments show that TPD significantly improves late-concept realization while preserving temporal coherence and visual fidelity, and that the targeted suppression recurs across distinct text-to-video backbones.
Taewon Kang, Matthias Zwicker
University of Maryland at College Park, United States
Generating high-quality videos from complex temporal descriptions that contain multiple sequential actions is a key unsolved problem. Existing methods are constrained by an inherent trade-off: using multiple short prompts fed sequentially into the model improves action fidelity but compromises temporal consistency, while a single complex prompt preserves consistency at the cost of prompt-following capability. We attribute this problem to two primary causes: 1) temporal misalignment between video content and the prompt, and 2) conflicting attention coupling between motion-related visual objects and their associated text conditions. To address these challenges, we propose a novel, training-free attention mechanism, Temporal-wise Separable Attention (TS-Attn), which dynamically rearranges attention distribution to ensure temporal awareness and global coherence in multi-event scenarios. TS-Attn can be seamlessly integrated into various pre-trained text-to-video models, boosting StoryEval-Bench scores by 33.5% and 16.4% on Wan2.1-T2V-14B and Wan2.2-T2V-A14B with only a 2% increase in inference time. It also supports plug-and-play usage across models for multi-event image-to-video generation. The source code and project page are available at https://github.com/Hong-yu-Zhang/TS-Attn.
Hongyu Zhang, Yufan Deng, Zilin Pan +6
Peking University, Shenzhen Graduate School · Zhejiang University · Nanjing University +3