Pith. sign in

REVIEW 3 major objections 5 minor 62 references

Text-to-video models lock onto the opening scene and suppress later events; a training-free lower-bound on guidance can restore them.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 23:26 UTC pith:ZHRTXUMF

load-bearing objection Clean restorative dual to subtractive CFG feasibility work; real failure mode, solid ablations on Mochi, but backbone-agnostic gains are still mostly an argument plus qualitative frames. the 3 major comments →

arxiv 2607.26706 v1 pith:ZHRTXUMF submitted 2026-07-29 cs.CV

TPD: Temporal Prior Decoupling for Text-to-Video Diffusion Models

Taewon Kang , Matthias Zwicker This is my paper
classification cs.CV
keywords text-to-video diffusiontemporal prior suppressionclassifier-free guidanceinference-time guidancefeasibility constraintscounterfactual conditioningtemporal control
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

When a single prompt asks for an early scene that should persist while a new event appears over it—a sandcastle that a wave then washes away—current text-to-video diffusion models often render only the opening scene for the whole clip. The paper names this Temporal Prior Suppression: the early scene’s strong prior captures cross-attention along time and starves the late event of guidance. It treats the failure as a restorative feasibility problem, not a negation or attribute-morph problem. Temporal Prior Decoupling builds a counterfactual by conditioning only on the early segment, takes the difference from the full-prompt prediction as the suppressed direction, and projects classifier-free guidance onto a lower bound scheduled jointly over denoising step and video frame. The result is a closed-form, minimal correction that adds signal only where late frames under-realize the event, without retraining and without depending on a particular backbone architecture.

Core claim

Temporal Prior Suppression is a recurring failure of current text-to-video diffusion models: an early-scene prior dominates the guidance trajectory and under-expresses a requested late event in its intended frames. The paper shows this omission can be corrected at sampling time by enforcing a lower-bound feasibility constraint on the guidance update along the full-prompt versus early-only counterfactual direction, scheduled over both diffusion timestep and frame index.

What carries the argument

Temporal Prior Decoupling (TPD): the suppressed direction a_τ = ε_text − ε_counter (full prompt minus early-segment-only conditioning), with a minimal-energy projection that enforces a_τᵀ δ ≥ c(τ, i) via a two-dimensional (timestep × frame) lower-bound schedule—restorative rather than subtractive guidance correction.

Load-bearing premise

The method depends on a one-shot language-model split of the prompt into early and late parts plus estimates of transition point and prior strengths, none of which have ground truth; if those estimates are wrong, restoration is mis-timed or under-scaled.

What would settle it

On prompts that specify a persisting early scene and a genuine late emergence, run the same frozen backbone with and without TPD and check whether late-window frames realize the late concept (coverage and violation rate) while early frames and overall fidelity stay intact; if late realization does not improve, or improves only by discarding the early scene, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Upper-bound suppression, drift removal, and lower-bound restoration form one feasibility family for semantic guidance correction; the temporal case was the missing member.
  • Position-resolved emergence—not only whole-clip text similarity—becomes a measurable evaluation target for video generators.
  • Because the fix lives purely in classifier-free guidance space, the same restoration recipe can transfer across video backbones without retraining.
  • Attribute-morph and pre-segmented multi-prompt methods do not cover single-prompt emergence over a persisting scene; TPD addresses that gap at inference time.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Image generators with spatial concept neglect may admit an analogous lower-bound restoration if a ‘where’ axis replaces the video-frame axis.
  • Multi-event or overlapping timelines would need several simultaneous lower bounds rather than one early/late split.
  • If decomposition parameters were estimated from the backbone’s own predictions instead of an external language model, the pipeline would become fully self-contained at inference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper identifies Temporal Prior Suppression (TPS): in text-to-video diffusion, an early-scene prior can dominate the CFG trajectory so that a late emergent event specified in a single unsegmented prompt fails to appear in the intended frames. It proposes Temporal Prior Decoupling (TPD), a training-free method that builds a temporal counterfactual by conditioning on the early segment alone, defines a_τ = ε_text − ε_counter as the suppressed direction, and restores it via a minimal-energy lower-bound projection a_τ^⊤ δ ≥ c(τ,i) with a joint (diffusion-timestep × video-frame) schedule. The formulation is cast as the restorative dual of subtractive feasibility methods (NEGATE, DCR), with closed-form KKT solution. Evaluation uses a new 400-prompt emergence suite and reports gains on CLIP-early, TCS, and TVR on Mochi, with qualitative recurrence of TPS on HunyuanVideo and CogVideoX.

Significance. If the claims hold, the work makes three useful contributions: (1) a crisp formalization of a failure mode that standard clip–text metrics miss and that is specific to temporal placement rather than negation or attribute morphs; (2) a clean, architecture-neutral correction entirely in CFG space, with a closed-form projection and ablations that isolate frame axis, timestep schedule, and ρ-adaptive strength; (3) an evaluation suite organized by emergence mechanism rather than surface scene type. Completing the upper-bound / drift-removal / lower-bound family is a coherent conceptual advance. The training-free, backbone-agnostic design is practically valuable if quantitative transfer is demonstrated.

major comments (3)
  1. [Abstract; Table 1; Appendix H] Abstract, Introduction, and §Experiments claim backbone-agnostic improvement and that TPS/TPD transfer across distinct T2V systems. Table 1 reports standard-CFG baselines for HunyuanVideo and CogVideoX, and Appendix D shows qualitative recurrence of TPS, but every controlled TPD result—main metrics, Always-Add vs projection, and the three schedule ablations (Tables 1–3, Fig. 3, App. F)—is on Mochi only. Appendix H explicitly defers quantitative transfer and calls retuning (w, α0, β, q) “engineering rather than formulation.” For a central claim of backbone-agnostic restoration defined purely in CFG space, at least one other backbone needs full TPD numbers (TCS/TVR/CLIP-early ± ablations), or the claim must be narrowed to “failure recurs; method validated on one backbone.”
  2. [Appendix E; Appendix G; Table 1] GPT-4o is used both for D(p) = {p_early, p_late, γ, ρ_early, ρ_late} (App. G) and for the primary judged metrics TCS/TVR/EPS (App. E), with only protocol mitigations (temperature 0, separate queries, fixed rubric). Appendix A admits these scalars have no ground truth and that poor estimates cause under-correction. Because generation localization (frame sigmoid, α ∝ max(0, ρ_early−ρ_late)) and the headline compliance scores share the same model family, circularity risk is load-bearing. A human study on a subset ranking methods on late-concept realization, and/or an open/local decomposer ablation with frozen judge, is needed to support that gains are not judge–decomposer agreement.
  3. [Table 3; Appendix A; Conclusion] On CAUSAL (Table 3), TPD reaches TCS 2.36 and TVR 0.80—still very low absolute realization—consistent with Appendix A’s admission that guidance-space restoration cannot force trajectories the backbone will not follow. The paper’s strongest qualitative examples (e.g., sandcastle) work, but the claim “significantly improves late-concept realization” should be qualified by mechanism: large gains on ONSET/ENTRY/SCENE, limited mitigation on causal collapse. Without that stratification in the abstract/conclusion, the aggregate TCS/TVR can overstate reliability on the hardest and most distinctive regime the dataset was built to stress.
minor comments (5)
  1. [Figure 2; Appendix D] Figure 2 and Appendix D figures are informative, but several multi-panel sets (Figs. 4–6) are hard to read at print scale; consider fewer frames per row or higher-resolution crops of the late window only.
  2. [Method] Notation switches between t and τ for the diffusion step (Eqs. 1–13); unify to one symbol in the main method section.
  3. [Related Work] Related Work cites several 2025–2026 arXiv preprints appropriately, but a short explicit comparison table (segmented supervision vs morph vs TPD input assumptions) would help readers place the setting faster.
  4. [Appendix J; Implementation Details] Table 4 runtime: full TPD is ~2.2× Mochi (248s vs 114s). Stating this cost clearly in the main Experiments paragraph, not only Appendix J, would set expectations.
  5. [Appendix D figures] Typos/formatting: “CASUAL” appears in figure labels where “CAUSAL” is meant (Figs. 4–6, 11, 19, 27); “acrossh categories” in Fig. 6 caption; occasional missing spaces in compound words from PDF extraction should be checked in source.

Circularity Check

1 steps flagged

Core TPD derivation is independent; only mild self-citation props up judge metrics, not the method itself.

specific steps
  1. self citation load bearing [Appendix E, On the Reliability of Automated Judging; also Experiments/Quantitative Evaluation]
    "Prior work on constrained video generation has found that judge-based compliance scores and human perceptual rankings agree on method ordering for closely related tasks (Kang and Lin 2026), supporting the use of these scores as a proxy for human assessment of temporal faithfulness."

    TCS/TVR/EPS carry the main quantitative claim of late-concept restoration. Their human validity is justified primarily by a citation to overlapping-author work (NEGATE, Kang & Lin 2026) on related constrained-generation tasks, not by a fresh human study in this paper. This is load-bearing for trusting the judged metrics, not for the math of the projection; the derivation of δ* remains independent. Mild, pattern-3 only.

full rationale

Walking the claimed chain—TPS as early-prior dominance of CFG, aτ := ε_text − ε_counter from a temporal counterfactual, lower-bound half-space aτ⊤δ ≥ c(τ,i), closed-form KKT projection, 2-D (timestep × frame) schedule—shows no step that is true by definition of its inputs. aτ is an empirical denoiser difference, not a quantity fitted to the reported metrics; µ is zero when the constraint already holds; ablations (frame axis, timestep schedule, ρ-adaptive α, Always-Add) change distinct metrics in distinct ways, which would be impossible if success were tautological. The dual framing with NEGATE/DCR is conceptual scaffolding from overlapping-author preprints, not a uniqueness theorem that forces TPD. Shared use of GPT-4o for decomposition and TCS/TVR/EPS, and reuse of the same γ for schedule and late-frame partition, are methodological confounds and evaluation alignment choices, not derivation circularity: late-concept content must still appear in pixels. The only pattern that lightly matches is self-citation used to underwrite automated-judge reliability (Kang & Lin 2026). That supports confidence in the scores, not the projection formula. Backbone-agnostic transfer is under-demonstrated empirically (Mochi-only controlled tables) but that is a scope/evidence gap, not circularity. Score 2.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The claim rests on standard diffusion/CFG math, the modeling assumption that late-event under-expression is captured by ε_text − ε_counter, LLM-derived schedule scalars without ground truth, hand-chosen schedule hyperparameters, and a custom evaluation distribution plus VLM judge. No new physical entities; invented constructs are methodological (TPS, TPD, suppressed direction, 2D bound).

free parameters (5)
  • α0 (global restoration strength) = 1.0
    Overall scale of the lower bound; set to 1.0 by hand for Mochi.
  • β (frame-sigmoid sharpness) = 10.0
    Controls how abruptly correction turns on around γ; set to 10.0.
  • q (timestep schedule exponent) = 2.0
    ϕ(τ)=(τ/T)^q concentrates correction in early denoising; q=2.0 chosen analogously to prior work.
  • ρearly, ρlate, γ (per-prompt LLM estimates) = per-prompt GPT-4o outputs in [0,1]
    Drive α and frame localization; no ground truth, fallback (0.5, 0.7, 0.3); central to when/how much TPD acts.
  • CFG scale w and backbone defaults = backbone defaults (scale-up ablation uses 9.0)
    Inherited host-model guidance magnitude; paper notes retuning needed across backbones.
axioms (5)
  • domain assumption Classifier-free guidance update δ_ref = w(ε_text − ε_uncond) is the right locus for semantic temporal correction.
    Stated throughout Method and Related Work; shared with SEGA/S-CFG/NEGATE-style steering.
  • domain assumption a_τ = ε_text − ε_counter isolates the suppressed late-segment contribution rather than mere magnitude rescaling of δ_ref.
    Core construction in Temporal Counterfactual Construction; claimed not collinear with δ_ref because pearly ≠ p.
  • standard math Minimal-energy projection onto the half-space a_τ^⊤ δ ≥ c(τ,i) restores omitted signal without harming reverse-time ODE stability.
    KKT closed form and non-expansive projection arguments in Appendix I.
  • ad hoc to paper Prompts are essentially two-part (one persisting scene + one emergent event); multi-event overlap is out of scope.
    Limitations (Appendix A) and dataset design principles.
  • domain assumption GPT-4o vision-language judgments with fixed rubric are a valid proxy for human temporal faithfulness rankings.
    Appendix E; supported by citation to agreement on related tasks, not new human study here.
invented entities (3)
  • Temporal Prior Suppression (TPS) independent evidence
    purpose: Name and formalize the failure mode of early-scene dominance over late-event frames.
    Defined in Introduction/Abstract as distinct from negation and compositional collapse; evidenced by cross-backbone qualitative/quantitative patterns on the custom suite.
  • Suppressed signal direction a_τ from temporal counterfactual no independent evidence
    purpose: Provide the restorative direction for the lower-bound constraint.
    Constructed operationally from frozen denoiser branches; not a new physical object, but a paper-specific guidance-space object.
  • Two-dimensional lower-bound schedule c(τ,i) no independent evidence
    purpose: Localize restoration to late frames and structure-forming timesteps, scaled by prior gap.
    Defined in Frame-Selective Amplification Schedule; value shown via ablations.

pith-pipeline@v1.2.0-daily-grok45 · 37529 in / 3883 out tokens · 81132 ms · 2026-07-30T23:26:44.585607+00:00 · methodology

0 comments
read the original abstract

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.

Figures

Figures reproduced from arXiv: 2607.26706 by Matthias Zwicker, Taewon Kang.

Figure 1
Figure 1. Figure 1: Overview of Temporal Prior Decoupling (TPD). An LLM decomposes the unsegmented prompt p into the early segment pearly, the late segment plate, the transition point γ, and the prior strengths ρearly, ρlate. (a) Restorative Guidance Pipeline: Three branches of the frozen denoiser are evaluated—unconditional ϵuncond, full-prompt ϵtext, and temporal counterfactual ϵcounter conditioned on pearly alone—yielding … view at source ↗
Figure 2
Figure 2. Figure 2: Representative qualitative comparison (STRUCTURAL). Prompt: “a tall sandcastle standing on a beach where a wave rushes in and washes it completely flat leaving only wet sand behind”; frames in temporal order. All baselines carry the intact structure to the final frame— HunyuanVideo renders breaking waves that never reach the castle, and CogVideoX degenerates into a washed-out frame losing the beach context… view at source ↗
Figure 3
Figure 3. Figure 3: Ablation study on the sandcastle prompt, frames in temporal order. CFG Scale-Up: the structure never falls— scaling an update does not redirect it. Always-Add: the castle stays intact throughout—a fixed additive push cannot over￾come a strong early-scene prior. w/o Frame Axis: water already intrudes in the opening frame and a clean early state is never established—the loss of where to intervene. w/o Timest… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison across categories (Set 1). We evaluate our method (TPD) against state-of-the-art video diffusion models (Mochi, HunyuanVideo, CogVideoX) on eight temporal emergence categories: ENTRY, SCENE, ONSET, STATE, CAUSAL, WEATHER_NIGHT, ACTIVITY, and STRUCTURAL. Each row shows three frames sampled from gener￾ated videos for a representative prompt per category. Baseline models consistently fa… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison across categories (Set 2). This figure presents a second set of representative prompts covering all eight temporal emergence categories (ENTRY, SCENE, ONSET, STATE, CAUSAL, WEATHER_NIGHT, ACTIVITY, STRUCTURAL). These examples include more challenging emergence scenarios such as a family of hippopotamuses entering a calm river bend, a hailstorm erupting over a highland plateau, and a … view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison acrossh categories (Set 3). This figure presents a third set of prompts spanning all eight temporal emergence categories, including a flock of flamingos landing in a bare lakeside beach, a lightning storm rolling into a Milky Way campsite, and a tall stack of diner plates shattering on the kitchen floor. These cases probe stronger early-scene priors—frozen surfaces, bare natural sett… view at source ↗
Figure 7
Figure 7. Figure 7: ENTRY). Prompt: “A quiet mountain lake at dawn with still water and empty shoreline where a herd of deer walks down to the bank and begins drinking.” [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: SCENE. Prompt: “A silent factory floor before the shift starts where workers stream in and machines start up.” [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: ONSET. Prompt: “A dry basin under a pale overcast sky where a haboob wall of dust rises on the horizon and races toward the camera.” [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: STATE. Prompt: “A pristine white snow field under clear skies where a flock of dark birds flies across the field in moving black shapes.” [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: CAUSAL. Prompt: “A smooth sea ice surface in arctic light where breaches from below and the ice cracks and collapses inward.” [PITH_FULL_IMAGE:figures/full_fig_p029_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: WEATHER_NIGHT. Prompt: “A calm bay with boats anchored under a clear night sky where a storm rolls in and lightning illuminates the churning water.” [PITH_FULL_IMAGE:figures/full_fig_p030_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: ACTIVITY. Prompt: “A completely empty highway at dawn where cars begin appearing and the road fills with flowing traffic.” [PITH_FULL_IMAGE:figures/full_fig_p031_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: STRUCTURAL. Prompt: “A tall sandcastle standing on a beach where a wave rushes in and washes it completely flat leaving only wet sand behind.” [PITH_FULL_IMAGE:figures/full_fig_p032_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: ENTRY. Prompt: “A calm river bend at dawn with no movement where a family of hippopotamuses enters the water from the bank.” [PITH_FULL_IMAGE:figures/full_fig_p033_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: SCENE. Prompt: “A quiet airport gate area before boarding where passengers stream in and the area becomes packed with travelers.” [PITH_FULL_IMAGE:figures/full_fig_p034_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: ONSET. Prompt: “A quiet highland plateau under a pale sky where a hailstorm erupts overhead and stones begin hammering the flat ground.” [PITH_FULL_IMAGE:figures/full_fig_p035_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: STATE. Prompt: “A blank white canvas on an easel where a painter begins applying bold strokes of red and blue paint.” [PITH_FULL_IMAGE:figures/full_fig_p036_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: CAUSAL. Prompt: “A flat glacier surface in cold white light where meltwater pooling on top drains suddenly and the surface cracks.” [PITH_FULL_IMAGE:figures/full_fig_p037_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: WEATHER_NIGHT. Prompt: “A still moonlit field where a thunderstorm rolls in from the horizon and lightning strikes the open ground.” [PITH_FULL_IMAGE:figures/full_fig_p038_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: ACTIVITY. Prompt: “A still port terminal where cargo trucks begin arriving and the loading bays fill with vehicles.” [PITH_FULL_IMAGE:figures/full_fig_p039_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: STRUCTURAL. Prompt: “A tall pile of autumn leaves raked into a heap in a garden that a child jumps into and scatters.” [PITH_FULL_IMAGE:figures/full_fig_p040_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: ENTRY. Prompt: “A bare lakeside beach at sunrise where a flock of flamingos glides in and lands in the shallows in a wash of pink.” [PITH_FULL_IMAGE:figures/full_fig_p041_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: SCENE. Prompt: “A bare city rooftop at sunrise where a film crew arrives and sets up equipment across the space.” [PITH_FULL_IMAGE:figures/full_fig_p042_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: ONSET. Prompt: “A calm frozen lake under a pale winter sky where a crack propagates from the near shore and splits the surface.” [PITH_FULL_IMAGE:figures/full_fig_p043_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: STATE. Prompt: “A pristine white tablecloth on an outdoor table where autumn leaves fall and scatter across the surface.” [PITH_FULL_IMAGE:figures/full_fig_p044_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: CAUSAL. Prompt: “A smooth frozen fen in pale light where the ice is lifted by a frost heave below and the surface shatters.” [PITH_FULL_IMAGE:figures/full_fig_p045_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: WEATHER_NIGHT. Prompt: “A clear night campsite under the Milky Way where a storm rolls in and lightning crashes nearby.” [PITH_FULL_IMAGE:figures/full_fig_p046_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: ACTIVITY. Prompt: “A bare motocross track before practice where riders appear and begin taking the first jumps.” [PITH_FULL_IMAGE:figures/full_fig_p047_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: STRUCTURAL. Prompt: “A tall stack of diner plates on a counter that slips and shatters on the kitchen floor.” [PITH_FULL_IMAGE:figures/full_fig_p048_30.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 24 linked inside Pith

  1. [1]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Ditctrl: Exploring attention control in multi-modal diffusion transformer for tuning-free multi-prompt longer video generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  2. [2]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Mind the time: Temporally-controlled multi-event video generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  3. [3]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    From Prompt to Progression: Taming Video Diffusion Models for Seamless Attribute Transition , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  4. [4]

    arXiv preprint arXiv:2604.10030 , year=

    Prompt Relay: Inference-Time Temporal Control for Multi-Event Video Generation , author=. arXiv preprint arXiv:2604.10030 , year=

    Pith/arXiv arXiv

  5. [5]

    Advances in Neural Information Processing Systems , volume=

    Sega: Instructing text-to-image models using semantic guidance , author=. Advances in Neural Information Processing Systems , volume=

  6. [6]

    arXiv preprint arXiv:2312.15964 , year=

    Semantic guidance tuning for text-to-image diffusion models , author=. arXiv preprint arXiv:2312.15964 , year=

    Pith/arXiv arXiv

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Rethinking the spatial inconsistency in classifier-free diffusion guidance , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  8. [8]

    arXiv preprint arXiv:2605.06512 , year=

    DCR: Counterfactual Attractor Guidance for Rare Compositional Generation , author=. arXiv preprint arXiv:2605.06512 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2603.06533 , year=

    NEGATE: Constrained Semantic Guidance for Linguistic Negation in Text-to-Video Diffusion , author=. arXiv preprint arXiv:2603.06533 , year=

    arXiv

  10. [10]

    arXiv preprint arXiv:2503.06310 , year=

    Text2story: Advancing video storytelling with text guidance , author=. arXiv preprint arXiv:2503.06310 , year=

    Pith/arXiv arXiv

  11. [11]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Frequency-Aware Flow Matching for High-Quality Image Generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  12. [12]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Bias Is a Subspace, Not a Coordinate: A Geometric Rethinking of Post-hoc Debiasing in Vision-Language Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  13. [13]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    TempoControl: Temporal Attention Guidance for Text-to-Video Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  14. [14]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Image generation from contextually-contradictory prompts , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  15. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Make it SING: Analyzing Semantic Invariants in Classifiers , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  16. [16]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Clip is shortsighted: Paying attention beyond the first sentence , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  17. [17]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    Pith/arXiv arXiv

  18. [18]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  19. [19]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  20. [20]

    Advances in neural information processing systems , volume=

    Photorealistic text-to-image diffusion models with deep language understanding , author=. Advances in neural information processing systems , volume=

  21. [21]

    ACM transactions on Graphics (TOG) , volume=

    Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models , author=. ACM transactions on Graphics (TOG) , volume=. 2023 , publisher=

    2023

  22. [22]

    European conference on computer vision , pages=

    Compositional visual generation with composable diffusion models , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  23. [23]

    Advances in neural information processing systems , volume=

    Video diffusion models , author=. Advances in neural information processing systems , volume=

  24. [24]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Align your latents: High-resolution video synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  25. [25]

    arXiv preprint arXiv:2406.08070 , year=

    Cfg++: Manifold-constrained classifier free guidance for diffusion models , author=. arXiv preprint arXiv:2406.08070 , year=

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2212.05032 , year=

    Training-free structured diffusion guidance for compositional text-to-image synthesis , author=. arXiv preprint arXiv:2212.05032 , year=

    Pith/arXiv arXiv

  27. [27]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Freeu: Free lunch in diffusion u-net , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  28. [28]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Universal guidance for diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  29. [29]

    International Conference on Learning Representations , volume=

    Cogvideox: Text-to-video diffusion models with an expert transformer , author=. International Conference on Learning Representations , volume=

  30. [30]

    arXiv preprint arXiv:2511.18870 , year=

    Hunyuanvideo 1.5 technical report , author=. arXiv preprint arXiv:2511.18870 , year=

    Pith/arXiv arXiv

  31. [31]

    2024 , month =

    Genmo AI , title =. 2024 , month =

    2024

  32. [32]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Dymo: Training-free diffusion model alignment with dynamic multi-objective scheduling , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  33. [33]

    arXiv preprint arXiv:2502.05625 , year=

    Training-free constrained generation with stable diffusion models , author=. arXiv preprint arXiv:2502.05625 , year=

    arXiv

  34. [34]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Winoground: Probing vision and language models for visio-linguistic compositionality , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  35. [35]

    arXiv preprint arXiv:2210.01936 , year=

    When and why vision-language models behave like bags-of-words, and what to do about it? , author=. arXiv preprint arXiv:2210.01936 , year=

    Pith/arXiv arXiv

  36. [36]

    Advances in neural information processing systems , volume=

    Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality , author=. Advances in neural information processing systems , volume=

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment , author=. Advances in Neural Information Processing Systems , volume=

  38. [38]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Vision-language models do not understand negation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  39. [39]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    NegVQA: Can vision language models understand negation? , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  40. [40]

    no" to say

    Learn" no" to say" yes" better: Improving vision-language models via negations , author=. arXiv preprint arXiv:2403.20312 , year=

    Pith/arXiv arXiv

  41. [41]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI 2026) , year=

    Not Just What's There: Enabling CLIP to Comprehend Negated Visual Descriptions Without Fine-Tuning , author=. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI 2026) , year=

    2026

  42. [42]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Erasing concepts from diffusion models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  43. [43]

    Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

    Unified concept editing in diffusion models , author=. Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

  44. [44]

    arXiv preprint arXiv:2311.15127 , year=

    Stable video diffusion: Scaling latent video diffusion models to large datasets , author=. arXiv preprint arXiv:2311.15127 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2210.02303 , year=

    Imagen video: High definition video generation with diffusion models , author=. arXiv preprint arXiv:2210.02303 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2209.14792 , year=

    Make-a-video: Text-to-video generation without text-video data , author=. arXiv preprint arXiv:2209.14792 , year=

    Pith/arXiv arXiv

  47. [47]

    SIGGRAPH Asia 2024 Conference Papers , pages=

    Lumiere: A space-time diffusion model for video generation , author=. SIGGRAPH Asia 2024 Conference Papers , pages=

    2024

  48. [48]

    arXiv preprint arXiv:2311.10709 , year=

    Emu video: Factorizing text-to-video generation by explicit image conditioning , author=. arXiv preprint arXiv:2311.10709 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2309.17444 , year=

    Llm-grounded video diffusion models , author=. arXiv preprint arXiv:2309.17444 , year=

    Pith/arXiv arXiv

  50. [50]

    2024 , month=

    OpenAI , title=. 2024 , month=

    2024

  51. [51]

    arXiv preprint arXiv:2412.03603 , year=

    Hunyuanvideo: A systematic framework for large video generative models , author=. arXiv preprint arXiv:2412.03603 , year=

    Pith/arXiv arXiv

  52. [52]

    2025 , eprint=

    Movie Gen: A Cast of Media Foundation Models , author=. 2025 , eprint=

    2025

  53. [53]

    arXiv preprint arXiv:2308.06571 , year=

    Modelscope text-to-video technical report , author=. arXiv preprint arXiv:2308.06571 , year=

    Pith/arXiv arXiv

  54. [54]

    arXiv preprint arXiv:2305.18264 , year=

    Gen-l-video: Multi-text to long video generation via temporal co-denoising , author=. arXiv preprint arXiv:2305.18264 , year=

    Pith/arXiv arXiv

  55. [55]

    arXiv preprint arXiv:2403.14773 , year=

    Streamingt2v: Consistent, dynamic, and extendable long video generation from text , author=. arXiv preprint arXiv:2403.14773 , year=

    Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2310.15169 , year=

    Freenoise: Tuning-free longer video diffusion via noise rescheduling , author=. arXiv preprint arXiv:2310.15169 , year=

    Pith/arXiv arXiv

  57. [57]

    arXiv preprint arXiv:2211.13221 , year=

    Latent video diffusion models for high-fidelity long video generation , author=. arXiv preprint arXiv:2211.13221 , year=

    Pith/arXiv arXiv

  58. [58]

    arXiv preprint arXiv:2205.15868 , year=

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers , author=. arXiv preprint arXiv:2205.15868 , year=

    Pith/arXiv arXiv

  59. [59]

    arXiv preprint arXiv:2210.02399 , year=

    Phenaki: Variable length video generation from open domain textual description , author=. arXiv preprint arXiv:2210.02399 , year=

    Pith/arXiv arXiv

  60. [60]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Preserve your own correlation: A noise prior for video diffusion models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  61. [61]

    International Journal of Computer Vision , pages=

    Lavie: High-quality video generation with cascaded latent diffusion models , author=. International Journal of Computer Vision , pages=. 2024 , publisher=

    2024

  62. [62]

    arXiv preprint arXiv:2506.12544 , year=

    Constrained diffusers for safe planning and control , author=. arXiv preprint arXiv:2506.12544 , year=

    Pith/arXiv arXiv