pytorch-image-translation-models 0.2.2

A PyTorch library for multi-modal image translation with diffusion bridges, GANs, and transformer backbones.

pytorch-image-translation-models

A PyTorch library for multi-modal image translation with diffusion bridges, GANs, and transformer backbones.

Installation

Install from PyPI

pip install pytorch-image-translation-models

Install from source

pip install -e .

With optional dependencies:

# With training extras (accelerate, peft, datasets, tensorboard)
pip install -e ".[training]"

# With metrics extras (torchmetrics, lpips, torch-fidelity, scipy)
pip install -e ".[metrics]"

# Everything
pip install -e ".[all]"

Note: PyTorch is listed as a dependency but you may want to install a specific CUDA build first. See PyTorch — Get Started for details.

Features

Models

• GAN generators — UNetGenerator (encoder-decoder with skip connections), ResNetGenerator (residual blocks)
• GAN discriminators — PatchGANDiscriminator (Markovian patch-level classifier)
• StegoGAN — ResnetMaskV1Generator, ResnetMaskV3Generator, NetMatchability (steganographic masking for non-bijective translation, CVPR 2024)
• Diffusion bridge — I2SBUNet (ADM-style U-Net for Image-to-Image Schrödinger Bridge)
• DiT backbone — SiTBackbone (Scalable Interpolant Transformer for diffusion bridges)

Schedulers

Scheduler	Description
I2SBScheduler	Symmetric beta schedule with forward/reverse bridge kernels for I2SB
DDBMScheduler	Karras sigma schedule with Heun/Euler sampling for DDBM (VP/VE modes)
BiBBDMScheduler	Brownian Bridge noise schedule with bidirectional sampling for BiBBDM
DDIBScheduler	Gaussian diffusion with DDIM forward/reverse steps for DDIB
BDBMScheduler	Bidirectional Brownian Bridge schedule for BDBM
DBIMScheduler	Faster bridge sampler with eta-controlled stochasticity for DBIM
CDTSDEScheduler	Dynamic domain-shift eta schedule for CDTSDE
LBMScheduler	Flow-matching bridge for single/few-step LBM translation

Pipelines

Pipeline	Description
I2SBPipeline	End-to-end inference for I2SB models
DDBMPipeline	DDBM bridge diffusion with Heun's method
BiBBDMPipeline	Bidirectional Brownian Bridge translation (b2a / a2b)
DDIBPipeline	Dual-model DDIM encode/decode translation
BDBMPipeline	Bidirectional diffusion bridge with context conditioning
DBIMPipeline	Fast DBIM bridge sampling with bridge preconditioning
CDTSDEPipeline	CDTSDE with dynamic domain-shift scheduling
LBMPipeline	LBM flow-matching for single/few-step image translation

All pipelines support "pt", "pil", and "np" output types.

Data

• PairedImageDataset / UnpairedImageDataset with configurable transform pipelines

Losses

• GANLoss (vanilla / LSGAN / hinge), VGG-based PerceptualLoss

Training

• Pix2PixTrainer — Paired GAN training with checkpoint save/load
• StegoGANTrainer — StegoGAN unpaired training with steganographic masking and consistency losses
• I2SBTrainer — I2SB bridge model training (in examples/i2sb/)

Metrics

• compute_psnr, compute_ssim, compute_lpips, compute_fid

Community Pipelines

Self-contained, single-file modules contributed by the community (inspired by diffusers community pipelines):

Pipeline	Paper	Description
parallel_gan.py	Wang et al., TGRS 2022	SAR-to-Optical with hierarchical latent features

Quick Start

GAN-based translation (Pix2Pix)

import src

gen = src.UNetGenerator(in_channels=3, out_channels=3)
disc = src.PatchGANDiscriminator(in_channels=6)

from src.training import Pix2PixTrainer, TrainingConfig
config = TrainingConfig(epochs=100, device="cuda")
trainer = Pix2PixTrainer(gen, disc, config)
trainer.fit(dataloader)  # expects {"source": tensor, "target": tensor}

translator = src.ImageTranslator(gen, device="cuda")
result = translator.predict(pil_image)

Diffusion bridge translation (I2SB)

from src.models.unet import I2SBUNet, create_model
from src.schedulers import I2SBScheduler
from src.pipelines.i2sb import I2SBPipeline

# Create model and scheduler
model = create_model(
    image_size=256, in_channels=3, num_channels=128,
    num_res_blocks=2, attention_resolutions="32,16,8",
    condition_mode="concat",
)
scheduler = I2SBScheduler(interval=1000, beta_max=0.3)

# Inference pipeline
pipeline = I2SBPipeline(unet=model, scheduler=scheduler)
result = pipeline(source_tensor, nfe=20, output_type="pt")

DDBM bridge diffusion

from src.schedulers import DDBMScheduler
from src.pipelines import DDBMPipeline

scheduler = DDBMScheduler(pred_mode="vp", num_train_timesteps=40)
pipeline = DDBMPipeline(unet=my_unet, scheduler=scheduler)
result = pipeline(source_image, num_inference_steps=40, output_type="pil")

BiBBDM bidirectional translation

from src.schedulers import BiBBDMScheduler
from src.pipelines import BiBBDMPipeline

scheduler = BiBBDMScheduler(num_timesteps=1000, sample_step=100)
pipeline = BiBBDMPipeline(unet=my_unet, scheduler=scheduler)
# Source → Target
result = pipeline(source_tensor, direction="b2a", output_type="pt")
# Target → Source
result = pipeline(target_tensor, direction="a2b", output_type="pt")

DDIB dual-model translation

from src.schedulers import DDIBScheduler
from src.pipelines import DDIBPipeline

scheduler = DDIBScheduler(num_train_timesteps=1000)
pipeline = DDIBPipeline(source_unet=src_model, target_unet=tgt_model, scheduler=scheduler)
result = pipeline(source_image, num_inference_steps=250, output_type="pil")

LBM flow-matching translation

from src.schedulers import LBMScheduler
from src.pipelines import LBMPipeline

scheduler = LBMScheduler(num_train_timesteps=1000)
pipeline = LBMPipeline(unet=my_unet, scheduler=scheduler)
result = pipeline(source_image, num_inference_steps=1, output_type="pil")

DiT backbone (SiT) for diffusion bridges

from src.models.dit import SiTBackbone, SIT_CONFIGS

# Create a SiT-S/2 backbone (small, patch size 2)
depth, hidden_size, num_heads = SIT_CONFIGS["S"]
model = SiTBackbone(
    image_size=256, patch_size=2, in_channels=3,
    hidden_size=hidden_size, depth=depth, num_heads=num_heads,
    condition_mode="concat",
)
# Use as drop-in replacement for UNet in any bridge pipeline
output = model(noisy_sample, timestep, xT=source_image)

I2SB training with task configs

from examples.i2sb.config import sar2eo_config
from examples.i2sb.trainer import I2SBTrainer

cfg = sar2eo_config(resolution=256, train_batch_size=8)
trainer = I2SBTrainer(cfg)
model = trainer.build_model()
scheduler = trainer.build_scheduler()

# Single-step loss computation
loss = I2SBTrainer.compute_training_loss(model, scheduler, source_batch, target_batch)
loss.backward()

StegoGAN non-bijective translation

from src.training import StegoGANTrainer, StegoGANConfig

cfg = StegoGANConfig(
    input_nc=3, output_nc=3, ngf=64,
    lambda_reg=0.3, lambda_consistency=1.0,
    resnet_layer=8, fusionblock=True,
    device="cuda",
)
trainer = StegoGANTrainer(cfg)
# Run a single training step with unpaired data
losses = trainer.train_step(real_A_batch, real_B_batch)

Package Structure

src/                                 # ← Core library (single source of truth)
├── __init__.py                      # Public API
├── models/
│   ├── generators.py                # UNetGenerator, ResNetGenerator
│   ├── discriminators.py            # PatchGANDiscriminator
│   ├── unet/
│   │   ├── i2sb_unet.py            # I2SBUNet (native ADM-style backbone)
│   │   ├── unet_2d.py              # create_model factory
│   │   └── diffusers_wrappers.py   # DDBMUNet, DDIBUNet, … (diffusers UNet2DModel wrappers)
│   ├── dit/
│   │   └── sit.py                  # SiTBackbone (Diffusion Transformer)
│   └── stegogan/
│       ├── generators.py           # ResnetMaskV1Generator, ResnetMaskV3Generator
│       └── networks.py             # NetMatchability, mask_generate, ResnetBlock
├── schedulers/                      # One scheduler per method
│   ├── i2sb.py                     # I2SBScheduler
│   ├── ddbm.py                     # DDBMScheduler
│   ├── bibbdm.py                   # BiBBDMScheduler
│   ├── ddib.py                     # DDIBScheduler
│   ├── bdbm.py                     # BDBMScheduler
│   ├── dbim.py                     # DBIMScheduler
│   ├── cdtsde.py                   # CDTSDEScheduler
│   └── lbm.py                      # LBMScheduler
├── pipelines/                       # One pipeline per method
│   ├── i2sb.py                     # I2SBPipeline
│   ├── ddbm.py                     # DDBMPipeline
│   ├── bibbdm.py                   # BiBBDMPipeline
│   ├── ddib.py                     # DDIBPipeline
│   ├── bdbm.py                     # BDBMPipeline
│   ├── dbim.py                     # DBIMPipeline
│   ├── cdtsde.py                   # CDTSDEPipeline
│   └── lbm.py                      # LBMPipeline
├── data/
│   ├── datasets.py                 # PairedImageDataset, UnpairedImageDataset
│   └── transforms.py               # get_transforms, default_transforms
├── losses/
│   ├── adversarial.py              # GANLoss
│   └── perceptual.py               # PerceptualLoss
├── training/
│   ├── trainer.py                  # Pix2PixTrainer, TrainingConfig
│   └── stegogan_trainer.py         # StegoGANTrainer, StegoGANConfig
├── inference/
│   └── predictor.py                # ImageTranslator
└── metrics/
    └── image_quality.py            # PSNR, SSIM, LPIPS, FID
examples/                            # ← Training/inference scripts (import from src/)
├── community/                       # Community-contributed pipelines (single-file)
│   └── parallel_gan.py             # Parallel-GAN (Wang et al., TGRS 2022)
├── i2sb/                            # I2SB paper-oriented training code
│   ├── config.py                   # TaskConfig, sar2eo_config, etc.
│   └── trainer.py                  # I2SBTrainer
└── inference/
    └── run_inference.py            # Unified inference script for all methods

Credits

Reference papers

• I2SB: Image-to-Image Schrödinger Bridge (ICML 2023)
• DDBM: Denoising Diffusion Bridge Models (ICLR 2024)
• DDIB: Dual Diffusion Implicit Bridges (ICLR 2023)
• BBDM: Image-to-Image Translation with Brownian Bridge Diffusion Models (CVPR 2023)
• DBIM: Diffusion Bridge Implicit Models (2024)
• LBM: Latent Bridge Matching for Fast Image-to-Image Translation (2025)
• SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers (2024)
• StegoGAN: Leveraging Steganography for Non-Bijective Image-to-Image Translation (CVPR 2024)
• Parallel-GAN: SAR-to-Optical Image Translation with Hierarchical Latent Features (TGRS 2022)
• CUT: Contrastive Unpaired Translation (ECCV 2020)
• CycleGAN (ICCV 2017)
• img2img-turbo (2024)

License

MIT