Genμ 2.0 (2026) — Multimodal Identity Unlearning Challenge @ ECCV 2026

Challenge open. Final submission deadline: 10 July 2026. Evaluation and results declaration: 25 July 2026.

Challenge open · Genμ 2.0 · ECCV 2026

Multimodal Identity Unlearning Challenge

The 2026 update expands the original Genμ concept-unlearning benchmark into a broader privacy-preserving suite co-located with the U&ME Workshop at ECCV 2026. Three tracks test whether a generative model can forget one specified identity or concept without damaging nearby identities, attributes, styles, or general generation quality.

Submissions are collected through three track-specific Google Forms. Each entry must provide a shareable drive link to the submitted weights or checkpoint. The link may use any accessible cloud-storage provider. A complete GitHub repository and clear reproduction/testing instructions are also required.

Face problem statement Speech problem statement Concept problem statement Submission instructions

Task 1 — Face UnlearningRemove one face identity from any suitable face generator while retaining nearby identities and non-identity facial attributes. Read the full task →

Task 2 — Speech UnlearningRemove one speaker identity from any suitable TTS or voice-cloning model while retaining content, quality, prosody, and related speakers. Read the full task →

Task 3 — Concept UnlearningErase a target visual concept from the legacy Stable Diffusion benchmark while preserving adjacent concepts and robustness. Read the full task →

Challenge timeline

Key dates for the 2026 edition.

Challenge status

Open now

Teams may submit to one or more tracks.

Final submission deadline

10 July 2026

Submit the track form with weights, code, and complete testing instructions.

Evaluation & results

25 July 2026

Evaluation is completed and final results are declared.

Leaderboard

Updated during the challenge

Verified submissions are added to the website leaderboard as evaluation progresses.

Publication opportunities

Submit a method paper to the workshop

Challenge participants may independently submit a paper describing their method to the U&ME Workshop through the regular review process. The workshop paper deadline is 9 July 2026. Paper submission and challenge artifact submission happen in parallel; entering the challenge does not automatically submit a paper.

Challenge report co-authorship

The two highest-ranked eligible teams will be invited to co-author the Genμ 2.0 competition report. A team is eligible only if it finishes in the top two and exceeds the relevant official baseline for its track. The report will be prepared after evaluation for the workshop camera-ready stage.

How to submit

Use the Google Form for the selected track. A separate form is used for each track.

1. Weights

Upload the final unlearned weights, checkpoint, or adapter to a drive or cloud-storage provider and enable access for evaluation.

2. Complete code

Provide a GitHub repository containing unlearning code, inference code, environment files, and the exact commit or release tag.

3. Testing instructions

Document setup, commands, inputs, expected outputs, hardware requirements, and any track-specific preprocessing needed to reproduce the submission.

Face Track Google Form — link pending Speech Track Google Form — link pending Concept Track Google Form — link pending

For any submission-related query, contact genmu.challenge.2025@gmail.com.

Challenge tracks

Face and speech participants may use any suitable generative model. The suggested models are examples only; every submission must identify its exact architecture, upstream checkpoint, conditioning interface, and inference procedure.

Task 1 · Face

Face Identity Unlearning

Given a face generator and a specified target identity, modify the model so that its normal identity-conditioning route can no longer reproduce or verify that person. At the same time, preserve the model's ability to generate the provided retain identities, preserve non-identity attributes, and maintain visual quality.

Model choice: open
Public data: CelebA official dataset; FFHQ official dataset and download tools are also supported
Example models: Arc2Face, IDiff-Face, DCFace, EG3D
Primary metric: ERB, the harmonic mean of erasure accuracy and retain-identity accuracy

Full problem statement & evaluation Suggested models Face submission form — link pending

Task 2 · Speech

Speaker Identity Unlearning

Given a TTS or voice-cloning model and a target speaker, modify the model so that generated speech is no longer recognized as that speaker. Preserve the transcript, naturalness, prosody, accent or style attributes required by the subtask, and the voices of acoustically similar retain speakers.

Model choice: open
Public data: LibriTTS (OpenSLR 60) / LibriTTS-R (OpenSLR 141)
Example models: StyleTTS2-LibriTTS, XTTS-v2
Primary metric: ERB, the harmonic mean of erasure accuracy and retain-speaker accuracy

Full problem statement & evaluation Suggested models Speech submission form — link pending

Task 3 · Concept

Visual Concept Unlearning

Given Stable Diffusion v1.4 and a target concept, erase the concept for direct, indirect, and adversarial prompts. Preserve unrelated and semantically adjacent concepts, general image-generation utility, and robustness with minimal unnecessary parameter change.

Reference model: Stable Diffusion v1.4
Baselines: ESD, CA, FMN, FADE
Primary metric: ERR score

Full problem statement & evaluation Published baselines Concept submission form — link pending

How Face and Speech are ranked

For both identity tracks, the only leaderboard ranking metric is ERB: the harmonic mean of erasure accuracy \((100-\mathrm{FA})\) and retain accuracy. Geometry, attributes, prosody, quality, and intelligibility are reported separately for analysis and do not contribute to rank. See the exact Face formula and Speech formula.

Subtasks in 2026

A1
Identity unlearning under high inter-identity similarity
Forget the target while preserving highly similar neighboring identities. Details →

A2
Identity with attribute preservation
Forget identity while preserving attributes such as pose, smile, beard, or eyewear. Details →

B1
Selective speaker unlearning
Forget the target while retaining acoustically similar speakers; prosody is assessed separately. Details →

B2
Speech attribute preservation
Forget speaker identity while preserving accent, emotion, speaking style, intelligibility, and quality. Details →

Useful links

Challenge resources

Suggested model references

These are optional starting points, not mandatory base models. See the Baselines page for descriptions and official references.