Pith. sign in

REVIEW 4 major objections 5 minor 31 references

CNNs for Style Transfer of Digital to Film Photography

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A simple CNN reproduces the color of Cinestill 800T film, but not its grain or halation.

desk verdict A modest but honest dataset-plus-baseline paper whose quantitative claims are undercut by evaluation on the training set; the dataset alone justifies a serious referee. read the letter →

arxiv 2411.15967 v1 pith:LN4EAMHT submitted 2024-11-24 cs.CV

Pierre Mackenzie , Mika Senghaas , Raphael Achddou This is my paper · ORCID
classification cs.CV
keywords styletransferfilmphotographyCinestill800TU-Netperceptuallosspaireddatasetimageenhancement
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

This paper asks whether a plain convolutional network can translate digital photographs into the look of Cinestill 800T film, an iconic tungsten-balanced stock known for its teal-blue cast, pronounced grain, and red halation around bright lights. The authors collected 41 paired digital–film shots of the same scenes, aligned them with keypoint matching and luminance histogram equalization, and trained a small U-Net under different loss functions and training augmentations. Their central finding is that a combination of pixel-wise MSE and VGG perceptual loss, trained on randomly resized crops, reproduces the film's color transformation well enough to beat the baseline on most metrics. However, the same model produces only partial, low-quality grain and never produces halation. The paper attributes this to the small, varied dataset and releases the aligned pairs to encourage further work.

What carries the argument

The carrying mechanism is the U-Net encoder–decoder with skip connections, trained end-to-end on paired digital–film images. The learning signal is a weighted sum of pixel-wise losses (MSE or MAE), a perceptual VGG-19 loss computed on conv1_2, conv2_2, conv3_2 features, and an optional relative total-variation loss that compares noise levels between prediction and target. Two training augmentations are load-bearing: a random noise channel concatenated to the input, and resizing randomly cropped patches to the network input size, which the authors find significantly improves the learned color mapping.

What would settle it

Apply the same luminance-histogram matching used in preprocessing directly to the raw digital test images and compare their PSNR/SSIM against the film ground truth; if the histogram-matched images match the film images as closely as, or closer than, the MSE/VGG network outputs, then the learned color effect is largely inherited from the preprocessing rather than from the network.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a shallow fully convolutional U-Net, trained with a weighted MSE and VGG-19 feature loss on resized random crops, learns the color appearance of Cinestill 800T film from a few dozen aligned pairs, matching or exceeding the unprocessed digital baseline on PSNR and SSIM, while the textural effects of film—grain and halation—remain largely out of reach. Adding a noise input channel helps produce some variation but not authentic grain; a relative total-variation loss nudges the output toward noisier textures without matching the film's grain statistics; and no configuration in any experiment yields halation. The paper frames this as evidence that pure deep learning, at this dataset scale, is sufficient for the color transfer but not the full photographic effect.

Load-bearing premise

The entire comparison rests on the preprocessing pipeline producing pairs aligned well enough for pixel-wise losses to be meaningful; if residual misalignment remains, the network may be fitting alignment artifacts rather than the film style.

Editorial extensions

If this is right

  • If the finding holds, a lightweight CNN with an MSE/VGG loss is a viable starting point for digital-to-film color grading, at least for Cinestill 800T, and could run in seconds on high-resolution images.
  • The negative result for grain and halation implies that pixel-wise and perceptual losses alone do not drive a network to synthesize film texture, so applications that need authentic grain will require additional losses or explicit texture synthesis.
  • The released paired dataset gives the community a standardized testbed for comparing film-emulation methods under controlled alignment and luminance normalization.
  • The observation that resized crops improve color suggests that dataset patch sampling strategies can matter more than architecture choice for small training sets.

Reading between the lines

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

  • A likely reason halation never appears is not just data scarcity but the absence of a loss that rewards the characteristic bright-light glow; a training objective that separately targets halated regions (or an adversarial loss) would be a direct test of this interpretation.
  • The alignment pipeline (ORB+FLANN+RANSAC plus histogram matching) is the hidden variable in every quantitative comparison; if residual misalignment differs across scene types, the reported superiority of MSE/VGG could be partly an artifact of which scenes align well.
  • A hybrid pipeline—learned color transformation from this approach combined with statistical film-grain synthesis—would likely outperform either pure approach alone, and the paper's own appendix suggests such a direction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a pure deep-learning approach for translating digital photographs to the look of Cinestill 800T film. It contributes a small paired dataset of 38 aligned digital/film image pairs, a U-Net architecture, and an experimental study of loss functions (MSE, MAE, VGG, color, TV-Rel and combinations), an optional noise input channel, and random-resized patch training. The main positive claim is that MSE/VGG with resizing yields the best color production and the best quantitative metrics, while grain is only partially produced and halation is not produced. The paper also releases code and the dataset.

Significance. If the central claim were supported, the paper would provide a simple, reproducible baseline for film-style transfer and a useful public dataset for a relatively under-studied computational photography task. The dataset and code release are valuable: the authors explicitly share aligned paired images and PyTorch code, and they are honest about the limitations of their approach. The paper also makes falsifiable observations, such as the failure to produce halation and the limited grain quality. However, the quantitative evidence for the main claim is undermined by evaluation-on-training-data and train-on-test-image protocols, so the current manuscript does not establish the claimed superiority of MSE/VGG with resizing as a generalizable result.

major comments (4)
  1. [Section 3.4, Tables 4 and 5] The full-dataset evaluation is performed on the full dataset, which includes the training images. Section 3.4 states: 'The full dataset is split into a training, validation and test set using a 70-20-10 split ratio, but final evaluation is done on the full dataset.' Consequently, every metric in Tables 4 and 5, and the claim in Section 4.2 that MSE/VGG with resizing 'produces the best metrics across the board,' is computed on images used for optimization. These numbers cannot support a generalization claim. The authors should re-run evaluation on the held-out test split only, and report results separately for training, validation, and test images, along with variance across training runs.
  2. [Section 4.1, Tables 2 and 3] The single-image experiments train and evaluate on the same image, as stated in Section 3.4 ('we train and evaluate the model on the same image'). The relative ranking of losses in Tables 2 and 3 (e.g., Color/VGG/TV-Rel achieving the best SSIM without noise, and MSE/VGG improving with resizing) therefore reflects overfitting to one image, not a general property of the losses. This makes the selection of 'most promising candidates' for the full-dataset experiments (Section 4.2) unreliable. A proper validation split or cross-validation is needed to compare loss configurations.
  3. [Section 3.1, Tables 4 and 5] The preprocessing pipeline (ORB keypoint matching, FLANN, RANSAC homography, and luminance histogram matching) is assumed to produce alignment sufficient for pixel-wise and small-patch losses, but no residual alignment error is reported. If misalignment remains, pixel-wise losses like MSE and VGG features will be penalized by geometric mismatch, and the apparent advantage of MSE/VGG over the baseline could be an artifact of particular alignment errors rather than a learned film-style mapping. The authors should report alignment residuals (e.g., keypoint reprojection error or a visual alignment check) and, ideally, test robustness by evaluating on an alignment-robust loss or by perturbing the alignment.
  4. [Tables 4 and 5, Section 4.2] The dataset has only 38 pairs, and no error bars or repeated training runs are reported. The reported differences are small: SSIM 0.64 vs 0.64 and PSNR 22.87 vs 21.68 for the best model versus baseline. Without statistical significance or variance estimates, the claim that MSE/VGG 'produces the best metrics across the board' is not supported beyond the specific run. The authors should provide confidence intervals or at least standard deviations over multiple seeds, and acknowledge when differences are within noise.
minor comments (5)
  1. [Throughout] There are several typos: 'applicatino' (Section 1), 'baheviour' (Section 3.1), 'hyperpamaters' (Section 6), and 'Guassian' (Section 3.4 footnote 6).
  2. [References] References [11] and [12] are the same paper (Johnson et al., 2016), which creates confusion when both are cited for different purposes in Section 2. Also, [31] is cited for the contextual bilateral loss, but the cited paper title 'Zoom to learn, learn to zoom' does not appear to describe that loss; the attribution should be verified.
  3. [Appendix Table 7] In Table 7, the row 'MSE Yes No' appears twice with different values (0.63/22.20/0.35/2.31 and 0.59/19.68/0.48/3.38). Based on Table 4, the second row appears to be MAE with no noise and no resize, so the loss label is likely a typographical error.
  4. [Figure 4 caption] The caption states 'the best performing model is MSE/VGG with resizing,' but that configuration is not shown in Figure 4 and appears only in Table 3; the caption should either include that configuration or refer to Table 3.
  5. [Section 3.5] The PieAPP metric is described only as 'lower indicates higher similarity'; the range and interpretation should be stated for readers unfamiliar with it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the loss functions, hyperparameters, and evaluation baseline are externally fixed, and the central comparison is an empirical outcome rather than a definitional reduction.

full rationale

The paper's derivation chain is empirical rather than definitional. A U-Net is trained on paired digital-film images with losses defined in Eqs. 1-6; the loss weights are set to fixed values ('We simply use equal weights in all experiments with beta_i = 1 for all i' in Section 3.3) and the VGG layer weights are stated constants (0.4, 0.4, 0.2), not fitted to the target result. The claim that MSE/VGG with resizing gives the best colour production is supported by Tables 4 and 5, which compare against a digital-input baseline, so the scores are not forced by construction. No parameter is fitted to the metric used to declare success, and no prediction is the renaming of an input. There are no self-citations: all cited works are external ([1], [9], [24], [26], etc.), and the paper does not invoke any author-derived uniqueness theorem or ansatz. The paper's own limitation statements in Section 5 ('Our results are severely limited by the small size and great variety of our dataset') and the evaluation protocol in Section 3.4 ('final evaluation is done on the full dataset') are methodological concerns about generalization and train-set leakage, but they do not make the derivation circular. The central ranking is an empirical finding, not an equation-level identity. Therefore no significant circularity is present.

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

The central claim depends on standard deep learning assumptions (architecture, loss, optimizer) and on preprocessing assumptions about alignment and luminance handling. No new physical entities are introduced, and the free parameters are mostly standard hyperparameters or hand-set loss weights.

free parameters (4)
  • VGG layer weights (lambda_k) = 0.4, 0.4, 0.2
    Weights for VGG loss layers conv1_2, conv2_2, conv3_2 chosen by hand in Section 3.3.
  • Loss weights beta_i = 1 for all i
    Equal weights for all loss components, chosen arbitrarily in Section 3.3.
  • Gaussian blur kernel for color loss = kernel size 7, sigma 3
    Kernel parameters for color loss, borrowed from prior work [9] in Section 3.3.
  • Model hyperparameters = 64/128/256 filters, depth 3, patch size 256, lr 1e-3
    Standard values from related work [24, 9, 1], not fitted to this dataset but set for the experiments.
assumptions (4)
  • domain assumption ORB keypoint detection, FLANN matching, and RANSAC homography produce sufficiently aligned image pairs.
    Used in preprocessing (Section 3.1); if alignment is poor, pixel-wise losses are corrupted.
  • domain assumption Histogram matching on luminance removes brightness differences without distorting the film color characteristics.
    Preprocessing step in Section 3.1 that alters the film look before training.
  • domain assumption VGG-19 features are a valid perceptual similarity measure for optimizing this style transfer task.
    VGG loss is used in all combined losses in Section 3.3; if VGG features do not align with film perception, the loss misguides training.
  • domain assumption A dataset of 38 image pairs captures the variety of Cinestill800T effects.
    The paper's own conclusion states sample size severely limits results (Section 5), undermining the generality of findings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CNNs for Style Transfer of Digital to Film Photography." pith.science (2026). https://pith.science/paper/LN4EAMHT

@misc{pith2026241115967,
  author       = {Pith},
  title        = {Pith review of: CNNs for Style Transfer of Digital to Film Photography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LN4EAMHT}},
  note         = {Machine review of arXiv:2411.15967}
}
read the original abstract

The use of deep learning in stylistic effect generation has seen increasing use over recent years. In this work, we use simple convolutional neural networks to model Cinestill800T film given a digital input. We test the effect of different loss functions, the addition of an input noise channel and the use of random scales of patches during training. We find that a combination of MSE/VGG loss gives the best colour production and that some grain can be produced, but it is not of a high quality, and no halation is produced. We contribute our dataset of aligned paired images taken with a film and digital camera for further work.

Figures

Figures reproduced from arXiv: 2411.15967 by the authors.

Figure 1
Figure 1. Raw Paired Image Dataset. Examples of raw image pairs from the dataset. A column shows a single scene captured with a digital camera (top) and a film camera (bottom). The images show a wide range of scenes and visual effects. the model to learn the differences in appearance between the two media. Images should be aligned, as our model should only transfer style and not learn move contents of the image around, and di… view at source ↗
Figure 2
Figure 2. Dataset Preprocessing. Example of a raw and processed image pair. We can see the luminance alignment in the film image (a) and the spatial alignment in the digital image (b). by a non-linear activation functions and pooling operations. The decoder integrates both feature and spatial information using up-convolutions and high-resolution features via skip connections. We selected UNet as it is a relatively simple mode… view at source ↗
Figure 3
Figure 3. Training Overview. We train our model on paired digital-film images. Intuitively, the VGG loss encourages the model to generate images that are similar to the ground truth images in terms of their high-level semantics such as edges, textures and patterns. Color Loss. Inspired by [9], we consider an alternative to the MSE loss, which we refer to as the color loss. The color loss computes the MSE between two images, X… view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: We also experiment with resizing cropped patches during training. This is done by taking a crop at a random scale from the image and resizing it to the patch size, 256 × 256. The results are in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 4
Figure 4. Figure 4: Single Image Select Samples. Outputs from select loss functions and configurations of noise and resizing. We see that the best performing model is MSE/VGG with resizing, which produces the best colour effect. We also see that LPIPS scores are inconsistent with SSIM and…
Figure 5
Figure 5. Figure 5: Single Image Noise Comparison. Outputs from select models without noise (above) and with noise (below), no resizing. We see that when we feed noise into the model, the model learns to produce some variation, especially when the loss contains a feature-based metric like…
Figure 6
Figure 6. Figure 6: Single Image Resizing. Outputs from MSE and MSE/VGG with and without resizing, and then with resizing and with and without noise. We see that resizing improves the colour effect, but the grain effect is not as strong as in the single image experiments. (a) MSE/VGG patc…
Figure 7
Figure 7. Figure 7: Single Image Resizing with Noise. Outputs from MSE and MSE/VGG with and without noise and resizing. We see that similar grain is still produced even through resizing. 4.2 Full Dataset Results We select the best performing losses and settings from the single image exper…
Figure 8
Figure 8. Figure 8: Full Dataset Select Samples. Outputs from select models on the full dataset. We see that the best performing model is MSE/VGG with resizing, which produces the best colour effect. We also see that the grain effect is not as strong as in the single image experiments. on…
Figure 9
Figure 9. Figure 9: Full Dataset Grain Comparison. Predictions for a patch of sky from select loss functions, all with noise. Similarly to the single image experiments, the models with losses that target some noise produce a more visually pleasing result. The best visual result is produce…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

  1. [1]

    CoRR abs/2106.13883 (2021), https://arxiv.org/abs/2106.13883

    Afifi, M., Abuolaim, A.: Semi-supervised raw-to-raw mapping. CoRR abs/2106.13883 (2021), https://arxiv.org/abs/2106.13883

    arXiv 2021

  2. [2]

    Bradski, G.: The OpenCV Library. Dr. Dobb’s Journal of Software Tools (2000)

    2000

  3. [3]

    Deep Colorization

    Cheng, Z., Yang, Q., Sheng, B.: Deep colorization. CoRR abs/1605.00075 (2016), http: //arxiv.org/abs/1605.00075

    work page Pith review arXiv 2016

  4. [4]

    2015 IEEE International Conference on Computer Vision (ICCV) pp

    Deshpande, A., Rock, J., Forsyth, D.A.: Learning large-scale automatic image colorization. 2015 IEEE International Conference on Computer Vision (ICCV) pp. 567–575 (2015), https: //api.semanticscholar.org/CorpusID:13195504

    work page 2015

  5. [5]

    Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks (2015)

    work page 2015

  6. [6]

    Galerne, B., Raad, L., Lezama, J., Morel, J.M.: Scaling painting style transfer (2022)

    work page 2022

  7. [7]

    In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2414– 2423 (2016). https://doi.org/10.1109/CVPR.2016.265

    doi:10.1109/cvpr.2016.265 2016

  8. [8]

    Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normaliza- tion (2017)

    work page 2017

Show all 31 references
  1. [9]

    CoRR abs/1704.02470 (2017), http://arxiv

    Ignatov, A., Kobyshev, N., Vanhoey, K., Timofte, R., Gool, L.V .: Dslr-quality photos on mobile devices with deep convolutional networks. CoRR abs/1704.02470 (2017), http://arxiv. org/abs/1704.02470

    2017 arXiv

  2. [10]

    Isola, P., Zhu, J.Y ., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversar- ial networks (2018)

    2018

  3. [12]

    Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super- resolution (2016)

    2016

  4. [13]

    Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for improved quality, stability, and variation (2018)

    2018

  5. [14]

    Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks (2019)

    2019

  6. [15]

    Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks (2016)

    2016

  7. [16]

    Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., Shi, W.: Photo-realistic single image super-resolution using a generative adversarial network (2017)

    2017

  8. [17]

    ACM Trans

    Liu, X., Wan, L., Qu, Y ., Wong, T.T., Lin, S., Leung, C.S., Heng, P.A.: Intrinsic colorization. ACM Trans. Graph. 27(5) (2008). https://doi.org/10.1145/1409060.1409105, https://doi. org/10.1145/1409060.1409105

    2008

  9. [18]

    Luan, F., Paris, S., Shechtman, E., Bala, K.: Deep photo style transfer (2017)

    2017

  10. [19]

    Mechrez, R., Talmi, I., Zelnik-Manor, L.: The contextual loss for image transformation with non-aligned data (2018)

    2018

  11. [20]

    Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photo upsampling via latent space exploration of generative models (2020)

    2020

  12. [21]

    Muja, M., Lowe, D.: Fast approximate nearest neighbors with automatic algorithm configuration. vol. 1, pp. 331–340 (01 2009)

    2009

  13. [22]

    Image Processing On Line 7, 165–183 (2017), https://doi.org/10.5201/ipol.2017.192 12

    Newson, A., Faraj, N., Galerne, B., Delon, J.: Realistic Film Grain Rendering. Image Processing On Line 7, 165–183 (2017), https://doi.org/10.5201/ipol.2017.192 12

    2017 doi

  14. [23]

    CoRR abs/1806.02067 (2018), http://arxiv.org/abs/1806.02067

    Prashnani, E., Cai, H., Mostofi, Y ., Sen, P.: Pieapp: Perceptual image-error assessment through pairwise preference. CoRR abs/1806.02067 (2018), http://arxiv.org/abs/1806.02067

    2018 arXiv

  15. [24]

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation (2015)

    2015

  16. [25]

    Rublee, E., Rabaud, V ., Konolige, K., Bradski, G.: Orb: an efficient alternative to sift or surf. pp. 2564–2571 (11 2011). https://doi.org/10.1109/ICCV .2011.6126544

    2011

  17. [26]

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recogni- tion (2015)

    2015

  18. [27]

    Wang, X., Yu, K., Dong, C., Loy, C.C.: Recovering realistic texture in image super-resolution by deep spatial feature transform (2018)

    2018

  19. [28]

    Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y ., Dong, C., Loy, C.C., Qiao, Y ., Tang, X.: Esrgan: Enhanced super-resolution generative adversarial networks (2018)

    2018

  20. [29]

    Wang, Y .D., Swietojanski, P., Armstrong, R.T., Mostaghimi, P.: Pixel co-occurence based loss metrics for super resolution texture recovery (2020), https://openreview.net/forum?id= rylrI1HtPr

    2020

  21. [30]

    CoRR abs/1801.03924 (2018), http://arxiv.org/ abs/1801.03924

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. CoRR abs/1801.03924 (2018), http://arxiv.org/ abs/1801.03924

    2018 arXiv

  22. [31]

    CoRRabs/1905.05169 (2019), http://arxiv.org/abs/1905.05169

    Zhang, X.C., Chen, Q., Ng, R., Koltun, V .: Zoom to learn, learn to zoom. CoRRabs/1905.05169 (2019), http://arxiv.org/abs/1905.05169

    2019 arXiv

  23. [32]

    The code is written in Python and uses PyTorch, Lightning and Hydra for config- uration

    Zhu, J.Y ., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle- consistent adversarial networks (2020) 13 7 Appendix 7.1 Code Documentation The code for this project is available at https://github.com/mikasenghaas/sillystill/ tree/main. The code ...

    2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.