Pith. sign in

REVIEW 4 major objections 5 minor 116 references

This paper claims that the craft of film editing—choosing which shots to keep, where to cut, and how to pace them—can be learned by an autoregressive transformer trained on movies, and that the resulting model, FilmGPT, edits raw footage in

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 ↗

An autoregressive transformer with an explicit cut token and footage-constrained decoding edits raw video into sequences that people rate as better than two prior automated editing methods.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Genuinely new AR editing formulation, but both headline evals are setup-confounded; worth refereeing with requested ablations. the 4 major comments →

arxiv 2607.14645 v2 pith:ZQ4STNFG submitted 2026-07-16 cs.CV

Autoregressive Modeling of Film with Applications in Video Montage

Marcelo Sandoval-Casta\~neda , Fabian Caba Heilbron , Shiry Ginosar , Bryan Russell , Josef Sivic , Alexei A. Efros , Greg Shakhnarovich This is my paper
classification cs.CV
keywords video editingautoregressive transformerfilm grammarshot sequencingfootage-constrained decodingvideo montagenext-token predictionuser study
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.

The reading

The paper aims to establish that the 'grammar' of film editing can be learned from data rather than hand-coded. The authors train a long-context autoregressive transformer, called FilmGPT, on 6,200 hours of professionally edited movies, encoding each frame as discrete tokens and marking shot boundaries with an explicit cut token. At inference, instead of generating frames, a footage-constrained decoding algorithm searches the input raw footage for the next shot and trim point that maximizes the learned likelihood. The authors show this learned statistical model outperforms prior methods on a shot-sequence-ordering benchmark and, in a forced-choice user study, human participants prefer its edits to those of two prior text-driven editing systems by 83% and 61%. If correct, this means low- and mid-level editing decisions can be largely automated or assisted without explicit rules or language descriptions of the footage.

Core claim

On its own terms, the central claim is that next-token prediction over discrete visual tokens, interleaved with a dedicated cut token, captures enough of the structure of professional editing to drive shot selection, trimming, and ordering. FilmGPT tokenizes each frame of a movie into 32 discrete tokens via TiTok-L, inserts a single ⟨CUT⟩ token at every shot boundary, and trains a decoder-only transformer with sliding-window attention (up to 128k tokens of context) using a cut-weighted cross-entropy loss that upweights the cut token and the tokens immediately following it. Audio features are injected via time-aligned cross-attention. Because the model is trained purely on the sequence of fra

What carries the argument

The central machinery is an autoregressive sequence model over frame tokens plus a special cut token. FilmGPT treats an edited movie as a token sequence x_1,...,x_T where each frame is a block of 32 TiTok-L tokens and each shot boundary is a ⟨CUT⟩ token; it maximizes the factorized probability p(x)=∏ p(x_t | x_<t). At inference, the footage-constrained decoding algorithm enforces a hard subsequence constraint—every generated prefix must match a contiguous span of a video in the input collection—and finds the most likely shot and trim length by greedy beam search. The cut-weighted cross-entropy loss is what makes the model pay attention to editing decisions rather than just within-shot visual

Load-bearing premise

Editing decisions are determined by past context alone: FilmGPT and its decoding are strictly causal, so any editing idiom whose purpose is to set up a later shot cannot be learned or reproduced.

What would settle it

A benchmark or user study that includes edits whose correctness depends on future context (e.g., a point-of-view shot that only makes sense once the next shot is seen) would settle whether the causal, prefix-only model captures full editing grammar; if a non-causal or bidirectional model performs better on such items while matching FilmGPT on prefix-determined edits, the clean autoregressive claim would be weakened.

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

If this is right

  • Editing decisions can be cast as causal next-token prediction, so methods that scale language models (more data, longer context, better tokenizers) should directly improve editing quality.
  • An explicit cut token lets a single model jointly decide shot selection and trimming; removing it drops shot-ordering accuracy from 29.7% to 25.3% in the paper's ablations.
  • Extending context from 16k to 128k tokens raises shot-ordering accuracy from 34.5% to 44.5%, and adding audio conditioning pushes it to 50.5%, showing that long-range and audio cues carry much of editing's structure.
  • The model needs no manual annotations such as shot size or genre; it learns directly from raw movie sequences and preceding context, outperforming a method that uses such labels.
  • The same inference algorithm, with small constraint changes, handles practical tasks: trimming unusable footage, assembling stringouts, multi-camera live editing, and human-in-the-loop workflows.

Where Pith is reading between the lines

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

  • A causal, prefix-only model cannot represent editing idioms whose purpose is to set up a later shot (e.g., a point-of-view shot that only reads as POV because the next shot shows the subject looking); the paper's own limitation section acknowledges this, and it is a natural consequence of the greedy footage-constrained decoding.
  • Because FilmGPT never synthesizes frames, it cannot invent new action or hallucinate events outside the captured footage, which is a safety property for re-editing real video; the flip side is that it can only rearrange and re-time what was already shot.
  • The user study compares against two specific text-driven systems; extending the comparison to a broader set of editing tools and to human editors would test how much of the perceived quality is due to FilmGPT's learned grammar versus the inherent quality of the source footage.
  • The model's learned 'grammar' is an emergent property of the training distribution; if trained on a corpus with different conventions (e.g., mostly static surveillance footage), the resulting edits would presumably conform to that distribution, suggesting the grammar is learned rather than universal.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper introduces FilmGPT, a decoder-only autoregressive transformer trained on a large corpus of movies to model sequences of discrete visual tokens interleaved with an explicit ⟨CUT⟩ token. At inference, a footage-constrained decoding algorithm selects and trims shots from raw footage to assemble a montage. The authors claim state-of-the-art performance on the AVE shot-ordering benchmark (53.9% Acc@1 vs. 35.3% for UQNet) and report a user study in which participants prefer FilmGPT output over Transcript2Video (83.07%) and EditDuet (61.49%). They also demonstrate applications in trimming, stringouts, multi-camera editing, and human-in-the-loop editing, and argue that film-editing idioms such as establishing sequences, cutting on action, and point-of-view shots emerge from the learned model.

Significance. If the findings hold, this is a meaningful advance: it is among the first attempts to treat video montage as long-context next-token prediction over a discrete visual vocabulary, with an explicit cut token and a constrained decoding procedure. The scale of training data (6200 hours), the use of audio conditioning, and the external user-study signal are strengths, as is the internally consistent ablation sequence that isolates the contributions of the CUT token, cut-weighted loss, long context, and audio. However, the two headline claims—SOTA on AVE and user preference over prior systems—are currently under-evidenced: the AVE comparison changes the task protocol without re-baselining, the training/evaluation split is not stated, and the user-study winner is the full LLM+CLIP+FilmGPT pipeline rather than FilmGPT alone. The no-lookahead property of the causal model also limits the scope of the qualitative claims about emergent idioms. These issues are fixable with additional experiments and reframing, and the core idea remains promising.

major comments (4)
  1. [Sec. 4.2, Table 1] The AVE experiments never state whether the clips used for the shot-ordering evaluation are excluded from training. The text says the model is trained 'on AVE only' and then evaluated on 'AVE shot sequence ordering,' but no train/test split is specified. If the same scenes appear in both, the reported 53.9% Acc@1 may partly reflect memorization rather than learned editing structure. Please report the exact split (e.g., scene-disjoint or film-disjoint) and confirm no overlap between training and evaluation.
  2. [Sec. 4.2, Table 1, rows '+128k Context' onward] The comparison with UQNet is not controlled. Starting at the '+128k Context' row, the model receives the preceding shots as context, and the text explicitly states that 'previous approaches are unable to incorporate' this information. UQNet's 35.3% Acc@1 is obtained without such context. The claimed 18.6-point improvement therefore conflates model quality with a changed evaluation protocol. Please either evaluate UQNet under the same context-augmented protocol or report a context-free FilmGPT variant on the same task as the direct SOTA comparison.
  3. [Sec. 4.1, Fig. 5] The user study compares the full FilmGPT pipeline—LLM planner, CLIP-based top-50 footage preselection, and footage-constrained decoding—against Transcript2Video and EditDuet, which do not receive the same pre-filtered candidate set. The preference results therefore validate the complete system, not the contribution of FilmGPT's learned editing model. No ablation removes or randomizes FilmGPT while holding the LLM/CLIP preselection fixed. Please add such an ablation or explicitly reframe the claim as evaluating the full pipeline rather than FilmGPT alone.
  4. [Sec. 3.2 and Sec. 5.2, Fig. 7] Because decoding is strictly left-to-right and greedy, the model never conditions on future shots. Editing idioms whose identity depends on a subsequent shot—notably point-of-view shots—cannot be deliberately produced by this causal model; their appearance in the stringout examples is at best accidental. The paper should state this limitation explicitly in Sec. 6 and temper the qualitative claim that such idioms 'emerge' from the model. This is a scope limitation for the claimed cinematic grammar, not an error in the decoding algorithm itself.
minor comments (5)
  1. [Sec. 4.1, Fig. 5] The preference percentages are reported without confidence intervals. With 1242 data points and p<0.001 the result is strong, but the EditDuet comparison (61.49%) would benefit from a CI or a per-participant breakdown.
  2. [Sec. 4.2, Table 1] The Δ columns are not explicitly defined in the caption. State that they are relative to the immediately preceding row, or to the base model, to avoid ambiguity.
  3. [Sec. 3.2, Algorithm 1] The prose in Step 2 refers to 'highest per-token likelihood' while the pseudocode in FindCut minimizes NLL. Align the terminology (e.g., 'lowest per-token NLL').
  4. [Sec. 5.3] Typo in the first sentence: 'Mutli-camera' should be 'Multi-camera'.
  5. [Sec. 3.1.2] A one-sentence description of how the sliding-window attention is positioned across the 128k-token context and how positions are encoded would improve reproducibility, even if full details are in the supplement.

Circularity Check

0 steps flagged

No circular derivation: FilmGPT's outputs are evaluated against external human preference and a benchmark ordering task, not against its own training objective or fitted parameters.

full rationale

The paper's central claims are empirical. FilmGPT is trained with a standard autoregressive next-token objective (Eq. 1) and a cut-weighted loss (Eqs. 2-3); these do not encode the evaluation targets. The AVE shot-ordering task evaluates the model's ability to order held-out shots by likelihood, and the user study measures human preference for edited outputs. Neither reduces to the training loss by construction. The footage-constrained decoding algorithm is a constrained generation procedure that selects shots by model likelihood; it is not a fitted parameter renamed as a prediction. Self-citations to EditDuet are used as a baseline, as a source of the LLM planner, and as a user-study protocol reference, but they are not invoked as proof of the central result or to forbid alternatives, so they are not load-bearing. Potential concerns—no ablation of the LLM/CLIP preselection in the user study and no explicit statement of AVE train/test exclusion—are attribution and rigor issues, not circularity by construction, and should be addressed in correctness review. The paper's limitations section also acknowledges concrete failure modes (e.g., cutting on unrelated actions), which further indicates the evaluation is not tautological.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 1 invented entities

The paper's contribution is a modeling framework; it introduces no physical entities. Its free parameters are architectural/hyperparameter choices. The load-bearing assumption is that editing grammar is statistically learnable from discrete frame tokens.

free parameters (7)
  • tokens per frame = 32 (TiTok-L)
    Each frame encoded to 32 tokens from a 4096-vocab; determines temporal resolution of context (128k ≈ 4000 frames).
  • sliding window size = 16k tokens
    Self-attention window; with 128k context, chosen to balance cost and long-range context.
  • cut-weight w and decay window = w=32, decay over w tokens
    Empirically set; upweights cut tokens and following 32 tokens in CE loss.
  • audio drop probability = 0.4
    Drop audio conditioning with p=0.4 to prevent over-reliance.
  • CLIP top-k for LLM planner = 50
    User-study pipeline restricts video collection to top 50 CLIP matches from LLM-generated descriptions.
  • max shot extension = 10 seconds
    Beam search considers extensions up to 10s after anchor point.
  • beam width = k (not stated in main text)
    Beam width for constrained search; value not given in paper (supplementary).
axioms (6)
  • domain assumption Film editing grammar is learnable from sequences of frame tokens and cut tokens
    Sec. 1; the entire approach rests on this statistical learnability.
  • domain assumption Sliding-window attention over 16k tokens with 128k context adequately approximates full attention for editing decisions
    Sec. 3.1.2; cuts/context beyond 2.5 min are ignored.
  • domain assumption TiTok-L's 32-token reconstruction preserves the visual information needed for editing (motion, continuity, subject identity)
    Sec. 3.1.1; if tokenizer discards fine-grained cues, editing decisions suffer.
  • domain assumption A single ⟨CUT⟩ token adequately represents a cut event
    Sec. 3.1.1; cuts have no duration/type in this representation.
  • domain assumption The ground-truth shot ordering in AVE measures editing quality
    Sec. 4.2; used as benchmark.
  • standard math Autoregressive factorization is valid
    Chain rule of probability, Eq. (1).
invented entities (1)
  • ⟨CUT⟩ token no independent evidence
    purpose: discrete marker separating shots, making cut placement a first-class prediction target
    A modeling construct; no external falsifiable handle.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Autoregressive Modeling of Film with Applications in Video Montage." pith.science (2026). https://pith.science/paper/ZQ4STNFG

@misc{pith2026260714645,
  author       = {Pith},
  title        = {Pith review of: Autoregressive Modeling of Film with Applications in Video Montage},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZQ4STNFG}},
  note         = {Machine review of arXiv:2607.14645}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This work introduces FilmGPT, an autoregressive transformer designed to address the challenge of video montage--turning a collection of raw, "unwatchable" footage into coherent cinematic sequences. Inspired by language learning in modern LLMs, we train a long-context autoregressive transformer on a large corpus of movies. The aim is to implicitly capture the "grammar" of film directly from data rather than from hand-coded rules. Unlike other generative models, FilmGPT does not generate any new video frames. Instead, at inference time, we introduce a footage-constrained decoding algorithm to select the best next shot from the input raw footage according to the statistical patterns learned from films. We first evaluate these learned statistics directly by using the FilmGPT autoregressive model for next shot prediction on a standard benchmark of shot sequence ordering, outperforming the previous state of the art. We then evaluate our footage-constrained decoding algorithm on the full film editing task via a user study, and find that our FilmGPT-based editing significantly outperforms previous approaches. Finally, we demonstrate the applicability of FilmGPT to a wide range of applications in video montage, from automatic video segment trimming to human-in-the-loop film editing.

Figures

Figures reproduced from arXiv: 2607.14645 by Alexei A. Efros, Bryan Russell, Fabian Caba Heilbron, Greg Shakhnarovich, Josef Sivic, Marcelo Sandoval-Casta\~neda, Shiry Ginosar.

Figure 1
Figure 1. Figure 1: From raw footage to edited video with emergent idioms. Our system takes as input a collection of raw footage, typically hundreds of minutes. Our autoregressive model FilmGPT selects, trims and assembles videos from this collection into plausible sequences of shots. We find that this approach produces edited sequences where standard film editing idioms, such as establishing sequences, cutting on action, and… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our computational video montage framework. (Left) During training, FilmGPT from movies via next token prediction. Given an edited video, each frame is individually tokenized using TiTok-L [Yu et al. 2024], and an additional ⟨CUT⟩ token is added at the boundaries of each shot. For the purposes of this visualization, each frame is tokenized to one single token. In our actual implementation, each … view at source ↗
Figure 3
Figure 3. Figure 3: Audio conditioning via cross-attention. For the purposes of this visualization, each audio token corresponds to one second and each video token corresponds to 0.25 seconds. Each video token only cross-attends to the audio vectors corresponding to the past and to the current second in time. The ⟨CUT⟩ token is shown with red diagonal stripes. additionally condition our model on audio features from the edited… view at source ↗
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Participant preference results in our user study. We compare our model with T2V [Xiong et al. 2022] and EditDuet [Sandoval-Castañeda et al. 2025]. 𝑝 < 0.001 for all results. Our method outperforms competing methods in terms of participant preference. as in EditDuet (Llama-3.1-8B-Instruct [Dubey et al. 2024]). The LLM receives a JSON file containing high level descriptions of every video in the raw footage … view at source ↗
Figure 6
Figure 6. Figure 6: Samples from our video segment trimming example applica￾tion. Grayed out portions on the film strips are segments not selected by our method. The model correctly assigns low probability to segments that do not look cinematic and with erratic movements. Output trimmed videos are provided in the supplementary material. general setup of a scene. A low-level task in video editing is to find the segments in eac… view at source ↗
Figure 7
Figure 7. Figure 7: Samples from our stringouts example application. Our method convincingly assembles short sequences of shots that respect simple movie idioms. Output stringout videos are provided in the supplementary material. Thumbnails are copyrighted and belong to EditStock. edited draft for a short film. As the user edits, our method produces candidate sequences to continue the edit, and the human editor is able to cho… view at source ↗
Figure 8
Figure 8. Figure 8: Sample from our multi-camera editing example application. Grayed out portions on the film strips are segments not selected by our method. Footage comes from the project “Before You Accuse Me”, where a duo of musicians with guitars play in front of a number of cameras. Our method cuts between streams focusing on sources of audio and rhythm. The output edited video is in the supplementary material. Thumbnail… view at source ↗
Figure 9
Figure 9. Figure 9: Sample from our human-in-the-loop video editing example application. The fourth frame in the strip belongs to the A-roll. Note the ability of our method to stick to thematic cross-cutting across B-roll sequences, where shots of the oven fire are interleaved with the process of preparing dough. See the supplementary material for the video outputs. Thumbnails are copyrighted and belong to EditStock [PITH_FU… view at source ↗
Figure 10
Figure 10. Figure 10: Failure mode: cutting on unrelated actions. Consecutive frames from sequences assembled by our method for “The Ovens of Cappoquin” and “Built by Life”. On the left, a shot of hands grabbing dough is cut to a shot with hands placing dough in a loaf pan. The dough, the location, and the lighting are clearly different, making a jarring cut. On the right, a man working the field with a rake is cut to a man wa… 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

116 extracted references · 13 linked inside Pith

  1. [1]

    Leake, Mackenzie and Davis, Abe and Truong, Anh and Agrawala, Maneesh , journal=TOG, volume=

  2. [2]

    Workshop on Intelligent Camera Control, Cinematography and Editing , pages=

    Gandhi, Vineet and Ronfard, R. Workshop on Intelligent Camera Control, Cinematography and Editing , pages=. 2015 , organization=

    2015

  3. [3]

    ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) , volume=

    Wu, Hui-Yin and Pal. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) , volume=. 2018 , publisher=

    2018

  4. [4]

    Interactive Storytelling: Fourth International Conference on Interactive Digital Storytelling, ICIDS , pages=

    Lino, Christophe and Chollet, Mathieu and Christie, Marc and Ronfard, R. Interactive Storytelling: Fourth International Conference on Interactive Digital Storytelling, ICIDS , pages=. 2011 , organization=

    2011

  5. [5]

    Jhala, Arnav and Young, Robert Michael , booktitle=AAAI, volume=

  6. [6]

    1993 , organization=

    Karp, Peter and Feiner, Steven , booktitle=. 1993 , organization=

    1993

  7. [7]

    Christianson, David B and Anderson, Sean E and He, Li-wei and Salesin, David H and Weld, Daniel S and Cohen, Michael F , booktitle=AAAI, pages=

  8. [8]

    and Salesin, David H

    He, Li-wei and Cohen, Michael F. and Salesin, David H. , title =. 1996 , isbn =. doi:10.1145/237170.237259 , booktitle =

    arXiv 1996

  9. [9]

    Wu, Hui-Yin and Christie, Marc , booktitle=

  10. [10]

    2024 , pages =

    Argaw, Dawit Mureja and Soldan, Mattia and Pardo, Alejandro and Zhao, Chen and Heilbron, Fabian Caba and Chung, Joon Son and Ghanem, Bernard , title =. 2024 , pages =

    2024

  11. [11]

    arXiv preprint arXiv:1508.07593 , year=

    Ronfard, R. arXiv preprint arXiv:1508.07593 , year=

    Pith/arXiv arXiv

  12. [12]

    2020 , organization=

    Huang, Qingqiu and Xiong, Yu and Rao, Anyi and Wang, Jiaze and Lin, Dahua , booktitle=ECCV, pages=. 2020 , organization=

    2020

  13. [13]

    2022 , organization=

    Argaw, Dawit Mureja and Heilbron, Fabian Caba and Lee, Joon-Young and Woodson, Markus and Kweon, In So , booktitle=ECCV, pages=. 2022 , organization=

    2022

  14. [14]

    Rao, Anyi and Wang, Jiaze and Xu, Linning and Jiang, Xuekun and Huang, Qingqiu and Zhou, Bolei and Lin, Dahua , booktitle = ECCV, year=

  15. [15]

    Tapaswi, Makarand and Zhu, Yukun and Stiefelhagen, Rainer and Torralba, Antonio and Urtasun, Raquel and Fidler, Sanja , title =

  16. [16]

    Lei, Jie and Yu, Licheng and Bansal, Mohit and Berg, Tamara L , booktitle=EMNLP, year=

  17. [17]

    Ridouane Ghermi and Xi Wang and Vicky Kalogeiton and Ivan Laptev , journal=

  18. [18]

    Joseph Mascelli , title =

  19. [19]

    Daniel Arijon , title =

  20. [20]

    Roy Thompson and Christopher Bowen , title =

  21. [21]

    Christian Metz , title =

  22. [22]

    Frank Manchel , title =

  23. [23]

    1968 , publisher =

    Roy Knight , title =. 1968 , publisher =

    1968

  24. [24]

    Qihang Yu and Mark Weber and Xueqing Deng and Xiaohui Shen and Daniel Cremers and Liang-Chieh Chen , title =

  25. [25]

    Peters and Arman Cohan , journal=

    Iz Beltagy and Matthew E. Peters and Arman Cohan , journal=

  26. [26]

    Guangxuan Xiao and Yuandong Tian and Beidi Chen and Song Han and Mike Lewis , booktitle=ICLR, year=

  27. [27]

    arXiv preprint arXiv:2302.13971 , year=

    Touvron, Hugo and Lavril, Thibaut and Izacard, Gautier and Martinet, Xavier and Lachaux, Marie-Anne and Lacroix, Timoth. arXiv preprint arXiv:2302.13971 , year=

    Pith/arXiv arXiv

  28. [28]

    2303.08774 , archivePrefix=

    OpenAI , year=. 2303.08774 , archivePrefix=

    Pith/arXiv arXiv

  29. [29]

    Communications , number =

    Metz, Christian , doi =. Communications , number =. 1964 , Bdsk-Url-1 =

    1964

  30. [30]

    2016 , organization=

    Merabti, Billal and Christie, Marc and Bouatouch, Kadi , booktitle=CGF, volume=. 2016 , organization=

    2016

  31. [31]

    Learning Transferable Visual Models From Natural Language Supervision , author =

  32. [32]

    D\'idac Sur\'is and Sachit Menon and Carl Vondrick , booktitle=iccv, year=

  33. [33]

    Visual Programming: Compositional visual reasoning without training , author =

  34. [34]

    Manning and Peter Shaw and Mandar Joshi and Kenton Lee , booktitle=icml, year=

    Shikhar Murty and Christopher D. Manning and Peter Shaw and Mandar Joshi and Kenton Lee , booktitle=icml, year=

  35. [35]

    Gomez, Aidan and Cohere , year=2024, howpublished=

    2024

  36. [36]

    al , year=

    Abhimanyu Dubey and Abhinav Jauhri and et. al , year=. 2407.21783 , archivePrefix=

    Pith/arXiv arXiv

  37. [37]

    Xiong, Yu and Heilbron, Fabian Caba and Lin, Dahua , booktitle=acmmm, year=

  38. [38]

    Wang, Bryan and Li, Yuliang and Lv, Zhaoyang and Xia, Haijun and Xu, Yan and Sodhi, Raj , booktitle=

  39. [39]

    Madaan, Aman and Tandon, Niket and Gupta, Prakhar and Hallinan, Skyler and Gao, Luyu and Wiegreffe, Sarah and Alon, Uri and Dziri, Nouha and Prabhumoye, Shrimai and Yang, Yiming and others , journal=nips, volume=

  40. [40]

    Romera-Paredes, Bernardino and Barekatain, Mohammadamin and Novikov, Alexander and Balog, Matej and Kumar, M Pawan and Dupont, Emilien and Ruiz, Francisco JR and Ellenberg, Jordan S and Wang, Pengming and Fawzi, Omar and others , journal=

  41. [41]

    Chiquier, Mia and Mall, Utkarsh and Vondrick, Carl , journal=

  42. [42]

    Yang, John and Jimenez, Carlos E and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , journal=

  43. [43]

    Hu, Ziniu and Iscen, Ahmet and Jain, Aashi and Kipf, Thomas and Yue, Yisong and Ross, David A and Schmid, Cordelia and Fathi, Alireza , booktitle=icml, year=

  44. [44]

    Huang, Kaixuan and Qu, Yuanhao and Cousins, Henry and Johnson, William A and Yin, Di and Shah, Mihir and Zhou, Denny and Altman, Russ and Wang, Mengdi and Cong, Le , journal=

  45. [45]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , journal=

  46. [46]

    Kim, Joongwon and Paranjape, Bhargavi and Khot, Tushar and Hajishirzi, Hannaneh , journal=

  47. [47]

    VideoAgent: Long-form Video Understanding with Large Language Model as Agent , author=

  48. [48]

    Min, Juhong and Buch, Shyamal and Nagrani, Arsha and Cho, Minsu and Schmid, Cordelia , booktitle=cvpr, year=

  49. [49]

    and Zhang, Hao and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , title =. 2024 , publisher =

    2024

  50. [50]

    Ioannidis and Karthik Subbian and Jure Leskovec and James Zou , booktitle=NIPS, year=

    Shirley Wu and Shiyu Zhao and Qian Huang and Kexin Huang and Michihiro Yasunaga and Kaidi Cao and Vassilis N. Ioannidis and Karthik Subbian and Jure Leskovec and James Zou , booktitle=NIPS, year=. AvaTaR: Optimizing

  51. [51]

    2025 , organization=

    Video Question Answering with Procedural Programs , author=. 2025 , organization=

    2025

  52. [52]

    2025 , organization=

    V-Trans4Style: Visual Transition Recommendation for Video Production Style Adaptation , author=. 2025 , organization=

    2025

  53. [53]

    2022 , organization=

    Autotransition: Learning to recommend video transition effects , author=. 2022 , organization=

    2022

  54. [54]

    2025 , organization=

    VideoStudio: Generating Consistent-Content and Multi-scene Videos , author=. 2025 , organization=

    2025

  55. [55]

    2024 , booktitle =

    Han Lin and Abhay Zala and Jaemin Cho and Mohit Bansal , title =. 2024 , booktitle =

    2024

  56. [56]

    arXiv preprint arXiv:2411.04925 , year=

    StoryAgent: Customized Storytelling Video Generation via Multi-Agent Collaboration , author=. arXiv preprint arXiv:2411.04925 , year=

    Pith/arXiv arXiv

  57. [57]

    arXiv preprint arXiv:2405.01535 , year=

    Prometheus 2: An open source language model specialized in evaluating other language models , author=. arXiv preprint arXiv:2405.01535 , year=

    Pith/arXiv arXiv

  58. [58]

    Gonzalez and Ion Stoica , booktitle=NIPS, year=

    Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica , booktitle=NIPS, year=. Judging

  59. [59]

    arXiv preprint arXiv:2401.06591 , year=

    Prometheusvision: Vision-language model as a judge for fine-grained evaluation , author=. arXiv preprint arXiv:2401.06591 , year=

    Pith/arXiv arXiv

  60. [60]

    2023 , eprint=

    GPT-4V(ision) as a Generalist Evaluator for Vision-Language Tasks , author=. 2023 , eprint=

    2023

  61. [61]

    GPT-4V(ision) Technical Work and Authors , author=

  62. [62]

    Sarfraz, Saquib and Murray, Naila and Sharma, Vivek and Diba, Ali and Van Gool, Luc and Stiefelhagen, Rainer , title =

  63. [63]

    LLaVA-NeXT: A Strong Zero-shot Video Understanding Model , url=

    Zhang, Yuanhan and Li, Bo and Liu, haotian and Lee, Yong jae and Gui, Liangke and Fu, Di and Feng, Jiashi and Liu, Ziwei and Li, Chunyuan , month=. LLaVA-NeXT: A Strong Zero-shot Video Understanding Model , url=

  64. [64]

    arXiv preprint arXiv:2307.09702 , year=

    Efficient Guided Generation for LLMs , author=. arXiv preprint arXiv:2307.09702 , year=

    Pith/arXiv arXiv

  65. [65]

    2024 , eprint=

    Transcrib3D: 3D Referring Expression Resolution through Large Language Models , author=. 2024 , eprint=

    2024

  66. [66]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    Towards mitigating LLM hallucination via self reflection , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    2023

  67. [67]

    Adobe Premiere Pro , author=

  68. [68]

    Avid Media Composer , author=

  69. [69]

    2023 , publisher=

    Eventfulness for interactive video alignment , author=. 2023 , publisher=

    2023

  70. [70]

    2014 , publisher=

    Videosnapping: Interactive synchronization of multiple videos , author=. 2014 , publisher=

    2014

  71. [71]

    2014 , publisher=

    Automatic editing of footage from multiple social cameras , author=. 2014 , publisher=

    2014

  72. [72]

    2022 , publisher=

    PopStage: The Generation of Stage Cross-Editing Video based on Spatio-Temporal Matching , author=. 2022 , publisher=

    2022

  73. [73]

    2015 , publisher=

    Gaze-driven video re-editing , author=. 2015 , publisher=

    2015

  74. [74]

    2012 , publisher=

    Tools for placing cuts and transitions in interview video , author=. 2012 , publisher=

    2012

  75. [75]

    Iberian Conference on Pattern Recognition and Image Analysis , pages=

    Film Shot Type Classification Based on Camera Movement Styles , author=. Iberian Conference on Pattern Recognition and Image Analysis , pages=. 2022 , organization=

    2022

  76. [76]

    arXiv preprint arXiv:1308.0850 , year=

    Generating sequences with recurrent neural networks , author=. arXiv preprint arXiv:1308.0850 , year=

    Pith/arXiv arXiv

  77. [77]

    2016 , organization=

    Ssd: Single shot multibox detector , author=. 2016 , organization=

    2016

  78. [78]

    2020 , organization=

    A unified framework for shot type classification based on subject centric lens , author=. 2020 , organization=

    2020

  79. [79]

    Searching for mobilenetv3 , author=

  80. [80]

    arXiv preprint arXiv:2411.12293 , year=

    Generative Timelines for Instructed Visual Assembly , author=. arXiv preprint arXiv:2411.12293 , year=

    Pith/arXiv arXiv

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.