LifeGenIP Workshop

Overview

Recent advances in diffusion models have revolutionized video generation, enabling the generation of high-quality videos. Both open-source and commercial models have shown increasingly impressive results. These models enable users to specify the subject of the generated video, a capability commonly referred to as video customization. Such customization typically follows two approaches. Tuning-based methods fine-tune a video diffusion model on a small set of video clips, for example, using Low-Rank Adaptation (LoRA), and can then generate diverse new videos. Reference-based methods, such as image-to-video (I2V) models, take a reference image and a text prompt, offering pixel-level control.

However, these models raise serious concerns about about personal privacy, identity ownership, and intellectual property protection. A malicious user can collect public video clips of a target person and exploit current models to generate misleading or harmful content about that person: not only by finetuning a video diffusion model on the whole collection through tuning-based customization, but also by extracting a single frame from any of these videos as the reference image for reference-based customization. These threats may lead to fake news and violations of portrait rights. Protecting video content from being misused thus becomes a pressing concern.

Existing anti-diffusion works have been mostly developed in the image domain, covering both tuning-based customization of T2I models and reference-based I2I editing. Protection on the video side, however, has received much less attention: the few existing efforts target only reference-based I2V pipelines, while protection against the tuning-based customization of T2V models remains unexplored. This leaves a critical gap, since tuning-based customization is another pathway through which a malicious user can exploit.

Competition Guidelines

Task Definition

Participants are given collections of short, face-centric video clips of individuals and must return protected (“unlearnable”) versions of those clips. The goal is that, if a malicious attacker takes the protected clips and runs them through diffusion-based video customization, the generated videos either fail to reproduce the target identity or fail to follow the attacker's text prompt, while the protected clips remain perceptually close to the originals. We consider the two customization threats below.

Reference-based customization (I2V): the attacker extracts a single frame from a protected clip and uses it, together with a text prompt, as the conditioning image of an image-to-video model.
Tuning-based customization (fine-tuning): the attacker fine-tunes a video diffusion model (for example, via LoRA) on the whole protected video set and then samples new videos using text prompts.

A submission is a set of protected videos in one-to-one correspondence with the released set: only pixel values may change (resolution and frame count are preserved), within a specified perturbation budget.

Phases & Tracks

The competition uses three evaluation tracks:

Effectiveness: evaluated on the diffusion model that the organizers disclose to participants.
Transferability: evaluated on additional diffusion models that are not disclosed.
Robustness: protected videos are first passed through temporal attacks that are not disclosed, and are then used for customization.

Phase 1 — Preliminary Round (open to all teams)

Setting: reference-based customization (I2V). A randomly sampled frame, together with held-out prompts, is fed to an image-to-video model.
Evaluation: submissions are scored on the Effectiveness and Robustness tracks.
Outcome: each team may submit up to two times, and top teams advance to Phase 2.

Phase 2 — Final Round (top teams from Phase 1)

Settings: reference-based customization (I2V) and tuning-based customization (fine-tuning). For reference-based, a randomly sampled frame, together with held-out prompts, is fed to image-to-video models. For tuning-based, the protected video set is used to fine-tune video diffusion models, which then generate videos using held-out prompts.
Evaluation: each setting is scored on all three tracks (Effectiveness, Transferability, Robustness).
Outcome: each team may submit up to two times, and the final ranking is determined by Phase 2, aggregating all tracks across both settings.

Key Dates

Phase 1 Release Jun 15, 2026 —
Phase 1 Evaluation Start Jun 20, 2026 —
Phase 1 Evaluation End Jul 15, 2026 —
Phase 2 Release Jul 15, 2026 —
Phase 2 Evaluation Start Jul 20, 2026 —
Phase 2 Evaluation End Jul 30, 2026 —
Notification of Final Ranking Aug 1, 2026 —

All deadlines are in Anywhere on Earth (AoE) time.

Evaluation Metrics

Submissions are judged on two criteria: protection effectiveness (measured on the generated videos) and imperceptibility (measured on the protected videos against the originals).

Participation & Awards

Teams may have up to 10 members.
Prize-winning teams must release reproducible code. They are also invited to submit a report describing their method, which will be accepted by the LifeGenIP workshop at ECCV 2026.
Prizes: cash prizes for the top-ranked teams (amount to be announced), sponsored by TBA.

Useful Links

Submission Portal: Codabench.
Starting Kit: coming soon.
Leaderboard: coming soon.

Competition Organizers

Contact

For workshop coordination, submission questions, speaker-related inquiries, and sponsorship opportunities, please contact the organizing committee: lifegenip_workshop@googlegroups.com

LifeGenIP Competition: Unlearnable Videos against Diffusion-based Customization

Overview

Competition Guidelines

Task Definition

Phases & Tracks

Key Dates

Evaluation Metrics

Participation & Awards

Useful Links

Competition Organizers

Ms. Yuxin Huang

Mr. Gaoge Han

Dr. Lixu Wang

Dr. Shanshan Ye

Sponsors

Contact

WeChat Group