Pixel-Level Domain Transfer

• The paper proposes pixel-level transfer that directly generates a target image from a semantically related source, rather than adapting model parameters in feature space.Source: §1 · pp. 1–3
• Because one source can admit many possible targets, the encoder-decoder converter must preserve semantic meaning while generating realistic pixels, as in dressed-person-to-clothing transfer.Source: §1 · pp. 1–2
• A domain discriminator judges whether source-target pairs are associated, while a real/fake discriminator encourages realistic outputs; both supervise the converter during adversarial training.Source: §1 · p. 2
• On the dressed-person-to-top-product task, LookBook contains 84k images, including 75k human images associated with 10k top product images, and evaluation verifies the domain discriminator’s effectiveness.The framework’s two discriminators are no longer needed after training; the converter is the final model.Source: §1 · pp. 2–3

2 Related Work

• Image-generation work spans parametric models, which struggle with training complexity and realistic natural images, and adversarial GANs [15] [17], which can be unstable.GANs train a generator to confuse a discriminator that distinguishes generated from real images.Source: §2 · p. 3
• Conditional GANs [15] and coarse-to-fine models extend GANs, while DCGANs [17] improve training stability and generate high-quality images.Recurrent and deconvolutional network architectures have also been proposed for image generation.Source: §2 · p. 3
• GAN advances motivate an image-conditioned converter, unlike class-conditional [15] and attribute-conditional [25] generators; nondeterministic targets preclude the mean-square loss used for video prediction [14], motivating a domain discriminator.The video model conditions on frames to predict subsequent frames and uses mean-square loss to relate inputs to predictions.Source: §2 · p. 3
• Domain adaptation transfers models or learns invariant features, including clothing recognition and retrieval across fashion and unconstrained images, whereas this work generates target images directly from source images.The cited clothing work addresses attribute recognition and product retrieval.Source: §2 · pp. 3–4

• GAN frames image generation as adversarial training between a generator mapping random vectors to pixel-level images and a discriminator distinguishing real from generated images.The cited framework is used for visual data [17] [14].Source: §3 · p. 4
• Training alternates between fitting the discriminator on real and generated images and updating the generator to fool it, repeating until convergence.The discriminator outputs a scalar probability that is high for real images and low for generated ones, and its loss is binary cross-entropy.Source: §3 · p. 4
• GAN’s real/fake supervision can produce realistic images without strong fine-grained supervision, motivating the question of whether it can transfer images between distinct domains.The authors report succeeding at this task and defer the method to Section 4.Source: §3 · pp. 4–5

4 Pixel-Level Domain Transfer

• The task is to transfer a source image into a target-domain image of the same pixel dimensions, using a convolutional converter trained on paired source–target examples.The output is an image rather than a feature vector.Source: §4 · p. 5

4.1 Converter Network

• The end-to-end converter encodes a source into a 64-dimensional semantic code, then decodes it through five fractional-strided convolutional layers into a relevant target image.Its encoder has five convolutional layers and is intended to capture attributes such as category and color for decoding.Source: §4.1 · p. 5

4.2 Discriminator Networks

• MSE can produce blurry images and force the converter toward one target despite the many valid shapes and viewpoints for a source [14].Because pixel-level targets are non-unique, the authors argue that supervision should tolerate target diversity and small geometric misalignment.Source: §4.2 · pp. 5–7
• A real/fake discriminator encourages realistic targets [17], while a domain discriminator penalizes outputs unrelated to their source.The domain discriminator judges source–target pairs, treating ground-truth pairs as associated and inferred or irrelevant targets as unassociated; adversarial training pits the converter against it. The discriminator networks have different objectives, and paired images are stacked across channels.Source: §4.2 · pp. 7–8

4.3 Adversarial Training

• The converter and both discriminators are trained with their respective losses under the adversarial training procedure of.The section defines losses for the real/fake and domain discriminators before describing the training procedure.Source: §4.3 · p. 8

A Update the converter C:

• A target sample is stochastically selected with equal probability from the paired target samples.The selection is performed by sel(·), described as a random selection function with equal probability.Source: A Update the converter C · p. 9
• The real/fake discriminator is trained on the target batch, while the domain discriminator is trained on both source and target batches; their updated parameters are then frozen.These discriminator updates reduce the losses in Eq. (2) and Eq. (5), respectively.Source: A Update the converter C · p. 9
• With the discriminators frozen, the converter is trained to increase the losses of both discriminators.The passage says the converter loss is represented afterward, but its formula is not included in the supplied text.Source: A Update the converter C · p. 9

5.1 LookBook Dataset

• LookBook pairs clean-background top-product images with matching fashion-model images, providing 84,748 images across 9,732 products and about eight model images per product.The images were collected from five online fashion shopping malls; the authors suggest the dataset may support broader domain-adaptation research.Source: §5.1 · pp. 9–10

5.2 Experiment Details

• Images were resized to 64 pixels on the longer side, and 5% of products each were reserved for validation and testing; training used SGD with batch size 128.The learning rate was 0.0002 for 25 epochs, then 0.00002 for five more; baseline training details were identical to the proposed method.Source: §5.2 · p. 10

5.4 Quantitative evaluation by user study

• Our method outperforms the baselines on all user-study criteria, with better attribute and category capture than C+MSE and a higher realism score than C+RF.The authors attribute the improvement to the domain discriminator and its associated/unassociated supervision, which encourages realistic, relevant targets.Source: Table 3 · p. 11, §5.4 · p. 12

5.5 Quantitative evaluation by pixel-level (dis)similarity

• C+MSE has the lowest RMSE, but our method outperforms all baselines on SSIM and outperforms retrieval by DD-score, supporting generalization to unseen items.The authors also report that our method outperforms C+RF+DD−Neg, which omits irrelevant input pairs from domain-discriminator supervision.Source: Table 3 · p. 11, §5.5 · pp. 12–13

5.3 Qualitative evaluation

• Despite varied backgrounds, viewpoints, poses, and self-occlusions, the converter produces less-varying targets that reflect the source clothing attributes and categories.Qualitative test-set results are described as more relevant and realistic than the baselines, with sharp product boundaries and generally well-rendered small details; C+RF looks realistic but irrelevant, while C+MSE is blurry.Source: §5.3 · p. 11

5.4 Quantitative evaluation by user study

• The user study compared our method with the two top baselines using 100 test-set source images, 25 participants, and ratings for realism, attribute capture, and category match.Realism and attribute capture used 0–2 scores; category match used a binary score.Source: §5.4 · pp. 11–12

6 Conclusion

• The domain discriminator trains semantic relations between domains, enabling the converter to generate decent target images.The framework is based on Generative Adversarial Nets.Source: §6 · p. 13
• The paper presents a large dataset intended to contribute to domain adaptation research.Source: §6 · p. 13
• The framework is not limited to specific problems, and the authors expect to extend it from low-level image processing to high-level synthesis.Source: §6 · p. 13