ParticleGAN

Learnable particle priors and GAN building blocks for PyTorch.

API reference · Minimal GAN loop · Minimal DDGAN + UCD loop

Installation

Requires Python 3.10+ and PyTorch. Install from PyPI:

python -m pip install particlegan

For development and the repository's research experiments:

git clone https://github.com/255BITS/ParticleGAN.git
cd ParticleGAN
python -m pip install -e '.[experiments,dev]'
# Image experiments also need the images extra:
# python -m pip install -e '.[experiments,images,dev]'

CI tests Python 3.10–3.12 and builds installable distributions. See CI and PyPI releases for the automated publishing setup.

Use in your PyTorch project

Use individual components in your existing loop. You own the networks, data, optimizers, backward calls, devices, logging, and checkpoints. No trainer is required, and the loss helpers never call backward or step an optimizer.

See the minimal GAN loop, minimal DDGAN + UCD loop, and API reference for complete examples and contracts.

Our examples and experiment trainers consume these same public primitives and recipe factories. See the migration and compatibility checks for existing-config GPU smoke tests and checkpoint comparisons.

The single-transition example defaults to the winning MisGAN-inspired encoder: G1 -> st, G2 -> at, G3 -> st+1, plus E(st, at) -> z -> G3 -> st+1. It uses 1,024 MoG components, bcap, joint/action critics and a shared state critic. Run python -u examples/transition_gan.py. See the toy demo, architecture and losses, and leaderboard.

from particlegan import ParticlePrior, GANLoss, GradientPenalty, ParticleRegularizer

prior = ParticlePrior().to(device)  # 20,000 learnable particles, z_dim=4
adversarial = GANLoss()            # relativistic-paired logistic loss
penalty = GradientPenalty()       # exact L2 cap penalty, weight=1, cap=1
spread = ParticleRegularizer()    # variance/covariance regularization, weight=1

# Customize with ordinary keyword arguments:
prior = ParticlePrior(num_particles=4096, z_dim=16).to(device)
adversarial = GANLoss(loss_type="hinge", mode="vanilla")

prior.sample(batch_size) returns (z, indices), with z shaped [B, z_dim]. Include prior.parameters() in your generator optimizer to learn the particles. Use GaussianPrior(z_dim=16) for fresh Gaussian samples with the same sampling interface; its indices are None and it needs no particle regularization.

Mixture-of-Gaussians particles (0.3.0)

MoGParticlePrior gives each learned particle a continuous Gaussian neighborhood. Its defaults use the selected compact configuration: 400 components, z_dim=4, sigma_rel=1/40, standardized means. Sigma is shared and fixed after calibration.

from particlegan import MoGParticlePrior, get_recipe

prior = MoGParticlePrior().to(device)
z, component_ids = prior.sample(256)

recipe = get_recipe("mog")  # 400 components, 28k steps, prior LR 0.06, beta1=0.5
prior = recipe.make_prior().to(device)
opt_g, opt_d = recipe.make_optimizers(G, D, prior)
# Regularize raw means, not noisy draws. Use the full table at N <= 1024.
prior_loss = recipe.make_prior_regularizer()(prior.z)

Use ParticlePrior for atoms. Existing GAN/DDGAN presets and examples keep their defaults; mog and ddgan_mog are dedicated opt-in recipes. The MoG recipe's prior_lr_mult=100 is relative to G's LR, giving 0.06. It is equivalent to the experiment's 10× multiplier on the original prior LR of 0.006.

Run the default MoG benchmark with the existing experiment trainer:

python -u experiments/train_100gaussians.py --config configs/mog/default.toml
# During training:
tail -F results/mog/default/log.txt

The 400-component model passed the C0 acceptance envelope at 28k steps on the 100-Gaussian benchmark, using 50× fewer components and 4× the original updates. This is a single-seed result; both MoG and atoms remain useful. Results and tradeoffs · MoG API and checkpoint details · Changelog.

Core MoG sampling and calibration work with PyTorch alone. For faster calibration of large low-dimensional tables, install the optional extra with python -m pip install 'particlegan[mog]' (or python -m pip install -e '.[mog]' from this checkout).

Particle AE-GAN, VAE-GAN and AE-DDGAN (0.5.0)

from particlegan import get_recipe

recipe = get_recipe("vae_gan")  # also: ae_gan, ae_ddgan
prior = recipe.make_prior()
# query = E(x): [batch, recipe.z_dim], produced by your encoder
# encoded = recipe.encode(query, prior)
# x_hat = G(encoded.codes)  # [batch, draws, observed dimensions]
# loss = encoded.reconstruction_loss(x_hat, x)

vae_gan selects one particle and adds fixed-sigma Gaussian noise. Its joint KL is constant, so it needs no KL regularizer in training. AE uses a deterministic bounded offset instead. You own the networks, loop and loss composition; reconstruction never silently adds KL. See the guide and runnable example for explicit ELBO reporting, optional categorical inference, AE-DDGAN integration and measured limits. Genuine VAE results are toy-only; AE-GAN and AE-DDGAN have matched CIFAR32 evidence.

DDGAN with MoG particles (0.4.0)

Select the combined recipe with one name:

from particlegan import DDGAN, get_recipe

recipe = get_recipe("ddgan_mog")
prior = recipe.make_prior().to(device)
process = DDGAN(recipe.alpha_bar).to(device)
opt_g, opt_d = recipe.make_optimizers(G, D, prior)

z, indices = prior.sample(batch_size)
prior_loss = recipe.make_prior_regularizer()(prior.z[indices.unique()])

Defaults: 400 components, z_dim=4, sigma_rel=0.025, standardized reads, 100k updates, constant LR, prior LR 0.06, prior betas (0.5, 0.999), and the existing four-step DDGAN/class-only UCD settings. Override any recipe field with keywords, for example get_recipe("ddgan_mog", num_classes=8, z_dim=16).

These are the package hyperparameters from the 100k study, which used a 32-wide generator and 128-wide discriminator. You supply the networks, data, training/sampling loop and EMA; the recipe does not select architectures.

Add a loss to an existing pipeline

Components are independent. For example, add a critic penalty or a particle spread term to losses your pipeline already computes:

# In your discriminator update:
d_loss = existing_d_loss + penalty(D, real, fake.detach(), step=step)

# In your generator/prior update, when sampled particle indices are available:
g_loss = existing_g_loss + spread(prior.z[indices.unique()])
# Your code calls backward() and optimizer.step().

To use the adversarial objective itself, call adversarial.d_loss(D(real), D(fake.detach())) for D and adversarial.g_loss(D(fake), D(real).detach()) for G. Freeze D's parameters for the G update while retaining gradients through D(fake).

Selected defaults, with easy overrides

from particlegan import get_recipe

recipe = get_recipe()  # Recommended GAN defaults.
recipe = recipe.replace(z_dim=16, num_particles=4096, lr=3e-4)
prior = recipe.make_prior().to(device)
adversarial = recipe.make_loss()
penalty = recipe.make_gradient_penalty()
spread = recipe.make_prior_regularizer()
opt_g, opt_d = recipe.make_optimizers(G, D, prior)  # after moving modules to device
print(recipe.to_dict())  # inspect every resolved value

The optional optimizer helper returns ordinary Adam optimizers. The G optimizer has separate generator and particle groups. You can build your own optimizers using the recipe's fields instead. Recipes are immutable; .replace(...) returns a new one. Unknown options raise errors.

Default	get_recipe() / gan	ddgan
Generation	One-shot GAN	Four-step DDGAN
Prior	20,000 learned particles, dimension 4	Same
GAN loss / critic penalty	Rp logistic / exact L2 cap, weight 1	Same
Particle regularizer	VICReg, weight 1, unique sampled rows	Same
Adam learning rates: G / D / prior	0.0006 / 0.0009 / 0.006	Same
Adam betas / EMA decay	(0, 0.999) / 0.995	Same
Schedule	Hold 60%, cosine to 5%	Same
Batch size / training updates	256 / 7,000	256 / 56,000
Conditioning	Unconditional	Class-only UCD, 4 classes, CE weight 0.02

These defaults come from the selected 100-Gaussians and denoising experiments. Networks remain application choices: the reference toy benchmarks use MLPs and two Fourier frequencies in D. Changing the architecture or dataset changes the experiment; the recipe alone does not establish convergence on a new problem.

The executable PyTorch loop shows optimizer setup, D freezing/restoration, unique-particle regularization, the learning-rate schedule, and EMA for G and the prior. It uses small MLPs and synthetic data, requires no research dependencies, and writes one flushed JSON record per log line:

mkdir -p runs/api
python -u examples/pytorch_loop.py --steps 5 --batch-size 16 > runs/api/smoke.log 2>&1
# In another terminal while a longer run is active:
tail -f runs/api/smoke.log

TOML is just constructor arguments

Load TOML with your preferred parser and unpack a section into a constructor. The library does not require a parser or configuration framework:

[particlegan]
# Omit name for GAN defaults; use name = "ddgan" for DDGAN + UCD.
z_dim = 16
num_particles = 4096
lr = 0.0003

# Alternatively, configure independent primitives:
[prior]
z_dim = 16
num_particles = 4096

[loss]
loss_type = "logistic"
mode = "rp"

try:
    import tomllib  # Python 3.11+
except ModuleNotFoundError:
    import tomli as tomllib  # Python 3.10: pip install tomli

with open("model.toml", "rb") as f:
    config = tomllib.load(f)

recipe = get_recipe(**config["particlegan"])
# Or construct components directly:
prior = ParticlePrior(**config["prior"])
adversarial = GANLoss(**config["loss"])
# Explicit code overrides are ordinary dictionary merges:
recipe = get_recipe(**{**config["particlegan"], "lr": 1e-4})

toml.load(...) dictionaries work too. Choose either recipe-owned values or per-component sections for your application. Try the supplied configuration:

python -u examples/pytorch_loop.py --config examples/api.toml --steps 5

The primary research trainers also accept TOML or YAML using their existing flat experiment schema (separate from the constructor sections above):

python experiments/train_100gaussians.py --config configs/100gaussians/default.toml
python experiments/train_denoising.py --config configs/denoising/default.toml

Running python experiments/train_denoising.py with no arguments loads configs/denoising/default.toml. Install the experiments extra for these trainers; the denoising trainer requires CUDA. See the experiment runner guide for grids and recorded effective configurations.

To use MoG latents with the DDGAN trainer:

mkdir -p results/denoising/mog
python -u experiments/train_denoising.py --config configs/denoising/mog.toml > results/denoising/mog/log.txt 2>&1
# From another terminal:
tail -F results/denoising/mog/log.txt

Set prior = "mog"; sigma_rel and standardize control its fixed noise and read standardization. Optional prior_betas sets separate Adam betas for the component means. The supplied config uses the one-shot MoG recipe's 400 components and prior optimizer settings. The matched small-generator studies at 14k updates and 100k updates report the quality tradeoffs; the supplied full-width 56k configuration is not a selected DDGAN benchmark winner. Set generator_hidden to vary generator width while keeping discriminator width controlled by hidden. The trainer regularizes raw means and preserves the calibrated noise in EMA and checkpoints. Forward diffusion and the separately configured reverse noise source retain their existing behavior.

DDGAN and UCD compose independently

DDGAN supplies Gaussian forward pairs and reverse transitions. UCD selects class scores from a logit network; it does not inject class labels into that network. Neither owns your training loop. Here is the D-loss portion of a conditional denoising pipeline, with caller-defined G, logit_network, data, labels, and device:

import torch
from particlegan import DDGAN, UCD, ucd_loss

recipe = get_recipe("ddgan", num_classes=4)
prior = recipe.make_prior().to(device)
adversarial = recipe.make_loss()
penalty = recipe.make_gradient_penalty()
process = DDGAN(alpha_bar=recipe.alpha_bar).to(device)
critic = UCD(logit_network, num_classes=recipe.num_classes).to(device)

t = torch.randint(1, process.steps + 1, (len(real),), device=device)
rng = torch.Generator(device=device).manual_seed(123)
x_prev, xt = process.forward_pair(real, t, rng)
z, indices = prior.sample(len(real), generator=rng)
x0_hat = G(z, labels, xt=xt, t=t)  # G predicts clean data
fake_prev = process.reverse(x0_hat, xt, t, torch.randn_like(xt))
real_score, real_logits = critic(x_prev, labels, xt=xt, t=t)
fake_score, fake_logits = critic(fake_prev.detach(), labels, xt=xt, t=t)
d_loss = adversarial.d_loss(real_score, fake_score)
d_loss += ucd_loss(real_logits, fake_logits, critic.ucd_labels(labels, t),
                   weight=recipe.ucd_weight)
d_loss += penalty(lambda x: critic(x, labels, xt=xt, t=t)[0],
                  x_prev, fake_prev.detach())

For class-only UCD, the network receives network(x, xt=xt, t=t) and returns [B, C] logits. Labels and times are [B] long tensors; times run from 1 to T. For joint time/class heads, use UCD(network, num_classes=C, target="time_class", num_steps=T); the network receives network(x, xt=xt) and returns [B, T*C] logits. UCD also works without diffusion as critic(x, labels) over a network that accepts only x.

Recompute critic scores after its update, freeze critic parameters, and keep fake_prev attached for the G/prior adversarial loss. Class CE belongs to D. The default schedule is (1, .9, .5, .05, .0001); corruption and reverse noise are Gaussian and separate from learned latent particles. See the API reference for the full composition rules.

Teacher/student pipelines

A teacher can produce the target batch in your existing training pipeline. Use matching conditioning for a paired supervised loss:

teacher.eval()
with torch.no_grad():
    targets = teacher(inputs)
z, indices = prior.sample(len(inputs))
fake = student(z, inputs)
# Update your critic using targets as reals and fake.detach() as fakes.
# Then freeze the updated critic's parameters for this student/prior loss:
student_loss = supervised_weight * supervised_loss(fake, targets)
student_loss += adversarial_weight * adversarial.g_loss(
    D(fake), D(targets).detach(),
)
student_loss += spread(prior.z[indices.unique()])
# Your student/prior optimizer performs backward and step; restore D afterward.

The teacher, supervised objective, loss weights, and update order belong to your application. You can also use only the prior or regularizers without an adversarial objective.

Inference without a critic or optimizer

Save G and the learned prior, ideally their EMA states. Keep the architecture configuration needed to reconstruct G alongside the checkpoint:

torch.save({"generator": G.state_dict(), "prior": prior.state_dict()}, "model.pt")

# In another application, reconstruct your generator architecture and prior:
G = build_generator(z_dim=16).to(device)
prior = ParticlePrior(num_particles=4096, z_dim=16).to(device)
state = torch.load("model.pt", map_location=device, weights_only=True)
G.load_state_dict(state["generator"])
prior.load_state_dict(state["prior"])
G.eval()
prior.eval()
with torch.inference_mode():
    z, _ = prior.sample(64)
    samples = G(z)

Use the same particle count and latent dimension as training. Conditional G also receives labels or inputs. DDGAN inference additionally reconstructs its schedule (or loads its state_dict) and starts from Gaussian x_T; loop over T, ..., 1, drawing fresh latent samples and reverse noise at each step: x = process.reverse(G(z, labels, xt=x, t=t), x, t, noise). There is no critic or optimizer in inference.

The Problem

GANs can suffer from mode collapse: the generator produces only a subset of the data distribution. This project explores whether optimizing a finite latent particle cloud alongside the generator improves coverage on small, highly multimodal benchmarks.

The Insight

What if the prior could move too?

We introduce learnable "particles" in latent space. Both the generator and these latent vectors are optimized during training. The experiments examine how that extra flexibility interacts with discriminator regularization, optimizer dynamics, and sample quality. The results are empirical observations on these benchmarks, not a guarantee against collapse.

Historical Gaussian example

Historical visualization from the older Gaussian example. Its architecture and training recipe differ from the particle example above, so these GIFs are not a matched prior comparison.

Evidence and controls

The historical regularizer study compares discriminator penalties within the particle model. It does not establish that a fixed Gaussian prior necessarily collapses. The current examples share one training loop and matched defaults; the only training change for the Gaussian controls is removing the learned prior and its regularizer.

For a reproducible three-way comparison, run:

python experiments/compare_priors.py --study-dir runs/prior_comparison --run --device cuda:0

This runs learned particles, a frozen Gaussian table, and fresh Gaussian noise on paired seeds 23001–23003. It records configs, source revision, final samples, coverage, transport distances, and per-mode radial and covariance shape diagnostics. See prior controls and interpretation and reproducing the project.

The completed nine-run matched comparison reached 100/100 high-quality modes on every learned-prior seed, with a mean high-quality fraction of 98.6%, versus 8.1% for the frozen table and 6.4% for fresh Gaussian noise. This establishes a concentration advantage under this recipe. The report also shows remaining tail and covariance distortion, finite output support, and transport-metric tradeoffs; it does not establish complete Gaussian calibration or a general guarantee against collapse.

How It Works

1. Particle Prior: Instead of sampling z ~ N(0, I), we maintain a set of learnable latent vectors (particles). During training, we sample from this discrete set.

2. Joint Optimization: Particles are optimized alongside G and D. Their positions can adapt to the data modes.

3. VICReg Regularization: We apply variance-covariance regularization to prevent particles from collapsing to a single point, while allowing arbitrary topology (clusters, gaps, etc.).

Examples

Five Modes (Text Generation)

A minimal example demonstrating the core idea. Five words ("apple", "grape", "lemon", "melon", "berry") are encoded into a 2D latent space. Each word gets one particle.

python examples/five_modes.py

The visualization shows:

• Left: Loss curves for D and G/E/Prior
• Center: 2D latent space with particle positions (white stars) and encoded words (colored dots)
• Right: Reconstruction quality over training

100 Gaussians (2D Distribution)

The main benchmark. 100 Gaussian modes arranged on a 10×10 grid. This is a stress test for mode coverage.

python examples/100gaussians.py

The historical particle study reports runs with 100/100 modes and approximately 99% of samples within 3σ of a center after 7k steps. Coverage alone does not establish that the within-mode distribution is correct; the trainer also records shape and transport metrics.

The default recipe is RpGAN (relativistic, logistic) + a one-sided cap gradient penalty on D (relu(‖∇ₓD‖ − 1)² on reals and fakes, coeff 1.0), Fourier-feature D, EMA evaluation, Adam β1=0, base LR 6e-4 with a delayed cosine anneal. The cap won a 420-run bake-off against the zero-centered R1/R2 penalty, which is still available with --reg_arm a_r1r2 --reg_coeff 0.02. See FINDINGS.md for the study and docs/convergence-tips.md for the transferable reasoning behind each ingredient.

Without particle prior (baseline):

python examples/100gaussians_no_particle_prior.py

This entrypoint uses the same architecture, losses, learning rates, schedule, and EMA as the particle example, with fresh Gaussian noise. Use --prior frozen_gaussian for a finite frozen-table control. The outcome depends on the recipe and seed; the baseline does not assume collapse.

Project Structure

ParticleGAN/
├── particlegan/            # Installable PyTorch primitives and recipe helpers
│   ├── particle_prior.py   # Learnable particle cloud (nn.Module)
│   ├── gan_loss.py         # Flexible GAN losses (hinge, logistic, Wasserstein, LSGAN)
│   ├── grad_regularizers.py # D gradient penalties (cap, R1/R2, eikonal, ...)
│   ├── vicreg_loss.py      # Variance-covariance regularization
│   ├── diffusion.py        # DDGAN forward/reverse transitions
│   ├── conditioning.py     # UCD scores and class supervision
│   └── recipes.py          # Inspectable defaults and optional factories
├── lib/                    # Repository compatibility imports and research helpers
├── examples/
│   ├── pytorch_loop.py                  # Minimal caller-owned loop (Torch only)
│   ├── api.toml                         # Constructor/recipe configuration
│   ├── five_modes.py                    # Text generation toy problem
│   ├── 100gaussians.py                  # 100-mode benchmark (with particles)
│   └── 100gaussians_no_particle_prior.py # Baseline (without particles)
└── README.md

The grid-search infrastructure behind the study — config generation, the per-arm trainer, grid runner, and the analysis/leaderboard scripts — lives in experiments/, with the generated per-run configs in configs/.

The CIFAR DDGAN experiment scales the particle recipe to images. Its speed study compares exact/lazy/finite-difference bcap and backports the shared implementation to both toy trainers. The faster CIFAR default retains exact derivatives; FD is optional.

Notes

• The text experiments (five_modes.py) use the same recipe (RpGAN + one-sided cap penalty on the joint critic ∇₍ₓ,𝓏₎D, EMA, β1=0, cosine anneal)
• The 100-Gaussian experiments use the one-sided cap penalty (--reg_arm, default b_cap); a gradient penalty is what lets the sharp Fourier discriminator keep full mode coverage
• Particles use a higher learning rate (10×) than G/D for faster adaptation

Changelog

Versions before 0.2 tracked the default recipe of examples/100gaussians.py.

0.5.0 — 2026-09-17

• Add particle AE-GAN, constant-KL VAE-GAN and AE-DDGAN recipes and public encoding/reconstruction helpers with caller-owned training loops.
• Add optional encoder optimizer integration, explicit ELBO reporting and categorical inference, numerical tests and installed-wheel coverage.
• Preserve queued toy/CIFAR studies and publish their leaderboards and limits. See the full changelog.

0.4.0 — 2026-09-17

• Adds get_recipe("ddgan_mog"): four-step DDGAN with class-only UCD, 400 MoG components, z_dim 4, sigma_rel 0.025, standardized reads, 100k updates, constant LR, prior LR 0.06 and prior Adam betas (0.5, 0.999).
• Adds MoG support to train_denoising and checkpoint probes, raw-mean regularization, separate prior optimizer settings, and independent generator width through generator_hidden. Existing recipe defaults remain unchanged.
• Includes the matched 14k/100k capacity studies and frozen-noise interventions. At 100k, DDGAN+MoG reaches 79.57% joint HQ and all 100 modes; one-shot models retain higher HQ but cover 77 modes. These single-seed results use a small generator and do not isolate representational capacity. See the 100k readout.

0.3.0 — 2026-09-17

• Adds the public MoGParticlePrior: a uniform mixture with learned means and a shared, fixed Gaussian sigma calibrated from initial nearest-neighbor spacing (defaults: 400 components, z_dim 4, sigma_rel 1/40, standardized reads).
• Adds get_recipe("mog"), the selected 400-component, 28k-step recipe (prior LR 0.06, prior Adam betas (0.5, 0.999), existing GAN defaults), plus configs/mog/default.toml for the 100-Gaussian trainer. Recipe factories can select the prior and set prior betas separately.
• MoG supports explicit sampling generators, fixed epsilon snapshots, noisy module forward calls for DDP, raw-center regularization, EMA, and state-dict restoration of read configuration and calibrated noise. Legacy experimental checkpoints remain loadable.
• Core dependency stays PyTorch only; the optional mog extra installs SciPy for faster calibration of large tables, with an exact Torch fallback without it.
• Passed the frozen C0 envelope on 100 Gaussians at 28k steps (HQ/real 0.99953, width/real 0.92636, KL 0.02888) with 50× fewer components and 4× the steps of the original baseline. Single-seed result; see results/mog/COMPONENT_SCALE.md.

0.2.0 — 2026-09-16

• Adds the installable particlegan namespace, independent PyTorch primitives, immutable recipes, and direct use of loaded TOML dictionaries.
• Core runtime requires only Torch; research dependencies use the experiments extra.
• Repository trainers use the shared package while retaining their own loops.

0.1.2 — 2026-08-22

• Default gradient penalty switched to the one-sided cap (b_cap, relu(‖∇ₓD‖ − 1)², coeff 1.0) via lib/grad_regularizers.py; base LR 3e-4 → 6e-4; run length 5k → 7k steps.
• Chosen by a 420-run controlled study (FINDINGS.md): same game-damping as R1/R2, sharper modes (hq 0.986 vs a ~0.94 ceiling), honest per-mode core width (0.87), zero collapses. R1/R2 stays available via --reg_arm a_r1r2.
• Adds the experiments/ study infrastructure and the deterministic video renderer.

0.1.1

• R3GAN-style defaults (previously undocumented; commit b5529cb): RpGAN logistic objective + zero-centered R1+R2 (γ=0.02) + Fourier-2 features on D + EMA(0.995) on G and prior + Adam β1=0 + delayed cosine LR anneal + z_dim 4.
• Full coverage with ≥90% hq in ~3.5k steps.

0.1.0

• Original example: vanilla/hinge GAN, no gradient regularizer, plain MLP D, z_dim 2, Adam β1=0.5, no EMA, no LR anneal.
• Never converged on the 100-Gaussians benchmark: ~86–92/100 modes, ~30% hq at 12k steps (baseline row in docs/convergence-tips.md).

Citation

@software{particlegan2025,
  author = {Martyn Garcia},
  title = {ParticleGAN: Learnable Priors for Stable GANs},
  year = {2025},
  url = {https://github.com/255BITS/ParticleGAN}
}

License

MIT