---

name: "mmgeneration" description: "Routes MMGeneration tasks across sampling, training, evaluation, configuration, and latent-editing workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MMGeneration

MMGeneration is OpenMMLab's generative-model toolkit for GANs, translation models, SinGAN, and diffusion workflows.

Use this repo skill when the task mentions mmgen, MMGeneration, one of the model families below, or the repo's train/eval/demo/application scripts.

Start here

1. Read references/installation-and-compatibility.md if the environment is not already verified.
2. Read references/repo-provenance.md when you need to check whether this skill still matches the current checkout.
3. Use the route map below to choose the right sub-skill.
4. Use scripts/check_install.py for a fast import and backend smoke check.

Installation and verification

MMGeneration depends on a compatible PyTorch wheel, mmcv-full in the 1.x line, and mmcls in the 0.x line. A practical install order is:

1. Install PyTorch and torchvision.
2. Install a matching mmcv-full wheel for that torch/CUDA combination.
3. Install mmcls<1.0.0 and the runtime requirements used by this repo.
4. Install the repo itself in editable mode when working from a checkout.
5. Run python scripts/check_install.py.

For detailed compatibility notes, backend guidance, and common failure modes, read references/installation-and-compatibility.md and references/troubleshooting.md.

Route map

Sub-skill	Use it for	Typical signals
sub-skills/inference-and-sampling/	Load a generator and sample unconditional, conditional, translation, or DDPM outputs	init_model, sample_unconditional_model, sample_img2img_model, demo scripts, sample shapes, device selection
sub-skills/training-and-distribution/	Train or resume configs, choose launcher mode, or debug distributed training	tools/train.py, dist_train.sh, slurm_train.sh, DynamicIterBasedRunner, EMA, CPU training, DDP wrapper
sub-skills/evaluation-and-metrics/	Run evaluation, precompute inception stats, or reason about FID/IS/PPL/SWD/PR/MS-SSIM	tools/evaluation.py, inception_stat.py, translation_eval.py, metrics, online/offline evaluation
sub-skills/configuration-and-extension/	Edit configs, define datasets/pipelines, register new models/losses/hooks, or inspect config expansion	_base_, _delete_, custom_imports, cfg-options, registries, dataset layouts, print_config.py
sub-skills/applications-and-deployment/	Latent interpolation, projection, SeFa, StyleCLIP, or TorchServe packaging	apps/, mmgen2torchserver.py, projector.py, modified_sefa.py, styleclip.py

Public model families

MMGeneration's supported workflows cluster around these families:

• Unconditional GANs: DCGAN, LSGAN, WGAN-GP, PGGAN, StyleGANv1/v2/v3, MS-PIE StyleGAN2, ADA.
• Conditional GANs: SNGAN/Projection GAN, SAGAN, BigGAN, BigGAN-Deep.
• Image translation: Pix2Pix and CycleGAN.
• Internal learning: SinGAN and PESinGAN.
• Diffusion: Improved DDPM.

See references/model-overview.md for the repo-facing model map and where each family lives.

Key public APIs

The installed package exposes mmgen.apis helpers for model loading and sampling, mmgen.models builders and registries, mmgen.datasets builders and dataset classes, and mmgen.core evaluation, hook, optimizer, and runner utilities. Read references/api-reference.md for verified signatures and references/data-formats.md for dataset layouts and pipeline keys.

CLI and command references

The repo does not define a console entry point; use the repository scripts directly. Read references/cli-reference.md for the main command families and flag meanings.

Shared helpers

• scripts/check_install.py — verify that the installed package imports, the public modules are visible, and the expected backend is usable.
• sub-skills/configuration-and-extension/scripts/print_config.py — print a fully resolved config with optional overrides.
• sub-skills/inference-and-sampling/scripts/sample_mmgen.py — run a safe sampling helper across the common inference modes.

Troubleshooting

If a workflow fails, start with references/troubleshooting.md. The most common issues are:

• mmcv.runner import failures from an incompatible MMCV version.
• mmcv.ops/CUDA extension problems from the wrong wheel or missing GPU runtime.
• styleclip.py requiring the optional clip package.
• Dataset layout mismatches for paired or unpaired translation data.
• Metric/extraction commands that need cached inception or VGG assets.

Provenance

When the repo changes, compare it with references/repo-provenance.md before reusing this skill.