acceptodds
Withdrawn from ICLR 2027

Film Style Transfer

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In film post-production, a colorist needs to apply the visual look of a whole film — its film look — to target frames. The standard approach is manual colour grading, accurate but labour-intensive. We propose to automate this step by introducing a novel computer vision task, termed film style transfer (FST): a set of reference frames sampled from a film represents its film look, which needs to be transferred to target frames. One intuitive solution is to leverage off-the-shelf style transfer methods. However, this suffers several issues: (1) Existing methods often assume a single style image, whereas FST is defined over a set of reference frames; (2) applying the film look uniformly over target frames tends to degrade their fine structure, rendering such direct application suboptimal. To address these, we introduce a dedicated FST method, AnswerPrint, that starts with film look extraction, followed by a coarse-to-fine transformation: first transferring the global film look, then recovering the fine content details the global transfer destroys, echoing the spirit of the established colorist workflow. Our method integrates with existing image generative models (e.g., Stable Diffusion) without extra training. To enable evaluation, we build Reel-30, a new FST benchmark with thirty films as well as metrics in film identity (source-film recognisability) and structural fidelity (local structure preservation). Extensive experiments show that AnswerPrint excels over existing alternatives (e.g., LUT, StyleID). Code and data will be released.

Primary area:
generative models
Keywords:
Style Transfer, Diffusion Models, Training-free Image Editing, AI for Art

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_QrRGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies transferring static tone and texture from a set of film frames to a content image. AnswerPrint pools inverted-latent statistics across references, injects attention from a representative frame by default, restores content attention along a sparse structure graph, and suppresses empty-prompt cross-attention in a frozen diffusion model. Evaluation includes the new Reel-30 benchmark, diffusion baselines, component ablations, a painting-style benchmark, and colorist preferences. The scope is still images with a coherent reference look, not temporal consistency.

Critical strengths

The reference-set formulation has a concrete motivation. StyleID’s Film Top-1 ranges from 0.563 to 0.387 across eight reference ranks (Table A3, Appendix C.1, PDF p. 18). Reference choice therefore matters substantially under this evaluation. Giving all single-image baselines the same colour-central frame also controls a meaningful source of comparison variability, although it need not be every method’s best reference.

The ablations quantify both structural recovery and its cost. The default secondary pass raises edge F1 from 0.610 to 0.635 while reducing Film Top-1 from 0.618 to 0.580. Adding correction reaches 0.647 and 0.607, respectively (Table 2, PDF p. 9). This supports a useful, training-free preservation mechanism and makes clear that its benefit is a trade-off, not improvement on every axis.

Critical weaknesses

The perceived film-look advantage needs stronger validation. Table 1 reports gains over StyleID in Film Top-1 (0.607 versus 0.563) and style similarity (0.222 versus 0.213; PDF p. 6). However, the probe’s 0.987 ceiling measures classification of real frames, not validity on stylized outputs (Appendix A.2, PDF p. 15). Appendix B.3 attributes confusions with Isle of Dogs to probe errors without a targeted independent assessment separating evaluation failure from transfer failure (PDF p. 17). The reported 69% colorist look preference could provide independent support, but Section 6.2 omits participant and trial counts, per-baseline results, and uncertainty; the supplied appendices do not provide them either. The central perceptual claim remains under-supported, not disproven.

The graph-specific contribution is not isolated. The primary grade builds on StyleID’s AdaIN and attention injection (Section 4.1, PDF p. 4; Appendix D, PDF p. 19), making the added preservation design important to the methodological contribution. Table 2 shows that restoration helps mainly where it operates, but this establishes locality rather than the value of the selected graph over simpler masks. Appendix C.2 tests curved contours, and Appendix E reports alternative graph-injection routes, but neither supplies a coverage-matched preservation control at comparable film identity (PDF pp. 18–19). It remains unclear how much the graph improves the trade-off beyond generic partial restoration of content.

Questions for authors

Human-study clarification: How many colorists and trials underlie Section 6.2, and how were films and contents sampled? Please provide per-baseline preferences, uncertainty accounting for repeated ratings, and exact metric–human agreement rates alongside an always-choose-the-majority-method baseline. These details could materially strengthen the film-look claim.

Minimal additional validation: Could you compare the graph with one equal-coverage random or simple-edge mask at comparable film identity? This would test whether graph selection improves the trade-off, rather than merely whether partial content restoration helps.

Evaluation clarification: Which edge detector, thresholds, and spatial matching tolerance produce the reported F1 scores, and how is graph-restricted scoring implemented? Section 5 refers to an independent detector, but Section 6.3 and Appendix A.2 do not specify these settings. Algorithm 1 describes anchor construction, not evaluation.

1 comment

We thank the reviewer for the precise, page-referenced review. We ran the requested control and answer each question below.

W2 / Q2: Isolating the graph from generic partial restoration. We ran exactly this control. With the injected layers, schedule, dose (wa=0.5w_a{=}0.5) and reel fixed, we replaced only the matte. Every alternative covers the same number of tokens as the graph on each image.

Matte (matched coverage)Film Top-1 ↑Edge F1F_1 ↑
Graph (ours)0.6070.647
Random tokens (3 draws)0.5940.636
Learned contours (PiDiNet)0.5920.647
Semantic boundaries (Mask2Former)0.5920.649
Canny edges*0.5970.660

*Circular: edge F1F_1 is itself computed on Canny edges.

At comparable film identity, the graph improves edge F1F_1 over random restoration of equal size (+0.012, 95% CI [0.006, 0.019]). Placement on structure therefore improves the trade-off beyond generic partial restoration. The graph matches learned and semantic mattes while remainig training-free and weight-free. We also measured graph coverage and edge-F1F_1 gain per content type:

ContentGraph coverageEdge F1F_1 gain
Built scenes12.7%+0.046
People10.9%+0.049
Animals3.8%+0.021

Performance improves as coverage increases (Spearman ρ = 0.56); the method remains effective even at the lowest levels of coverage, albeit with a relatively weaker effect.

W1 / Q1: Validity of the film-look evaluation and the human study. To clarify the role of the 0.987 ceiling: it is measured on held-out frames disjoint from those injected, so it shows that the probe separates the 30 films by their look rather than by scene content; we do not offer it as evidence of validity on stylised outputs. For that, three results already bear on it.

  • (i) Negative controls: modern image editors (details in our response to AVca) that leave the target nearly unchanged score close to chance (FLUX.2 0.060, Kontext 0.083; chance 0.033), and colour-only transfer (in Rdx2) reaches just 0.187, so the probe credits neither an untouched target nor a palette alone.
  • (ii) An independent backbone: style similarity uses DINOv2, a backbone independent of the CSD-based probe, and it ranks the top five methods in the same order as film Top-1.
  • (iii) Human judgement: in the blind two-alternative study with trained colourists, AnswerPrint is preferred on look in 69% of judgments, pooled over baselines. We are completing this study with our partner film academy and will report its full statistics, including how often the probe agrees with the raters' majority choice, which tests the probe on stylised outputs directly. A completed time-to-acceptance study points the same way: starting from AnswerPrint, colourists reached an accepted match in a median of 3 min, against 7.5 min from MKL colour transfer and 12 min from the ungraded plate, and 75% of AnswerPrint starts were accepted without correction (details in response to ZGSC).

Q3: Evaluation specification.

  • Edge F1F_1. We follow the condition-reconstruction protocol of ControlNet++. Canny edges are computed with a 3×3 blur and thresholds 100/200, and edges are matched between output and content within a 2-pixel tolerance.
  • Graph-restricted scores. The same computation is applied inside and outside the matte's 8-pixel token footprint.
  • Separation from the matte. The Canny evaluator is independent of the LSD-and-corner detector that builds the graph.

We will state these settings in the main paper.

Reviewer_ZGSCOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulates film style transfer (FST), where the visual look of a film is represented by a set of reference frames and transferred to a target image while preserving target structure. The proposed training-free method, AnswerPrint, combines pooled AdaIN statistics and self-attention injection for a primary grade, sparse structure-graph anchoring for a secondary grade, and attenuation of unconditional cross-attention outputs; the paper also introduces the Reel-30 benchmark and film-identity and edge-preservation metrics.

Critical strengths
  • The paper identifies a practically motivated distinction between conventional single-reference artistic style transfer and a multi-frame film-look setting. Figure 1 communicates this motivation effectively: the close-up and long-shot examples illustrate why a film's recognizable look cannot simply be equated with the content or palette of one still.

  • The coarse-to-fine decomposition is intuitive and technically coherent at a high level. In particular, and self-attention injection establish a global appearance, while the sparse content graph selectively restores local information. Equations (5) and (6) provide a reasonably concrete implementation of this local intervention through layer-specific rasterized masks.

  • The authors make a useful effort to separate style identity from content preservation.

  • Figure 5, right, also appropriately reports film-level paired bootstrap comparisons rather than only aggregate point estimates.

  • The qualitative comparisons are useful and show some real strengths of the method. Figure 6 covers both animation and live-action settings, and visually supports the claim that some baselines either under-transfer appearance or introduce content contamination.

  • The paper is unusually candid that FST is presently a still-image method and does not solve temporal coherence or broader cinematic storytelling. This limitation is important and is stated explicitly rather than hidden behind the post-production framing.

Critical weaknesses
  1. The paper's positioning substantially understates prior work on photorealistic style transfer and automated video color grading. Section 2 frames the comparison primarily around diffusion style-transfer methods, then states that “all share a single-image reference” and positions reel-level transfer as a new practical setting. However, reference-based photorealistic video style transfer, localized video stylization, neural LUT generation, and diffusion-assisted video color grading are highly relevant precedents.

  2. The empirical comparison does not fairly isolate the claimed benefit of using a reel. The proposed method receives Fa=8 frames for pooled AdaIN statistics, whereas single-image baselines are evaluated with one selected medoid frame. Giving single-image methods a favorable reference is considerate, but it does not answer the core causal question: does AnswerPrint succeed because it uses a set, because it uses a particular architecture, or because the baselines are unable to consume the available reference information? Section 6.3 shows that pooling more frames helps AnswerPrint, but the paper does not adapt plausible baselines to the same multi-reference setting, for example by aggregating feature statistics, averaging style embeddings, independently stylizing with multiple references and selecting or blending outputs, or applying a multi-reference LUT/color-transfer control. Since the contribution is explicitly reel-based style transfer, this is a central missing control rather than a cosmetic baseline request.

  3. The human evaluation is too underspecified to validate the claimed practical superiority. Section 6.2 reports that trained colorists preferred APrint in 69% of look-match and 83% of structure comparisons, but the main paper does not state the number of colorists, number of film-content pairs, number of comparisons per pair, rater agreement, whether the same raters evaluated all baselines, or confidence intervals/significance tests. The assertion that the metrics agree with “trained-eye ground truth” in >60%>60\% and >85%>85\% of pairs is also difficult to interpret without a formal definition of agreement and a measure of inter-rater reliability. These omissions matter because the paper relies on professional judgment to establish that the proposed metrics correspond to the desired production criterion.

  4. The paper overstates the post-production implication relative to the evaluation setting. Reel-30 stylizes static MSCOCO images, not raw footage or shots from a production pipeline. The paper acknowledges the temporal limitation near the end, but the abstract and conclusion still claim that it automates an important post-production step and is competitive with professional grading. That conclusion needs to be narrowed, or supported with video experiments and a controlled comparison to professional grades.

  5. The evidence for structural robustness is narrower than the headline claim. The anchor graph is based on line segments and corners, which is likely effective for architecture and rigid objects but not necessarily for faces, hair, animals, foliage, or other organic content. The paper should report edge and perceptual-structure results stratified by content category, graph coverage, and object type. It should also compare the proposed sparse graph to simpler edge masks, semantic boundaries, or modern contour detectors, since it is currently unclear whether the particular LSD-plus-corners graph is necessary.

Minor comment: The title "Film Style Transfer" is somewhat vague and does not specify what the paper does. It does not capture what the paper produces. This kind of very short titles are more appropriate for a book/tutorial on this topic. Even if the authors proposed task of FST is novel, the title seems to indicate that they are providing a full/best solution to this problem, which is not true. So, a title could be better scoped.

Questions for authors

Can the authors provide a controlled multi-reference comparison in which the strongest applicable baselines receive the same Fa=8 reference frames?

How does anchor coverage and edge F1 vary across portraits, animals, landscapes, interiors, and architecture? Does the method remain beneficial when the graph covers very few content tokens?

2 comments

We thank the reviewer for the detailed and constructive review. We ran the requested controls as bellow, and will all be added to the paper.

W2 / Q1: Controlled multi-reference comparison with the same Fa=8F_a{=}8 frames. We agree this is the central control, and we provide it at two levels.

(a) Across methods: every multi-reference baseline receives the same eight frames as AnswerPrint.

Method (same 8 frames)Film Top-1 ↑Edge F1F_1 ↑
StyleGallery0.3370.729
Ruta et al. (reimpl.)0.6350.491
StyleID + pooled statistics0.6130.606
Qwen-Image-Edit0.1880.509
Kontext0.0830.470
FLUX.20.0600.778
AnswerPrint0.6070.647

(b) Within the strongest baseline: StyleID receives the reel in every way the reviewer lists.

StrategyFilm Top-1 ↑Edge F1F_1 ↑Cost
Single central frame0.5630.5821×
Aggregate feature statistics (pooled AdaIN)0.6130.6061×
Average keys/values0.2200.0671×
Concatenate keys/values0.4850.6551×
Stylise per frame, select by colour0.5150.5998×
Stylise per frame, select by CSD*0.6500.5498×
Stylise per frame, blend outputs0.5530.5978×
Multi-reference LUT0.1450.628-

*Selects in the same embedding the film probe is trained on, so its identity score is circular.

This answers the causal question:

  • Reel access accounts for the identity gain (0.563 → 0.613).
  • Our secondary pass accounts for the structure gain (0.606 → 0.647) at near-equal identity.
  • Only pooled statistics improves both over the single frame without multiplying the cost.

W1: Positioning. We will extend the related work to colour transfer and example-based grading (Reinhard et al., the Monge–Kantorovich map behind LUT grading, HistoGAN, image-adaptive 3D LUTs, Bonneel et al.), photorealistic style transfer, localised video stylisation and diffusion-assisted video colour grading. We will also remove the statement that "all share a single-image reference", and two closed-form baselines (MKL, pooled LUT) are now evaluated (film Top-1 0.187 and 0.145).

W3: Underspecified human evaluation. We agree with the reviewer. Together with our partner film academy, we are completing the blind preference study. We will report:

  • the number of colourists and film–content pairs, and the comparisons per pair;
  • whether the same raters judged all baselines;
  • per-baseline preference rates with Wilson intervals and binomial tests;
  • inter-rater agreement (Fleiss' κ);
  • a formal definition of metric–human agreement, compared against an always-choose-AnswerPrint baseline.

In addition, we ran a fully specified time-to-acceptance study:

  • Design. One professional colourist graded 8 Reel-30 films, stratified by film category, in DaVinci Resolve from three starting points: the ungraded plate, the MKL colour transfer and AnswerPrint. Each film was paired with a different content for each starting point (24 plates in total). Task order was randomised and balanced across starting points, and the method behind each start was hidden.
  • Acceptance. A second colourist confirmed every accepted match.
  • Results. Median grading time was 12 min from the ungraded plate, 7.5 min from MKL (50% accepted without correction) and 3 min from AnswerPrint (75% accepted; Wilcoxon signed-rank over plates, p≈0.008p ≈ 0.008).

We will enlarge the rater pool for the final version.

W4: Overstated post-production implication. We will narrow the claims. The abstract will state that we study the still-image form of the matching step, and AnswerPrint will be positioned as a starting grade that shortens a colourist's turnaround (3.6× per shot including its own ≈20 s), not as a replacement for professional grading at this moment, and the future research efforts will be pushing this value further. The new video diagnostic (8 DAVIS clips; warping error 0.0048 vs. 0.0056 for StyleID, 0.0193 for StyleGallery, and 0.0093 for FLUX.2, with no visible flicker) will be added as a first step toward footage. For the results, please see: https://anonymous.4open.science/w/supp-vis-2027-18C1/index.html#video.

W5 / Q2: Robustness by content type, and comparison with simpler masks.

ContentGraph coverageEdge F1F_1 gain
Built scenes12.7%+0.046
People10.9%+0.049
Animals3.8%+0.021

Gain follows coverage (Spearman ρ = 0.56), and the method stays beneficial even at the lowest coverage, though less so. At matched coverage and fixed dose:

Matte (same token count)Film Top-1 ↑Edge F1F_1 ↑
Graph (ours)0.6070.647
Random tokens (3 draws)0.5940.636
Learned contours (PiDiNet)0.5920.647
Semantic boundaries (Mask2Former)0.5920.649
Canny edges*0.5970.660

*Circular: edge F1F_1 is computed on Canny edges.

Placement on structure is what matters (+0.012 over random, 95% CI [0.006, 0.019]). The graph matches learned and semantic mattes while being training-free and weight-free. We will add both tables.

Minor: Title. We will retitle the paper "Film Look Transfer: From a Reel of Frames to a New Shot", which scopes both the task and the setting.

Reviewer_AVcaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method that transfers a film’s style from reference frames to a user-provided still image. It uses a pretrained diffusion model to change the photo’s appearance, while using lines and corners automatically detected in the original photo to guide the preservation of local details. Experiments on a new benchmark covering 30 films show a better balance between matching the film’s style and preserving the original image content than the tested baselines.

Critical strengths
  1. Targeted preservation of fine details. The method uses lines and corners detected in the input photo to identify where original image features should be restored. The approach is simple, and the ablation supports its effectiveness.

  2. Useful findings on combining multiple references. The experiments provide a useful observation: pooling statistics from multiple frames helps, while expanding the attention bank does not. Since injection mechanisms are not new, I consider this a modest empirical contribution.

Critical weaknesses
  • The advance over existing methods is insufficiently established. The method mainly extends StyleID with several test-time improvements. Existing multi-reference methods such as StyleGallery are neither discussed nor evaluated. The current experiments do not clearly establish the added value of the proposed film-specific design.

  • The proposed structure protection has limited coverage. The default anchors mainly focus on straight structures and seem to provide little coverage for portraits and animals. Adding curved contours improves structure preservation but weakens film-style recognition. This limits the scope of the proposed detail-preservation mechanism.

  • Modern image-editing models already support style transfer and multi-image inputs. The paper does not establish what additional capability its dedicated pipeline provides, or whether it offers a meaningful advantage in quality, content preservation, or efficiency. Being faster than manual grading does not establish an advantage over these existing automated alternatives.

Overall, I felt the paper does not make a strong case for what important challenge the film-specific setting introduces beyond existing style transfer and image editing.

Questions for authors
  1. How does AnswerPrint compare with a modern image-editing model, such as Qwen-Image-Edit, in style matching, content preservation, and runtime?

  2. How does it compare with StyleGallery when both use the same content images and film references?

2 comments

We thank the reviewer for the frank assessment. We ran every comparison the reviewer asked for. All results below will be added to the paper. Visual comparisons for all results in this response are on this page: https://anonymous.4open.science/w/supp-vis-2027-18C1/index.html.

Q1 / W3: Comparison with modern image editors. We evaluated Qwen-Image-Edit-2509, FLUX.1 Kontext [dev] and FLUX.2 [dev] on Reel-30 (600 outputs each). Each received the same eight reel frames as a tiled sheet beside the target, with one fixed instruction to re-grade the target to the sheet's look while keeping its content.

MethodFilm Top-1 ↑Style sim. ↑Edge F1F_1 ↑s / image ↓Peak memory ↓
Qwen-Image-Edit0.1880.0730.50974.960.4 GB
FLUX.1 Kontext0.0830.0160.47013.736.4 GB
FLUX.20.0600.0080.77897.868.1 GB
AnswerPrint0.6070.2220.64721.24.8 GB

The editors fail in one of two ways:

  • They leave the target nearly unchanged. FLUX.2 and Kontext stay close to chance on film Top-1 (chance is 0.033).
  • They replace the target. In 24% of Qwen-Image-Edit's outputs, the target is replaced by a mosaic of reel frames.

AnswerPrint exceeds all three on film Top-1, with each paired-bootstrap interval clear of zero, and needs 7.6–14× less memory. Time and memory are measured on one GH200, amortised over 20 targets per film.

Q2 / W1: Comparison with StyleGallery under the same inputs. With the same content images and the same eight reel frames:

MethodFilm Top-1 ↑Edge F1F_1 ↑s / image ↓
StyleGallery0.3370.72968.9
AnswerPrint0.6070.64721.2

StyleGallery keeps structure but carries little of the film (difference 0.270, 95% CI [0.183, 0.357] in our favour). It also needs per-image optimisation and takes over three times as long. We will add StyleGallery to the related work and the main table.

W1: Advance over StyleID. We separate our contribution into two steps by giving StyleID the whole reel in seven ways (same eight frames for every row):

How the reel enters StyleIDFilm Top-1 ↑Edge F1F_1 ↑Cost
One central frame (StyleID)0.5630.5821×
Pooled AdaIN statistics (our primary grade)0.6130.6061×
Averaged keys/values0.2200.0671×
Concatenated keys/values0.4850.6551×
Per-frame runs, select by colour0.5150.5998×
Per-frame runs, pixel average0.5530.5978×
Pooled LUT0.1450.628–

Only pooled statistics improves identity and structure together at no extra cost. Our secondary pass then raises edge F1F_1 further, from 0.606 to 0.647, at near-equal identity (0.607). The key design on top of StyleID is a crafted use of the reel plus targeted structure restoration, neither of which StyleID provides.

W2: Coverage of the structure graph. We measured graph coverage and edge-F1F_1 gain per content type:

ContentGraph coverageEdge F1F_1 gain
Built scenes12.7%+0.046
People10.9%+0.049
Animals3.8%+0.021

Portraits benefit as much as architecture; animals benefit least (Spearman ρ = 0.56 between coverage and gain), which we will discuss as a limitation and a future work item. Denser curved coverage withdraws the film's look from more tokens, which is why the matte is kept sparse. At matched coverage, the graph beats random tokens and matches learned contours (PiDiNet) and semantic boundaries (Mask2Former) without any training (details in our response to QrRG).

On "faster than manual grading does not establish an advantage over automated alternatives". We agree, and added a time-to-acceptance study in which an automated alternative is one of the starting points. One professional colourist graded 8 Reel-30 films, stratified by film category, to an accepted match in DaVinci Resolve from three starting points: the ungraded plate, the MKL colour transfer, and AnswerPrint. Each film was paired with three different contents, one per starting point (24 plates in total). Task order was randomised and balanced across starting points, the method behind each start was hidden, and a second colourist confirmed every accepted match. We compare the active grading time to acceptance with a Wilcoxon signed-rank test paired over the 8 films:

Starting pointMedian time to accepted matchAccepted as-is (confirmed)
Ungraded plate12 min–
MKL color transfer7.5 min50%
AnswerPrint3 min75%

AnswerPrint is faster than the MKL start on 8 of 8 films (Wilcoxon signed-rank, p≈0.008p ≈ 0.008), so the advantage holds against an automated alternative, not only against hand grading. We will enlarge the study for the final version.

What the film setting introduces. Film Style Transfer makes four demands: (R1) read the look from a set; (R2) move texture as well as colour; (R3) keep the target's content and fine structure; (R4) work per shot without per-film training. Against the results above:

ParadigmR1R2R3R4Evidence on Reel-30
Colour transfer / LUT◦×◦✓Top-1 0.187 / 0.145
Single-ref. diffusion×✓◦✓StyleID 0.563 → 0.387 by frame
Multi-ref. diffusion✓✓◦◦StyleGallery Top-1 0.337; Ruta edge F1F_1 0.491
Image editors◦✓×◦Top-1 ≤ 0.188; 24% outputs replaced; 60+ GB
AnswerPrint✓✓✓✓Top-1 0.607, edge F1F_1 0.647, 4.8 GB
(✓ met, ◦ partly, × not met.)

Each existing paradigm misses at least one demand. This is the significant challenge that the "film-specific scene" setting introduces to our style transfer task. We will add this analysis to the paper.

Reviewer_Rdx2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper formulates Film Style Transfer (FST), aiming to transfer the visual characteristics of a film, represented by multiple reference frames, to a target image while preserving its original content and structure. The authors propose AnswerPrint, a training-free framework built on pretrained diffusion models. Inspired by professional color-grading workflows, the method consists of three components: (1) a primary grade that aggregates style statistics across reference frames and performs attention-based style injection; (2) a secondary pass that uses a sparse structural graph to restore fine-grained content details; and (3) a process correction that suppresses near-constant cross-attention outputs. The authors also introduce Reel-30, a benchmark covering 30 films, with evaluation metrics for film-style identity and structural fidelity. Experiments compare AnswerPrint against several existing style transfer methods, including StyleID and StyleSSP. Quantitative results, ablation studies, and evaluations by professional colorists demonstrate improvements in style matching and structural preservation, alongside competitive computational efficiency.

Critical strengths

(1) The paper addresses a relevant problem in film post-production by extending style transfer to multiple reference frames. The proposed coarse-to-fine framework combines multi-frame style aggregation with structure-aware refinement. The sparse graph-guided restoration and cross-attention correction are interesting designs that improve content preservation without additional training. (2) The authors introduce Reel-30 and evaluate their method using multiple metrics, comparisons with existing methods, ablation studies, and assessments by professional colorists. The results demonstrate improvements in both film-style matching and structural fidelity, providing meaningful evidence for the effectiveness of the proposed approach.

Critical weaknesses

(1) The paper repeatedly presents FST as a novel computer vision task. However, reference-based film color grading and multi-reference style transfer have already been explored in previous studies. While the proposed formulation is valuable, the authors should more clearly distinguish FST from these existing tasks and acknowledge closely related work, such as Example-Based Video Color Grading (SIGGRAPH 2013) and Leveraging Diffusion Models for Stylization using Multiple Style Images (ICCVW 2025). (2) Although the paper is motivated by film post-production, the experiments focus on independent static images. Temporal consistency across consecutive frames and visual consistency across different shots are not evaluated. A small-scale video experiment would help establish how well the proposed method translates to its intended application.

Questions for authors

(1) How does the proposed FST formulation fundamentally differ from existing reference-based film color grading and multi-reference style transfer methods, such as Example-Based Video Color Grading (SIGGRAPH 2013) and Leveraging Diffusion Models for Stylization using Multiple Style Images (ICCVW 2025)? Clarifying these distinctions would help establish the novelty and significance of the proposed task. (2) Since the paper emphasizes achieving a consistent film look across different shots, could the authors provide additional evidence of style consistency when applying the same reference reel to multiple target images with different content and lighting conditions? Such an evaluation would help substantiate the claimed practical benefits for film post-production, even within the paper's static-image setting.

1 comment

We thank the reviewer for the positive assessment and the pointers to related work.

W1 / Q1: How does FST differ from example-based video colour grading [Bonneel et al. 2013] and multi-reference stylisation [Ruta et al. 2025]? We agree both are closely related, and we will discuss them explicitly. We do not claim to be the first to use multiple references. Our claim is narrower: a film's look is a property of a set of frames, and transferring it requires application specific considerations at once:

  • (R1) reading the look from the set;
  • (R2) moving grain and rendering texture as well as colour;
  • (R3) keeping the target's content and fine structure intact;
  • (R4) doing so per shot without per-film training.

To make the distinction concrete and provide more extensive comparison, we evaluated both families on Reel-30 with the same eight reel frames as the style reference:

MethodFilm Top-1 ↑Edge F1F_1 ↑
MKL colour transfer0.1870.812
Pooled LUT0.1450.628
Ruta et al. (reimpl.)0.6350.491
AnswerPrint0.6070.647

Main observations include:

  • Colour-level methods, including example-based grading, are fast and keep geometry (Edge F1F_1 0.812), but cannot move grain or rendering texture (R2). They therefore carry little of the film.
  • Ruta et al. match AnswerPrint on film identity (difference −0.028, 95% CI [−0.142, 0.085]). Being designed for artistic styles, they have no mechanism to protect the target's fine structure (R3), and they keep markedly less of it (Δ edge F1F_1 = 0.156, CI [0.112, 0.192]).

We will add this analysis, together with a table placing each paradigm against R1–R4, to the related work.

W2: Temporal consistency and a small video experiment. We ran a video diagnostic. Each frame of 8 DAVIS-2017 clips (24 frames each; fur, plumage, a dancer, a rider, vehicles, a boat) was graded independently toward a fixed reel. We measured the warping error under RAFT flow with occlusion masking:

Ungraded clipAnswerPrintStyleIDFLUX.2StyleGallery
Warping error ↓0.00240.00480.00560.00930.0193

Every AnswerPrint frame is still recognised as its own film, and played back, AnswerPrint's clips show no visible flicker: tone, grain and painted texture hold from frame to frame. This is preliminary evidence that AnswerPrint is usable on short clips without a temporal module. We will add this experiment to the paper and leave dedicated temporal modelling to future work. Side-by-side videos of all eight clips (AnswerPrint, StyleID, StyleGallery and FLUX.2, three films each) are on this page: https://anonymous.4open.science/w/supp-vis-2027-18C1/index.html#video

Q2: Consistency when one reel is applied to many targets. Reel-30 applies each reel to 20 different contents, and film Top-1 asks whether every output is recognised as the same film by a probe trained only on that film's held-out frames (ceiling 0.987 on real frames). A film Top-1 of 0.607 therefore reflects consistency across contents. Following the suggestion, we also measured consistency directly. For each film, we computed the spread of its 20 outputs in the probe (CSD) embedding (mean pairwise cosine distance) and divided it by the spread of the film's own held-out frames, which show different scenes in the same look. We also report how far the outputs' centroid lies from the film's frame centroid, in units of the median distance between two films.

MethodSpread / reel spreadCentroid-to-film distance ↓Edge F1F_1 ↑
Ungraded contents1.29––
MKL colour transfer1.150.7500.812
FLUX.21.200.9240.778
StyleGallery1.080.6000.729
StyleID0.670.5940.582
Ruta et al. (reimpl.)0.380.5600.491
AnswerPrint0.750.5410.647

AnswerPrint's outputs centroid lies closest to the film among all 16 methods evaluated (pooled StyleID is a near tie at 0.542). Methods that keep the target's colours (MKL, the editors) spread almost as widely as the ungraded contents, so the reel does not hold their outputs together. Also, a lower spread alone is not better: Ruta et al. and InjectFusion (0.38) cluster more tightly by overwriting the target, at a cost in structure. We will add the full per-film table for all methods to the appendix.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response summary

We thank all reviewers for their careful and constructive reviews. During the rebuttal period we ran all the additional experiments requested; their results are reported in the individual responses below, and all of them will be added to the paper. In summary:

  1. Broader baselines on Reel-30. We added (i) closed-form colour transfer (MKL and a pooled LUT); (ii) multi-reference methods given the same eight reel frames (i.e., the style) as AnswerPrint (StyleGallery, our reimplementation of Ruta et al. ICCVW 2025, and StyleID with pooled statistics); and (iii) modern image editors given the same input frames as a tiled sheet (Qwen-Image-Edit, FLUX.1 Kontext, FLUX.2).
  2. Controlled multi-reference study. We fed the reel to StyleID in seven different ways to separate the effect of reel access from that of our design.
  3. Matte isolation. At matched token coverage, we compared the structure graph with Canny edges, semantic boundaries, learned contours and random tokens, and we report coverage and gain per content type.
  4. Human evaluation. We added a time-to-acceptance study with professional colourists. The blind preference study is being completed with our partner film academy, and we will report its full statistics.
  5. Video diagnostic. We added a video diagnostic on 8 DAVIS clips.
  6. Scope. AnswerPrint addresses the still-image form of the matching step, as a starting grade for colourists rather than a replacement for them.

All supplementary visual results, including the videos, are on an anonymous page: https://anonymous.4open.science/w/supp-vis-2027-18C1/index.html.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.