Cut2Next: Generating Next Shot via In-Context Tuning

Hongbo Liu; Jiajun Li; Jingwen He; Wanli Ouyang; Yu Qiao; Ziqi Huang; Ziwei Liu

REVIEW 4 major objections 5 minor 56 references

Cut2Next: Generating Next Shot via In-Context Tuning

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Cut2Next generates the next shot from one input frame while obeying cinematic edit patterns.

desk verdict A new task and a plausible method, but the benchmark's provenance and train/test disjointness are undocumented—checkable and central. read the letter →

arxiv 2508.08244 v2 pith:KUK7VZIM submitted 2025-08-11 cs.CV cs.AI

Jingwen He , Hongbo Liu , Jiajun Li , Ziqi Huang , Yu Qiao , Wanli Ouyang , Ziwei Liu This is my paper · ORCID

classification cs.CVcs.AI

keywords nextshotgenerationcinematiccontinuityeditingpatternsshot/reversecutawaycut-in/cut-outdiffusiontransformerin-contexttuning

verification ladder T0 review T1 audit T2 compute T3 formal

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's target is Next Shot Generation: given one film shot, produce the next shot that looks like a real editor cut to it, keeping characters and environment continuous while obeying a specified editing pattern such as shot/reverse shot, cut-in/out, cutaway, or multi-angle. The authors argue that current multi-shot generators optimize visual similarity but ignore the narrative grammar of the cut, and they propose Cut2Next, a diffusion-transformer model fine-tuned with LoRA that takes in a hierarchical prompt: a relational prompt describing the inter-shot connection and individual prompts describing each shot. Their experiments on the new CutBench benchmark and user studies claim that Cut2Next beats an adapted in-context LoRA baseline on visual consistency, text fidelity, and human preference for cinematic continuity and editing adherence. If true, this gives storyboard and video tools a controllable next-shot generator for narrative sequences rather than merely consistent single frames.

What carries the argument

The central mechanism is hierarchical prompt conditioning inside a diffusion transformer. A relational prompt encodes the overall scene, narrative interpretation, and edit type; two individual prompts describe the input shot and target shot's content and cinematography. CACI makes the transformer aware of token segment roles by giving distinct AdaLN-Zero conditions to clean visual tokens, noisy target tokens, relational text, and individual text. HAM is a fixed binary attention mask that keeps individual prompts isolated with their own visual segment and lets the relational prompt attend to both visual segments, preventing textual cross-talk. LoRA fine-tuning (rank 256) adapts the pretrained model to combine these signals; CACI and HAM themselves add no parameters to the base model.

What would settle it

Take a random sample of CutBench pairs, have a professional editor label the true editing pattern without seeing the automated annotations, measure agreement with the automated labels, and re-run the benchmark on only the pairs where human and automated labels agree; if Cut2Next's edge over the baseline shrinks or vanishes, the claim that it follows cinematic editing patterns is not established.

Watch

Extended reading notes

Core claim

Cut2Next is a Diffusion Transformer (built on FLUX.1-dev) that, after LoRA fine-tuning on adjacent shot pairs, generates a target shot conditioned on a clean input-shot latent plus hierarchical text. Context-Aware Condition Injection (CACI) applies separate AdaLN-Zero conditioning to clean input tokens, noisy target tokens, and the relational and individual prompt tokens; Hierarchical Attention Mask (HAM) then restricts cross-attention so each individual prompt only modulates its own visual segment while the relational prompt can bridge both visual segments. With these two parameter-free modifications, the paper claims, Cut2Next maintains character identity, environment, lighting, and color while executing the requested edit pattern. On CutBench, Cut2Next reports higher DINO and CLIP-I similarity, higher CLIP-T fidelity, and lower FID than the adapted IC-LoRA-Cond baseline, and a user study of 15 raters preferred Cut2Next for both cinematic continuity and adherence to the intended edit.

Load-bearing premise

The load-bearing premise is that the automated annotations of editing pattern and continuity attributes are correct, because the same annotation style is used both to train Cut2Next and to score it in CutBench, and the paper does not report human validation of those labels.

Editorial extensions

If this is right

Storyboard artists can specify a cut type such as reverse shot, cutaway, cut-in/out, or multi-angle and receive a next frame that follows that grammar while preserving characters and setting.
Because CACI and HAM work on top of a pretrained DiT without adding parameters, the conditioning pattern can be transplanted to other in-context generation settings that mix clean reference tokens with noisy target tokens.
Two-stage training on a large raw shot-pair corpus followed by a small curated set gives a practical data recipe: broad visual transitions first, professional edit patterns second.
The human preference results indicate that next-shot quality is judged not only by embedding similarity but by edit-pattern fidelity and cinematic continuity, which automatic similarity metrics alone do not capture.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

My inference: the relational/individual prompt split is a general recipe for controllable conditional generation, applicable beyond film shots to any task where one input provides stable context and another is being synthesized.
My inference: a direct test of whether Cut2Next actually understands cut grammar would be to condition it on the relational prompt alone, without the target individual prompt, and ask raters whether the edit type still comes through.
My inference: since the benchmark's editing-pattern labels come from the same automated annotator used in training, a human-validated subset of CutBench would separate genuine cinematic understanding from annotation-following.
My inference: the same masked-attention design could be extended to video by treating several clean context frames as the stable visual side and one future frame as the noisy target side.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Cut2Next is a solid empirical paper: it defines a new task (Next Shot Generation), builds two training datasets and a benchmark, and proposes parameter-free conditioning mechanisms (CACI and HAM) that fit the problem cleanly. The writing is clear, the ablations support the method's design, and the limitations are stated honestly.

The genuinely new bits are the task formulation itself—synthesizing a next shot that follows professional editing patterns—and the hierarchical prompting scheme with relational and individual prompts. The CACI and HAM mechanisms are simple and well-motivated: separating clean conditional latents from noisy target latents at the AdaLN level makes sense, and the attention mask prevents cross-talk between textual prompts. Nothing here is paradigm-shifting, but it is a useful step and the benchmark could become a reference point for follow-up work.

Now the soft spots. The stress-test note is right: the paper never says where CutBench images come from or how they are disjoint from RawCuts/CuratedCuts. Since RawCuts is built from MovieNet pairs and CutBench is described only as 'diverse movie shot images,' a reader cannot rule out overlap. The FID comparison against ground-truth targets is exactly where memorized pairs would inflate scores. This is a checkable gap, not a proven flaw—the authors may have used a disjoint set—but it is load-bearing and must be addressed.

Second, the evaluation leans on a single adapted baseline (IC-LoRA-Cond). Automatic metrics come without confidence intervals or significance tests, and the user study, while reporting strong preference, has fifteen participants and no inferential statistics. These are not fatal, but they cap the strength of the claims.

Third, the annotation pipeline uses Gemini-2.0-flash for both training and evaluation prompts. The paper does not report human validation of those editing-pattern labels. If the labels are noisy, the model may be learning annotation language rather than cinematic grammar. The authors do acknowledge some of this in the limitations, but not the benchmark-provenance part.

Who should read this: anyone working on multi-shot or narrative video generation, storyboarding tools, or editing-conditioned synthesis. It deserves a serious referee—the task is new and the benchmark is likely to be reused—but the authors should be asked to clarify CutBench construction, report train/test disjointness, add at least one more baseline, and run significance tests.

My recommendation: accept for peer review with major revision focus on the evaluation evidence.

Referee Report

4 major / 5 minor

Summary. The paper introduces Next Shot Generation (NSG), a task in which a model must synthesize a subsequent cinematic shot from a given input shot while preserving character, environment, lighting, and other continuity attributes and while conforming to a specified editing pattern (shot/reverse shot, cut-in/out, cutaway, multi-angle). The authors propose Cut2Next, built on FLUX.1-dev with LoRA fine-tuning and two parameter-free architectural additions: Context-Aware Condition Injection (CACI), which assigns timestep- and role-specific AdaLN conditioning to different token segments, and a Hierarchical Attention Mask (HAM), which restricts attention between relational prompts, per-shot individual prompts, and visual tokens. They construct two datasets, RawCuts and CuratedCuts, from MovieNet with automatic Gemini-2.0-flash hierarchical annotations, and introduce a benchmark, CutBench. Experiments compare Cut2Next against an adapted IC-LoRA-Cond baseline on DINO similarity, CLIP-I similarity, CLIP-T fidelity, and FID, and report a user study showing strong preference for Cut2Next in cinematic continuity and editing-pattern adherence.

Significance. If the empirical claims are sustained, the paper makes a useful contribution: NSG is a well-motivated task that bridges single-shot image generation and narrative multi-shot video generation, and the proposed CACI/HAM mechanisms are simple, parameter-free modifications that could transfer to other DiT-based in-context generation systems. The release of RawCuts, CuratedCuts, and CutBench would also be a valuable community asset, and the inclusion of a perceptual user study is a strength. However, the central generalization claim is currently supported by a single adapted baseline and by automatic metrics whose statistical reliability and benchmark provenance are not fully documented. These are fixable evidence gaps rather than fundamental flaws, but they need to be addressed before the paper's conclusions can be accepted.

major comments (4)

[Sections 3.1.1 and 4.1 (CutBench provenance and train/test disjointness)] CutBench is described only as 'hundreds of diverse movie shot images with our hierarchical prompts,' with no statement of where these images come from or whether they are disjoint from the RawCuts/CuratedCuts training data. Since RawCuts is built from MovieNet by pairing adjacent keyframes and CuratedCuts is a subset of RawCuts, it is critical to rule out near-duplicate overlap between CutBench and the training pairs; if overlap exists, the large FID gap in Table 1 (59.37 vs 80.43) and the user-study preferences could reflect memorization rather than next-shot generalization. Please document CutBench's source, construction, and explicit disjointness procedure, and report FID against a properly defined reference set of held-out ground-truth shots.
[Section 3.1.2 (Validity of automated annotations)] Both training and evaluation rely on Gemini-2.0-flash annotations for editing-pattern labels and continuity attributes, but the paper provides no human validation of these annotations. If the automated labels are noisy or systematically misidentify cut types, the model may be learning to follow annotation language rather than genuine cinematic continuity, and the CLIP-T fidelity metric (computed against prompts produced by the same pipeline) would be biased in the model's favor. Please add a human-annotation agreement study on a random subset of RawCuts/CuratedCuts and CutBench, reporting per-label accuracy or agreement for the editing-pattern and continuity attributes.
[Section 4.1 (Baseline comparison)] The only comparison baseline is IC-LoRA-Cond, an adaptation of a single prior method. Given the existence of other conditional multi-shot and in-context generation methods (e.g., CineVerse, SynCamMaster, and more recent IC-LoRA variants), the claim that Cut2Next 'significantly outperforms baselines' is not yet well supported. Please add at least one additional strong baseline adapted to NSG, or provide a principled justification for why no other baseline can be adapted.
[Sections 4.2-4.4 and Tables 1-4 (Statistical reliability)] All automatic metrics are reported as point estimates without confidence intervals or significance tests. The differences in DINO (0.4952 vs 0.4669), CLIP-I (0.7298 vs 0.7152), and CLIP-T (0.2979 vs 0.2805) are small and may be within sampling noise; the user-study preference rates in Table 4 are also presented without confidence intervals even though Section 4.4 states that the process was repeated four times to compute standard deviation. Please report bootstrap confidence intervals for each metric and a significance test (e.g., paired test over CutBench samples) for the Cut2Next-versus-baseline comparisons and for the ablation differences in Tables 2 and 3.

minor comments (5)

[Section 4.4] The phrase 'CutBench test split' is used but no test split is defined in Section 4.1; please specify how CutBench is partitioned and how the 100 user-study samples were selected.
[Section 3.1.1] Dataset sizes are reported imprecisely ('over 200k shot pairs' for RawCuts, 'thousands' for CuratedCuts); please state the exact numbers of pairs and the exact number of CutBench samples.
[Section 3.2.1, Eq. (1)] The notation for the concatenated input sequence does not make the segment boundaries explicit; please define the ordering and lengths of the text and visual token segments in the input sequence.
[Section 4.3, Figure 7] The claim that CACI variants 'converged similarly' is based only on training loss on RawCuts; please report final validation loss or a downstream metric, and state the number of training steps shown.
[Various] Several LaTeX macros appear to be broken in the compiled text (e.g., Section 3.2.2 and Table headers), producing garbled formulas; these should be fixed so that the architectural descriptions and metric names are readable.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity; the central next-shot generation claim is evaluated on held-out prompts and images, with only a minor non-load-bearing self-citation to the authors' ShotBench annotation schema.

other [Section 3.1.2, Hierarchical Prompt Annotation, Individual Prompts (P_shot)]
"P_shot offers a detailed description of its visual content and cinematographic characteristics [29]. This comprises: (1) a concise summary (subject, setting), a detailed description of visual content (appearance, posture, expression, costume, background); and (2) a structured set of cinematography attributes (e.g., shot size/framing, composition, camera angle, focal length)."

This is the one author-overlapping citation in the method chain: ref [29] (ShotBench) is prior work by several of the same authors, and it supplies the cinematography-attribute taxonomy used in the Individual Prompts. It is not a load-bearing circular step, because the taxonomy is an annotation convention rather than a predicted result; Cut2Next must still map held-out condition images and prompts to new target images, and the reported metrics compare against ground-truth targets and human judgments rather than reducing to the schema. It lowers the score only under the 'minor self-citation' category, not as a derivation that equals its own input.

full rationale

The paper's derivation chain is largely self-contained. Cut2Next is a LoRA fine-tune of FLUX.1-dev trained on RawCuts/CuratedCuts adjacent-shot pairs, and its central claims are tested on CutBench against an adapted IC-LoRA-Cond baseline plus human preference judgments. I found no equation or procedure that reduces a predicted quantity to a fitted input: the hierarchical prompts are generated by the same Gemini-2.0-flash pipeline for training and evaluation, but this is standard in-distribution evaluation rather than circularity, because the baseline receives the same prompts and the model can fail to follow them, with text fidelity measured by CLIP similarity to the target prompt and visual metrics measured against ground-truth targets. The only self-referential element is citation [29] (ShotBench), an author-overlapping prior work that supplies the cinematography-attribute schema for Individual Prompts; since this schema is a labeling convention and not the target of prediction, it does not force the results. The skeptical concern that CutBench may overlap with MovieNet-derived training pairs is a test-leakage and reproducibility risk, not a demonstrated circularity: the paper does not state CutBench's source or its disjointness from RawCuts/CuratedCuts, so the specific reduction required by the rules cannot be exhibited. Accordingly, the appropriate finding is no significant circularity aside from one minor non-load-bearing self-citation.

Assumptions & free parameters 0 free parameters · 4 assumptions · 3 invented entities

The method depends on the quality of the base FLUX model, automated text annotations, filtered movie data, and chosen metrics. These are domain assumptions rather than standard math axioms; the central claim is empirical.

assumptions (4)

domain assumption FLUX.1-dev is a capable text-to-image DiT baseline whose in-context generation ability can be repurposed for conditioned next-shot generation.
The method is built entirely on FLUX.1-dev (Section 3.2); if the base model could not handle clean-latent conditioning, the approach would not work.
domain assumption Gemini-2.0-flash annotations of shot pairs correctly capture the editing pattern and continuity attributes used for training and evaluation.
Section 3.1.2 uses automated annotations without human validation; the user study and CutBench assume these labels define the intended editing pattern.
domain assumption Low-motion keyframes sampled from MovieNet with aesthetic filtering constitute a valid distribution for learning cinematic transitions.
Section 3.1.1 filters shots by motion and aesthetics; this removes action sequences, and the resulting distribution is assumed sufficient to learn NSG.
domain assumption The evaluation metrics (DINO, CLIP-I, CLIP-T, FID) and user preference reflect cinematic continuity and editing-pattern adherence.
Sections 4.1 and 4.4 treat these metrics as proxies for the claimed cinematic quality; no calibration of metrics against expert ratings is provided.

invented entities (3)

CACI (Context-Aware Condition Injection)
purpose: Token-type-aware AdaLN conditioning for clean vs noisy latents and text prompts.
Validated only through internal ablations (Figure 7); no external replication or formal analysis.
HAM (Hierarchical Attention Mask)
purpose: Fixed binary attention mask guiding information flow between prompt and visual token segments.
Evaluated only within this paper; no independent test of the mask's necessity.
RawCuts, CuratedCuts, CutBench datasets
purpose: Training and evaluation resources for next-shot generation.
Not released in the manuscript; cannot be independently verified or reused.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cut2Next: Generating Next Shot via In-Context Tuning." pith.science (2026). https://pith.science/paper/KUK7VZIM

@misc{pith2026250808244,
  author       = {Pith},
  title        = {Pith review of: Cut2Next: Generating Next Shot via In-Context Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KUK7VZIM}},
  note         = {Machine review of arXiv:2508.08244}
}

read the original abstract

Effective multi-shot generation demands purposeful, film-like transitions and strict cinematic continuity. Current methods, however, often prioritize basic visual consistency, neglecting crucial editing patterns (e.g., shot/reverse shot, cutaways) that drive narrative flow for compelling storytelling. This yields outputs that may be visually coherent but lack narrative sophistication and true cinematic integrity. To bridge this, we introduce Next Shot Generation (NSG): synthesizing a subsequent, high-quality shot that critically conforms to professional editing patterns while upholding rigorous cinematic continuity. Our framework, Cut2Next, leverages a Diffusion Transformer (DiT). It employs in-context tuning guided by a novel Hierarchical Multi-Prompting strategy. This strategy uses Relational Prompts to define overall context and inter-shot editing styles. Individual Prompts then specify per-shot content and cinematographic attributes. Together, these guide Cut2Next to generate cinematically appropriate next shots. Architectural innovations, Context-Aware Condition Injection (CACI) and Hierarchical Attention Mask (HAM), further integrate these diverse signals without introducing new parameters. We construct RawCuts (large-scale) and CuratedCuts (refined) datasets, both with hierarchical prompts, and introduce CutBench for evaluation. Experiments show Cut2Next excels in visual consistency and text fidelity. Crucially, user studies reveal a strong preference for Cut2Next, particularly for its adherence to intended editing patterns and overall cinematic continuity, validating its ability to generate high-quality, narratively expressive, and cinematically coherent subsequent shots.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 52 canonical work pages

[1]

Accessed December 9, 2024 [Online]https://klingai.kuaishou.com/, 2024

Kling. Accessed December 9, 2024 [Online]https://klingai.kuaishou.com/, 2024

work page 2024
[2]

Accessed February 15, 2024 [Online]https://sora.com/library, 2024

Sora. Accessed February 15, 2024 [Online]https://sora.com/library, 2024

work page 2024
[3]

Multi-shot character consistency for text-to-video generation.�� , 2024

Yuval Atzmon, Rinon Gal, Yoad Tewel, Yoni Kasten, and Gal Chechik. Multi-shot character consistency for text-to-video generation.�� , 2024

work page 2024
[4]

Syncammaster: Synchronizing multi-camera video generation from diverse viewpoints.�� , 2024

Jianhong Bai, Menghan Xia, Xintao Wang, Ziyang Yuan, Xiao Fu, Zuozhu Liu, Haoji Hu, Pengfei Wan, and Di Zhang. Syncammaster: Synchronizing multi-camera video generation from diverse viewpoints.�� , 2024. 11

work page 2024
[5]

Talc: Time-aligned captions for multi-scene text-to-video generation.�� , 2024

Hritik Bansal, Yonatan Bitton, Michal Yarom, Idan Szpektor, Aditya Grover, and Kai-Wei Chang. Talc: Time-aligned captions for multi-scene text-to-video generation.�� , 2024

work page 2024
[6]

Lumiere: A space-time diffusion model for video generation

Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A space-time diffusion model for video generation. In�� , pages 1–11, 2024

work page 2024
[7]

Flux: Official inference repository for flux.1 models, 2024

Black Forest Labs. Flux: Official inference repository for flux.1 models, 2024. Accessed: 2024-11-12

2024
[8]

Stable video diffusion: Scaling latent video diffusion models to large datasets.�� , 2023

Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.�� , 2023

work page 2023

Show all 56 references

[9]

Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing

Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In �� , pages 22560–22570, 2023

2023
[10]

Skyreels-v2: Infinite-length film generative model.�� , 2025

Guibin Chen, Dixuan Lin, Jiangping Yang, Chunze Lin, Juncheng Zhu, Mingyuan Fan, Hao Zhang, Sheng Chen, Zheng Chen, Chengchen Ma, et al. Skyreels-v2: Infinite-length film generative model.�� , 2025

2025
[11]

FFmpeg.https://ffmpeg.org

FFmpeg Developers. FFmpeg.https://ffmpeg.org
[12]

An image is worth one word: Personalizing text-to-image generation using textual inversion.�� , 2022

Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion.�� , 2022

2022
[13]

Lcm-lookahead for encoder-based text-to-image personalization

Rinon Gal, Or Lichter, Elad Richardson, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Lcm-lookahead for encoder-based text-to-image personalization. In�� , pages 322–340. Springer, 2024

2024
[14]

Gemini: a family of highly capable multimodal models.�� , 2023

Gemini Google, Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.�� , 2023

2023
[15]

Long context tuning for video generation, 2025

Yuwei Guo, Ceyuan Yang, Ziyan Yang, Zhibei Ma, Zhijie Lin, Zhenheng Yang, Dahua Lin, and Lu Jiang. Long context tuning for video generation, 2025

2025
[16]

Id-animator: Zero-shot identity-preserving human video generation.�� , 2024

Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation.�� , 2024

2024
[17]

Lora: Low-rank adaptation of large language models.��, 1(2):3, 2022

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.��, 1(2):3, 2022

2022
[18]

In-context lora for diffusion transformers.�� , 2024

Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers.�� , 2024

2024
[19]

Movienet: A holistic dataset for movie understanding

Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In�� , pages 709–727. Springer, 2020

2020
[20]

Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning.�� , 2025

Yuzhou Huang, Ziyang Yuan, Quande Liu, Qiulin Wang, Xintao Wang, Ruimao Zhang, Pengfei Wan, Di Zhang, and Kun Gai. Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning.�� , 2025

2025
[21]

EasyOCR.https://github.com/JaidedAI/EasyOCR, 2024

Jaided AI. EasyOCR.https://github.com/JaidedAI/EasyOCR, 2024. 12

2024
[22]

Videobooth: Diffusion-based video generation with image prompts

Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In �� , pages 6689–6700, 2024

2024
[23]

Shotadapter: Text-to-multi-shot video generation with diffusion models.�� , 2025

Ozgur Kara, Krishna Kumar Singh, Feng Liu, Duygu Ceylan, James M Rehg, and Tobias Hinz. Shotadapter: Text-to-multi-shot video generation with diffusion models.�� , 2025

2025
[24]

Musiq: Multi-scale image quality transformer

Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In�� , pages 5148–5157, 2021

2021
[25]

Adam: A method for stochastic optimization.�� , 2014

Diederik P Kingma. Adam: A method for stochastic optimization.�� , 2014

2014
[26]

Hunyuanvideo: A systematic framework for large video generative models.�� , 2024

Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.�� , 2024

2024
[27]

aesthetic-predictor

LAION-AI. aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor , 2022

2022
[28]

Visualcloze: A universal image generation framework via visual in-context learning.�� , 2025

Zhong-Yu Li, Ruoyi Du, Juncheng Yan, Le Zhuo, Zhen Li, Peng Gao, Zhanyu Ma, and Ming- Ming Cheng. Visualcloze: A universal image generation framework via visual in-context learning.�� , 2025

2025
[29]

Shotbench: Expert-level cinematic understanding in vision-language models.�� , 2025

Hongbo Liu, Jingwen He, Yi Jin, Dian Zheng, Yuhao Dong, Fan Zhang, Ziqi Huang, Yinan He, Yangguang Li, Weichao Chen, et al. Shotbench: Expert-level cinematic understanding in vision-language models.�� , 2025

2025
[30]

One-prompt-one-story: Free-lunch consistent text-to- image generation using a single prompt.�� , 2025

Tao Liu, Kai Wang, Senmao Li, Joost van de Weijer, Fahad Shahbaz Khan, Shiqi Yang, Yaxing Wang, Jian Yang, and Ming-Ming Cheng. One-prompt-one-story: Free-lunch consistent text-to- image generation using a single prompt.�� , 2025

2025
[31]

Videostudio: Generating consistent-content and multi-scene videos

Fuchen Long, Zhaofan Qiu, Ting Yao, and Tao Mei. Videostudio: Generating consistent-content and multi-scene videos. In�� , pages 468–485. Springer, 2024

2024
[32]

Ace++: Instruction-based image creation and editing via context-aware content filling.�� , 2025

Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. Ace++: Instruction-based image creation and editing via context-aware content filling.�� , 2025

2025
[33]

Dinov2: Learning robust visual features without supervision.�� , 2023

Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.�� , 2023

2023
[34]

Moviecuts: A new dataset and benchmark for cut type recognition

Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcázar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In�� , pages 668–685. Springer, 2022

2022
[35]

Scalable diffusion models with transformers

William Peebles and Saining Xie. Scalable diffusion models with transformers. In�� , pages 4195–4205, 2023

2023
[36]

Cineverse: Consistent keyframe synthesis for cinematic scene composition.�� , 2025

Quynh Phung, Long Mai, Fabian David Caba Heilbron, Feng Liu, Jia-Bin Huang, and Cusuh Ham. Cineverse: Consistent keyframe synthesis for cinematic scene composition.�� , 2025

2025
[37]

Mask �dit: Dual mask-based diffusion transformer for multi-scene long video generation, 2025

Tianhao Qi, Jianlong Yuan, Wanquan Feng, Shancheng Fang, Jiawei Liu, SiYu Zhou, Qian He, Hongtao Xie, and Yongdong Zhang. Mask �dit: Dual mask-based diffusion transformer for multi-scene long video generation, 2025. 13

2025
[38]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In�� , pag...

2021
[39]

Exploring the limits of transfer learning with a unified text-to-text transformer.�� , 21(140):1–67, 2020

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.�� , 21(140):1–67, 2020

2020
[40]

Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In �� , pages 22500...

2023
[41]

Transnet v2: An effective deep network architecture for fast shot transition detection.�� , 2020

Tomáš Souˇcek and Jakub Lokoˇc. Transnet v2: An effective deep network architecture for fast shot transition detection.�� , 2020

2020
[42]

Ominicontrol: Minimal and universal control for diffusion transformer.�� , 2024

Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and universal control for diffusion transformer.�� , 2024

2024
[43]

Training-free consistent text-to-image generation.�� , 43(4):1–18, 2024

Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation.�� , 43(4):1–18, 2024

2024
[44]

NSFW Image Detection (Large)

TostAI. NSFW Image Detection (Large). https://huggingface.co/TostAI/ nsfw-image-detection-large, 2024

2024
[45]

Storyanchors: Generating consistent multi-scene story frames for long-form narratives.�� , 2025

Bo Wang, Haoyang Huang, Zhiyin Lu, Fengyuan Liu, Guoqing Ma, Jianlong Yuan, Yuan Zhang, and Nan Duan. Storyanchors: Generating consistent multi-scene story frames for long-form narratives.�� , 2025

2025
[46]

Dreamvideo: Composing your dream videos with cus- tomized subject and motion

Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with cus- tomized subject and motion. In�� ...

2024
[47]

Less-to- more generalization: Unlocking more controllability by in-context generation.�� , 2025

Shaojin Wu, Mengqi Huang, Wenxu Wu, Yufeng Cheng, Fei Ding, and Qian He. Less-to- more generalization: Unlocking more controllability by in-context generation.�� , 2025

2025
[48]

Videoauteur: Towards long narrative video generation.�� , 2025

Junfei Xiao, Feng Cheng, Lu Qi, Liangke Gui, Jiepeng Cen, Zhibei Ma, Alan Yuille, and Lu Jiang. Videoauteur: Towards long narrative video generation.�� , 2025

2025
[49]

Dynamicrafter: Animating open-domain images with video diffusion priors

Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamicrafter: Animating open-domain images with video diffusion priors. In�� , pages 399–417. Springer, 2024

2024
[50]

Cogvideox: Text-to-video diffusion models with an expert transformer.�� , 2024

Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.�� , 2024

2024
[51]

Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.�� , 2023

Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.�� , 2023

2023
[52]

Flashface: Human image personalization with high-fidelity identity preservation.�� , 2024

Shilong Zhang, Lianghua Huang, Xi Chen, Yifei Zhang, Zhi-Fan Wu, Yutong Feng, Wei Wang, Yujun Shen, Yu Liu, and Ping Luo. Flashface: Human image personalization with high-fidelity identity preservation.�� , 2024

2024
[53]

I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models.�� , 2023

Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models.�� , 2023. 14

2023
[54]

Moviedreamer: Hierarchical generation for coherent long visual sequence

Canyu Zhao, Mingyu Liu, Wen Wang, Weihua Chen, Fan Wang, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. �� , 2024

2024
[55]

Videogen-of-thought: A collaborative framework for multi-shot video generation.�� , 2024

Mingzhe Zheng, Yongqi Xu, Haojian Huang, Xuran Ma, Yexin Liu, Wenjie Shu, Yatian Pang, Feilong Tang, Qifeng Chen, Harry Yang, et al. Videogen-of-thought: A collaborative framework for multi-shot video generation.�� , 2024

2024
[56]

Storydiffusion: Consistent self-attention for long-range image and video generation.�� , 37:110315–110340, 2024

Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation.�� , 37:110315–110340, 2024. ��...

2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Accessed December 9, 2024 [Online]https://klingai.kuaishou.com/, 2024

Kling. Accessed December 9, 2024 [Online]https://klingai.kuaishou.com/, 2024

work page 2024

[2] [2]

Accessed February 15, 2024 [Online]https://sora.com/library, 2024

Sora. Accessed February 15, 2024 [Online]https://sora.com/library, 2024

work page 2024

[3] [3]

Multi-shot character consistency for text-to-video generation.�� , 2024

Yuval Atzmon, Rinon Gal, Yoad Tewel, Yoni Kasten, and Gal Chechik. Multi-shot character consistency for text-to-video generation.�� , 2024

work page 2024

[4] [4]

Syncammaster: Synchronizing multi-camera video generation from diverse viewpoints.�� , 2024

Jianhong Bai, Menghan Xia, Xintao Wang, Ziyang Yuan, Xiao Fu, Zuozhu Liu, Haoji Hu, Pengfei Wan, and Di Zhang. Syncammaster: Synchronizing multi-camera video generation from diverse viewpoints.�� , 2024. 11

work page 2024

[5] [5]

Talc: Time-aligned captions for multi-scene text-to-video generation.�� , 2024

Hritik Bansal, Yonatan Bitton, Michal Yarom, Idan Szpektor, Aditya Grover, and Kai-Wei Chang. Talc: Time-aligned captions for multi-scene text-to-video generation.�� , 2024

work page 2024

[6] [6]

Lumiere: A space-time diffusion model for video generation

Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A space-time diffusion model for video generation. In�� , pages 1–11, 2024

work page 2024

[7] [7]

Flux: Official inference repository for flux.1 models, 2024

Black Forest Labs. Flux: Official inference repository for flux.1 models, 2024. Accessed: 2024-11-12

2024

[8] [8]

Stable video diffusion: Scaling latent video diffusion models to large datasets.�� , 2023

Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.�� , 2023

work page 2023