---
kind: concept
slug: image-to-image-translation
title: Image-to-image translation
summary: >-
  Maps images from one domain to another while keeping the input's content. Covers ten variants: unpaired and paired GANs, one multi-domain GAN with style codes, training-free diffusion editing, flow-based style transfer, single-pair ViT-feature generation, and translation used for segmentation adaptation or anomaly detection. The choice depends on paired data, labels, and model access.
tags: [image-generation, gan, diffusion, medical-imaging, domain-transfer, multi-domain, fashion, segmentation, style-transfer, anomaly-detection, remote-sensing, vision-transformer]
papers: ["1611.02200", "1612.05362", "2211.12572", "1710.07346", "1903.12212", "2103.16877", "2201.00424", "1905.03198", "1904.07595", "1912.01865"]
related: [generative-adversarial-network, domain-adaptation, auto-context-model, diffusion-models, ddim-inversion, classifier-free-guidance, text-guided-image-editing, semantic-segmentation, adaptive-instance-normalization, medical-image-synthesis, anomaly-detection, neural-style-transfer, normalizing-flows, vision-transformer, self-supervised-learning]
status: draft
updated_at: 2026-10-19
---

## Overview

Image-to-image translation takes an image from a source domain and produces an
image in a target domain (another style, another imaging modality, another
rendering, another outfit, another sensor's view of the same city) that still
reflects the content of the input. The ten reuse-licensed cards agree on the
goal but differ in what ties the output to the input. The unpaired transfer
network keeps a fixed feature extractor's output unchanged [arxiv:1611.02200 §4].
The paired medical generator penalises pixel and gradient differences to a
ground-truth target [arxiv:1612.05362 §2.1]. The multi-domain generator uses a
cycle term to keep the input's structure while a style code sets the target
look [arxiv:1912.01865 §2.2]. The training-free method copies
internal features from a reconstruction run into the translation run
[arxiv:2211.12572 §4]. The clothing generator constrains its output with a
coarse map of the input's body layout [arxiv:1710.07346 §3.2]. The segmentation
method reads only a structure code of the input, so the translated image keeps
the input's layout [arxiv:1903.12212 §3.2]. The style method makes the transfer
invertible, so the input's content can be recovered [arxiv:2103.16877 §4.1]. The
single-pair method trains a new generator per image pair, and its losses come
from a frozen ViT [arxiv:2201.00424 §3.3]. The aerial method translates labelled
source images so that a segmenter can be fine-tuned on them [arxiv:1905.03198
§4.2]. The anomaly method is not a translation goal in itself: it uses a
label-to-image generator as a probe [arxiv:1904.07595 §3].

Most of the trained GAN variants (unpaired, paired, multi-domain, clothing and aerial) combine
an adversarial term that pushes outputs toward the target distribution with a
term that ties each output to its input. Without the second term, the output can
look realistic while losing the layout of the input. The GAN-only baseline in
the unpaired paper shows this: it reaches 13.71% on the digit transfer task
[arxiv:1611.02200 tab. 1]. Three variants have no adversarial term. The ArtFlow
training uses only feature losses [arxiv:2103.16877 §5.1], the single-pair
method uses frozen-ViT feature losses and no discriminator [arxiv:2201.00424
§3.3], and the training-free method has no training at all [arxiv:2211.12572
§4].

## Variants

### Unpaired transfer with a frozen feature extractor (Domain Transfer Network)

No paired examples are needed. The network learns G = g ∘ f from unlabelled
samples of two domains, where f is a fixed function. Only g and the
discriminator are trained, and the constraints are an f-constancy term on source
samples and an identity term on target samples [arxiv:1611.02200 §4] [eq. 5]
[eq. 6]. On the SVHN-to-MNIST digit task, an MNIST classifier reads 90.66% of the
transferred digits correctly [arxiv:1611.02200 tab. 1]. These are author-reported
numbers. The digit data are public, but the face experiment relies on a
non-public feature network and face set [arxiv:1611.02200 §5.2]. Card:
[Unsupervised Cross-Domain Image Generation](/papers/1611.02200).

### One generator for many domains with style codes (StarGAN v2)

This variant is for one model that translates between several domains, where
each input should yield several outputs. Separate models would need one
generator per ordered pair of domains, K(K-1) for K domains [arxiv:1912.01865
§1]. A single generator takes a style code. The code comes from a mapping
network on a random Gaussian latent, or from a style encoder on a reference
image, and it is injected by AdaIN [arxiv:1912.01865 §2.1] [app. E]. Training
adds a style-reconstruction term, a diversity term and a cycle term to the
adversarial loss [arxiv:1912.01865 §2.2] [eq. 5]. The diversity weight decays
linearly to zero over the 100K training iterations [arxiv:1912.01865 app. B].
Author-reported latent-guided FID is 13.7 on CelebA-HQ, against 33.1 for MSGAN,
and 16.2 on AFHQ, against 61.4 for MSGAN [arxiv:1912.01865 tab. 2]. With
reference images, FID is 23.8 and 19.8, against 39.6 and 69.8 for MSGAN
[arxiv:1912.01865 tab. 3]. The user study has 76 valid workers and no
significance test [arxiv:1912.01865 §3.2] [tab. 4]. The official code is
licensed CC BY-NC 4.0, so it cannot be reused commercially. Card:
[StarGAN v2: Diverse Image Synthesis for Multiple Domains](/papers/1912.01865).

### Paired 3D GAN with an auto-context stack (MR to CT)

This variant needs paired, voxel-aligned MR and CT volumes of the same subject
[arxiv:1612.05362 §3]. A 3D fully convolutional generator maps a 32³ MR patch to
a 16³ CT patch, trained with an adversarial loss, an L2 loss and a 3D
gradient-difference loss, with weights 0.5, 1 and 1 [arxiv:1612.05362 §2.1]
[§4]. Several such GANs are chained by the Auto-Context Model, so that later
stages see the previous CT estimate as a second input channel; two iterations
are used [arxiv:1612.05362 §2.2] [§4.2]. Author-reported PSNR on 16 brain
subjects is 27.6, against 26.3 for the structured random forest baseline
[arxiv:1612.05362 tab. 1]. On 22 pelvis subjects it is 34.1, against 32.1
[arxiv:1612.05362 tab. 2]. The gain from the GAN over a plain FCN is 24.7 to 25.9
PSNR in an ablation whose dataset is not stated [arxiv:1612.05362 §4.1]. The
gains from more auto-context iterations are only plotted [arxiv:1612.05362 §4.2]
[fig. 5]. The released repository does not implement the auto-context stage, so
the paper's pipeline must be rebuilt from the text [arxiv:1612.05362 §2.2].
Card: [Medical Image Synthesis with Context-Aware Generative Adversarial Networks](/papers/1612.05362).

### Training-free editing of a frozen diffusion model (Plug-and-Play)

No training is done. The guidance image is inverted with the empty prompt to a
noise latent x_T. The translation is then sampled from the same x_T with the
target prompt, while a decoder feature and the self-attention queries and keys
are replaced by those of the guidance run during the early steps
[arxiv:2211.12572 §4] [alg. 1]. The defaults are 50 sampling steps, with the
feature injected for 40 of them and attention for 25 [arxiv:2211.12572 app. C].
On the real-image benchmark, the structure distance rises from 0.058 to 0.090
without feature injection and to 0.097 without self-attention injection
[arxiv:2211.12572 tab. 1]. The main comparison with SDEdit, Prompt-to-Prompt,
DiffuseIT and VQGAN-CLIP is plotted, with no numbers given [arxiv:2211.12572
§5.1] [fig. 9]. Card: [Plug-and-Play](/papers/2211.12572).

### Two-stage text-conditioned clothing synthesis (Be Your Own Prada)

The input is a photo of a person and a text description of a new outfit. The
method needs no pairs of one person in two outfits; it trains on one photo and
one sentence per person [arxiv:1710.07346 §3.1]. The photo is reduced to a
seven-label segmentation map, then merged to four classes and down-sampled to
8x8. This keeps the body shape and pose, while the old garment's outline is
lost [arxiv:1710.07346 §3.2]. Stage 1 generates a new label map from noise, that
constraint and a design code made of attributes and the text embedding
[arxiv:1710.07346 §3.2] [eq. 2]. Stage 2 renders one channel per label and takes
each pixel from the channel of its label [arxiv:1710.07346 §3.3] [eq. 4]. The two
GANs are trained separately, because the argmax between them is not
differentiable [arxiv:1710.07346 §3.4]. Author-reported attribute mAP is 82.6,
against 71.8 to 74.7 for the baselines and 86.8 for real photos
[arxiv:1710.07346 tab. 1]. In a 50-volunteer ranking, the method has mean rank
1.544, against 2.286 for the best non-GAN baseline [arxiv:1710.07346 tab. 2].
Output is 128x128. Card: [Be Your Own Prada](/papers/1710.07346).

### Structure-texture split with cross-domain swap for segmentation (DISE)

This variant is a domain adaptation method for semantic segmentation, and
translation is its mechanism. Each image is encoded into a domain-invariant
structure code and a domain-specific texture code [arxiv:1903.12212 §3.1]. The
segmenter reads only the structure code [arxiv:1903.12212 §3.2]. Swapping
textures between a labelled source image and an unlabelled target image gives
translated images that inherit the source labels [arxiv:1903.12212 §3.1]
[§3.2]. Target labels are never used in the loss [arxiv:1903.12212 §3.2]. The
paper does not state the loss weights [arxiv:1903.12212 §3.2] [eq. 1]. The
author-reported result on GTA5 to Cityscapes is 45.4 mIoU, with a DeepLab-v2
ResNet-101 base, against 42.4 for the output-space adaptation of Tsai et al.
[arxiv:1903.12212 tab. 2]. The ablation gives 39.8 for source-only training, 44.1
without the label-transfer term and 45.4 with it [arxiv:1903.12212 tab. 4]. Card:
[All about Structure](/papers/1903.12212).

### Reversible-flow style transfer (ArtFlow)

Universal style transfer normally uses a fixed encoder, a style step and a
learned decoder. The authors argue that the decoder does not invert the encoder
exactly, so repeated stylization loses content, which they call content leak
[arxiv:2103.16877 §3.2]. ArtFlow replaces the encoder and decoder with one
invertible network, a Glow-style flow of two blocks with eight flows each
[arxiv:2103.16877 §5.1]. AdaIN or WCT acts on the latent features, and the
reverse pass returns pixels [arxiv:2103.16877 §4.1]. Training uses a content loss
and a feature-statistics style loss on VGG19 features, weighted 0.1 and 1
[arxiv:2103.16877 §5.1]. The content-leak evidence comes from 20 rounds of
repeated transfer, shown qualitatively [arxiv:2103.16877 §5.3]. Author-reported
SSIM is 0.45 for ArtFlow with AdaIN, against 0.29 for AdaIN alone
[arxiv:2103.16877 tab. 1]. At 512×512, its time is 0.408 s, against 0.064 s for
AdaIN [arxiv:2103.16877 tab. 1]. Card: [ArtFlow](/papers/2103.16877).

### Single-pair generator with frozen ViT feature losses (Splice)

The inputs are one structure image and one appearance image. No collection of
either domain is needed. A small U-Net is trained from scratch on that pair
[arxiv:2201.00424 §3] [app. A.1]. A frozen DINO ViT-B/8 is used only to compute
losses, and there is no discriminator [arxiv:2201.00424 §3] [app. A.2]. The
appearance term matches the deepest CLS token of the output to that of the
appearance image [arxiv:2201.00424 eq. 5]. The structure term matches the
self-similarity of the deepest-layer keys of the output to that of the structure
image [arxiv:2201.00424 eq. 2] [eq. 6]. An identity term keeps the keys of the
appearance image when the generator is fed that image [arxiv:2201.00424 eq. 7].
The structure and identity weights are both 0.1 [arxiv:2201.00424 §3.3]. Training
runs 2000 iterations at a constant learning rate of 2 × 10⁻³, about 20 minutes
for a 512 × 512 pair on one GPU [arxiv:2201.00424 app. A.3]. Author-reported
two-alternative forced-choice preference on Wild-Pairs is 79.0% against STROTSS
and 83.1% against WCT² [arxiv:2201.00424 tab. 1]. Because a generator is trained
per pair, a new pair needs new training [arxiv:2201.00424 §3]. The method fails
when the two images have no semantic match the ViT can use [arxiv:2201.00424
§4.3] [fig. 10]. Card: [Splicing ViT Features for Semantic Appearance
Transfer](/papers/2201.00424).

### Translation of labelled source images for aerial segmentation (CycleGAN-style)

No target labels are used to train the translator or the segmenter. An unpaired
translator with two generators and two discriminators is trained on the source
(Potsdam) and target (Vaihingen) image sets [arxiv:1905.03198 §4.1]. The
source-to-target generator restyles every labelled source image, and each
translated image keeps its source label [arxiv:1905.03198 §4.2]. A segmenter
trained on the source is then fine-tuned on the restyled images, starting from
its source weights [arxiv:1905.03198 §4.3] [eq. 9]. The translator loss has an
adversarial term for each direction and a reconstruction term that asks each
generator to undo the other [arxiv:1905.03198 eq. 3] [eq. 8]. The paper states no
weight for the reconstruction term [arxiv:1905.03198 eq. 8]. Author-reported
average accuracy on Vaihingen rises from 0.35 to 0.52 [arxiv:1905.03198 tab. 3].
The largest gains are for Building, from 0.23 to 0.71, and Tree, from 0.06 to
0.51. Low vegetation falls from 0.38 to 0.27 [arxiv:1905.03198 tab. 4].
Fine-tuning reports the best epoch on the target test set, so these numbers are
optimistic [arxiv:1905.03198 §5.2.4]. All results come from one source and target
pair and one segmenter [arxiv:1905.03198 §6]. Card: [Unsupervised Domain
Adaptation using Generative Adversarial Networks for Semantic Segmentation of
Aerial Images](/papers/1905.03198).

### Resynthesis as a probe for unexpected objects (discrepancy network)

This variant does not aim at a new image. A segmenter labels a street scene, a
pix2pixHD generator repaints the scene from those labels alone, and a
discrepancy network marks the pixels where the original and the repainting
disagree [arxiv:1904.07595 §3] [fig. 2]. The discrepancy network reads the
original image, the repainting and the label map. Image features come from a
pre-trained VGG16 with shared weights [arxiv:1904.07595 §3.1] [fig. 3]. It is
trained on repaintings of scenes in which some known object instances were
relabelled at random. Its target is the set of pixels whose label changed, so
it never needs unknown classes [arxiv:1904.07595 §3.2] [fig. 4]. The paper does
not state what fraction of instances is changed [arxiv:1904.07595 §3.2]. On
Lost and Found, the author-reported AUROC is 0.82 for all pixels except the
ego-vehicle and 0.92 to 0.93 for the road-only region. The RBM baseline reaches
0.74 and 0.86 [arxiv:1904.07595 §4.2] [fig. 5]. On a 60-image road anomaly set,
the AUROC is 0.79 to 0.83, against 0.59 for the RBM baseline [fig. 5]. The same
repainting also feeds an attack detector. It takes the L2 distance between HOG
descriptors of the input and of the repainting, and a logistic regression reads
that distance [arxiv:1904.07595 §3.3] [tab. 1]. Each frame needs one segmenter
pass, one generator pass and two VGG16 passes, and the paper gives no timing
[arxiv:1904.07595 §3.1]. Card: [Detecting the Unexpected via Image
Resynthesis](/papers/1904.07595).

## When to use which

| Situation | Use | Why |
|---|---|---|
| Paired, voxel-aligned volumes of two modalities, from a few dozen subjects | Paired 3D GAN with L2 and gradient-difference terms [arxiv:1612.05362 §2.1] | It is the only variant here with direct per-voxel supervision. Its reported PSNR is above the structured random forest baseline on 16 and 22 subjects, but the samples are small [arxiv:1612.05362 tab. 1] [tab. 2] |
| No paired data, and a pretrained extractor on the source domain whose representation must be kept | Unpaired transfer with a frozen extractor [arxiv:1611.02200 §4] | It trains from unlabelled samples of both domains, and the constancy term keeps the extractor's representation [arxiv:1611.02200 eq. 5] |
| Several domains with one model, unpaired images, and outputs that should vary for the same input | Multi-domain style-code GAN [arxiv:1912.01865 §2.1] [§2.2] | One generator serves all domains, and the diversity term gives different outputs per noise or reference. The paper's FID gains are against baselines trained per domain pair [arxiv:1912.01865 §3] [tab. 2] |
| No training budget, a frozen text-to-image model is available, and the guidance image shows content the text can name | Training-free diffusion editing [arxiv:2211.12572 §4] | It needs no training, but each real image costs about 1000 inversion and 1000 reconstruction steps before translation [arxiv:2211.12572 app. C] |
| The guidance image is a silhouette or a mask with arbitrary colours | Expect weak results from the diffusion method | The paper names this as a limitation, because the input carries no meaning the model can use [arxiv:2211.12572 §6] |
| Change clothing from a text description, keep the person's pose and body shape, and no paired photos exist | Two-stage text-conditioned GAN [arxiv:1710.07346 §3.2] [§3.3] | The coarse merged map carries pose and body shape. Arms and legs are merged into the "rest" class, so limb changes are not controlled by the constraint [arxiv:1710.07346 §3.2] |
| Segmentation on a target domain with no labels, a labelled source domain, and a shift mainly in texture | Structure-texture swap [arxiv:1903.12212 §3.1] [§3.2] | Target labels are not used. The reported gain over the output-space baseline is 3.0 mIoU on one benchmark [arxiv:1903.12212 tab. 2] |
| Labelled source images, unlabelled target images, and a shift mainly in sensor or appearance; GAN training and a second fine-tuning pass are affordable | CycleGAN-style source translation [arxiv:1905.03198 §4.2] | No target labels are used for training. The gain is measured on one pair and one segmenter, and the best epoch was picked on target test labels [arxiv:1905.03198 §5.2.4] [§6] |
| Stylization that is applied repeatedly, or must allow recovery of the input, with a budget near 0.4 s per 512×512 image | Reversible-flow style transfer [arxiv:2103.16877 §4.1] | The reverse pass recovers the content, but the method is slower than AdaIN and is not best on Gram loss [arxiv:2103.16877 tab. 1] |
| One structure image and one appearance image, no dataset of either domain, and about 20 minutes of GPU time per pair | Single-pair ViT-feature generator [arxiv:2201.00424 app. A.3] | No discriminator and no segmentation or correspondences are needed. Each new pair needs a new generator, and the method fails when the objects have no semantic match [arxiv:2201.00424 §3] [§4.3] |
| Flag unexpected objects in a street scene without training on them, and a segmenter and a generator for the same domain are available | Resynthesis discrepancy network [arxiv:1904.07595 §3] [§3.2] | Training uses only relabelled known objects. The authors name false positives as the main remaining issue [arxiv:1904.07595 §5] |
| You need code you can reuse directly | Check the licence first | The Plug-and-Play repository has no LICENSE file, so re-implement from the paper [arxiv:2211.12572 §4]. The medical repository's code differs from the paper's architecture [arxiv:1612.05362 §4]. The released DISE code is GPL-3.0 and the ArtFlow repository has no licence, so both papers' code should be read as reference only. The resynthesis code is GPL-3.0-or-later and omits the attack detector [arxiv:1904.07595 §4.3]. The Splice and aerial repositories have no licence file. The StarGAN v2 repository is CC BY-NC 4.0, so implement from the paper [arxiv:1912.01865 §2]. |

## Implementation sketch

The adversarial core of the trained GAN variants (unpaired, paired, clothing and
aerial), in simplified form. The unpaired paper uses a three-class discriminator
and the paired paper uses a binary one, so this sketch follows the paired
paper's binary form. The flow, single-pair, resynthesis and training-free
variants do not use this loop.

```pseudocode
# Shared core of trained translation  [arxiv:1612.05362 eq. 3] [eq. 5]
# Input:  source batch x, target batch y, tie_loss (variant-specific)
# Output: generator G

def translation_step(G, D, x, y, tie_loss):
    # Discriminator: real targets get label 1, generated outputs get 0
    fake = G(x)
    loss_d = bce(D(y), 1) + bce(D(fake), 0)
    update(D, loss_d)
    # Generator: fool D and stay tied to the input  [arxiv:1612.05362 eq. 5]
    loss_g = adv_weight * bce(D(fake), 1) + tie_loss(x, fake, y)
    update(G, loss_g)
```

Where the variants differ:

- **Unpaired transfer.** The tie term compares features of the source input
  and the output under a frozen extractor, and adds an identity term on target
  samples [arxiv:1611.02200 eq. 5] [eq. 6].
- **Multi-domain style-code GAN.** The tie is a cycle term that maps the output
  back with the input's own estimated style. Two more terms, style
  reconstruction and a diversity term, are added to the same objective
  [arxiv:1912.01865 eq. 2] [eq. 3] [eq. 4].
- **Paired 3D GAN.** The tie term is an L2 loss to the paired CT patch plus a
  gradient-difference loss [arxiv:1612.05362 eq. 5] [eq. 6] [eq. 7].
- **Clothing, two stages.** The tie is the coarse shape-map constraint. Each
  stage is a conditional GAN, trained on its own tuple, and the second stage
  composes pixels from per-label channels [arxiv:1710.07346 §3.2] [§3.3] [§3.4].
- **Structure-texture swap.** The tie is a texture swap with reconstruction
  losses. A label-transfer term gives translated images the source labels
  [arxiv:1903.12212 §3.2].
- **Aerial translator.** Two generators and two discriminators replace the single
  pair. Each generator's output must map back to its input under the other
  generator, and the reconstruction term is weighted [arxiv:1905.03198 eq. 6]
  [eq. 7] [eq. 8]. The source labels are then reused for the translated images,
  and the segmenter is fine-tuned on them [arxiv:1905.03198 §4.2] [§4.3].
- **Single-pair ViT features.** No discriminator. The tie is a sum of feature
  losses from a frozen ViT: a CLS-token term for appearance, a self-similarity
  term for structure and a key identity term [arxiv:2201.00424 eq. 4].
- **Reversible flow.** No adversarial term. The tie is a feature-statistics loss
  on the reversed output, and the style step runs in latent space
  [arxiv:2103.16877 §4.1] [§5.1].
- **Resynthesis probe.** Not a translation loop. A label-to-image generator
  repaints the scene, and a discrepancy network scores each pixel from the
  original, the repainting and the labels [arxiv:1904.07595 §3] [§3.1].
- **Training-free diffusion.** There is no training step. The tie comes from
  injecting the guidance run's features and attention during sampling, not from
  a loss [arxiv:2211.12572 alg. 1].

## Contradictions

No two papers report incompatible numbers for the same quantity. The ten papers
measure different things: classifier accuracy, PSNR, structure distance, SSIM,
attribute mAP, mIoU, AUROC, average accuracy, FID and a preference rate. They
also use different data. Their numbers should not be placed side by side as if
they measured the same quantity [arxiv:1611.02200 tab. 1] [arxiv:1612.05362 tab. 1]
[arxiv:2211.12572 tab. 1] [arxiv:1710.07346 tab. 1] [arxiv:1903.12212 tab. 2]
[arxiv:2103.16877 tab. 1] [arxiv:2201.00424 tab. 1] [arxiv:1905.03198 tab. 3]
[arxiv:1904.07595 §4.2] [arxiv:1912.01865 tab. 2].

One paper is internally inconsistent. The DISE GTA5 table gives the method of
Tsai et al. 42.4 mIoU, while the ablation calls the same method "seg-map
adaptation" at 42.6 [arxiv:1903.12212 tab. 2] [tab. 4] [§4.3]. Use the table
value only with that caveat.

The aerial paper also gives different "before" figures. The abstract and its
results table give 0.35 for the source model on Vaihingen, while the fine-tuning
text gives 34% and 39% to 52% [arxiv:1905.03198 tab. 3] [§5.2.4]. The gain is
quoted both as 17 points and as 5% to 17% [arxiv:1905.03198 §5.2.4] [§6]. Its
discussion says unaffected classes stay the same, but the per-class table shows
Low vegetation falling [arxiv:1905.03198 §6] [tab. 4]. Use the table values only
with those caveats.

## Papers

- [Unsupervised Cross-Domain Image Generation](/papers/1611.02200) (Taigman et al.): unpaired transfer that keeps a frozen extractor's representation, shown on digits and faces.
- [Medical Image Synthesis with Context-Aware Generative Adversarial Networks](/papers/1612.05362) (Nie et al.): paired 3D GAN with auto-context stages for MR-to-CT synthesis.
- [StarGAN v2: Diverse Image Synthesis for Multiple Domains](/papers/1912.01865) (Choi et al.): one generator for several domains, with style codes from a mapping network or a reference image.
- [Plug-and-Play](/papers/2211.12572) (Tumanyan et al.): training-free translation with a frozen Stable Diffusion, using feature and attention injection.
- [Be Your Own Prada: Fashion Synthesis with Structural Coherence](/papers/1710.07346) (Zhu et al.): two-stage conditional GAN that redresses a person from a text description, through a segmentation map.
- [All about Structure: Adapting Structural Information across Domains for Boosting Semantic Segmentation](/papers/1903.12212) (Chang et al.): structure and texture split with cross-domain swap for unsupervised segmentation adaptation.
- [ArtFlow: Unbiased Image Style Transfer via Reversible Neural Flows](/papers/2103.16877) (An et al.): universal style transfer with an invertible flow that reduces content leak under repeated stylization.
- [Splicing ViT Features for Semantic Appearance Transfer](/papers/2201.00424) (Tumanyan et al.): single-pair generator trained with frozen DINO ViT losses, with no adversarial training.
- [Unsupervised Domain Adaptation using Generative Adversarial Networks for Semantic Segmentation of Aerial Images](/papers/1905.03198) (Benjdira et al.): CycleGAN-style translation of labelled source patches to the target sensor look, followed by segmenter fine-tuning.
- [Detecting the Unexpected via Image Resynthesis](/papers/1904.07595) (Lis et al.): resynthesis from a segmenter's label map, compared with the image by a discrepancy network, for unexpected-object detection.