Genμ 2.0
Baselines and suggested starting models
Published baselines and optional model references
Task 3 retains four published concept-unlearning baselines. Tasks 1 and 2 are model-agnostic: the face and speech models listed here are examples that help teams start quickly, not mandatory architectures.
Task 1 — Face model references
Participants may use any face generator. These references cover identity-conditioned diffusion, synthetic face generation, reference-image personalization, and 3D-aware GANs. Read the full Face Track problem statement before selecting a model.
| Model | Model family and why it is relevant | Official resources |
|---|---|---|
| Arc2Face | Identity-conditioned diffusion model driven by ArcFace embeddings. It offers a direct identity interface and an official downloadable checkpoint. | Hugging Face checkpoint · official code · paper |
| IDiff-Face | Identity-conditioned latent diffusion system designed to generate synthetic identities with controlled intra-identity variation. | official code · paper |
| DCFace | Dual-condition diffusion face generator that separates identity and style conditions, making it suitable for studying selective identity removal while retaining variation. | official code · paper |
| EG3D | Efficient Geometry-aware 3D GAN. It is useful for teams studying identity unlearning in a 3D-aware face generator with camera/view control. | official code and pretrained networks · project page · paper |
| InstantID | Reference-image identity-preserving diffusion approach. Teams using it must clearly separate the fixed base model from the identity-conditioning components they unlearn. | Hugging Face · official code |
| IP-Adapter-FaceID | Face-ID adapter for diffusion models. Suitable for adapter-level or joint base-model/adapter unlearning when the submitted loading procedure is fully documented. | Hugging Face files and instructions |
Task 2 — Speech model references
Participants may choose any reproducible TTS or voice-cloning architecture. Read the full Speech Track problem statement and document sample rate, speaker conditioning, text normalization, and decoding settings.
| Model | Why it is relevant | Official resource |
|---|---|---|
| StyleTTS2-LibriTTS | Style-based TTS starting point aligned with LibriTTS-style evaluation and controllable speaker/style representations. | Hugging Face checkpoint |
| XTTS-v2 | Multilingual voice-cloning model with reference-audio speaker conditioning; useful for studying target-speaker removal while preserving text and naturalness. | Hugging Face checkpoint |
Other TTS, zero-shot voice-cloning, diffusion, flow-matching, autoregressive, or codec-language models are welcome if the complete checkpoint and evaluation interface can be reproduced.
Task 3 baselines — Concept Unlearning
The table below lists the four published methods retained from the original Genμ challenge. Read the full Concept Track problem statement.
| Method | Full name | Setting | Citation |
|---|---|---|---|
| ESD | Erased Stable Diffusion | Model-weight concept erasure using negative guidance as the training signal. | Erasing Concepts from Diffusion Models, Gandikota et al., 2023 |
| CA | Concept Ablation | Maps a target concept distribution toward an anchor concept while seeking to preserve related concepts. | Ablating Concepts in Text-to-Image Diffusion Models, Kumari et al., 2023 |
| FMN | Forget-Me-Not | Targeted removal of identities, objects, or styles while retaining unrelated generation capability. | Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models, Zhang et al., 2023 |
| FADE | Fine-Grained Attenuation for Diffusion Erasure | Adjacency-aware concept unlearning using concept-neighborhood and mesh components. | Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models, Thakral et al., 2025 · code |