Pleaseprovideadescription.531 We want to provide an open-source large-scale music dataset for the research com-532 munity. Such large datasets do not yet exist in this domain, and we believetheyare533 neededtodemocratize innovationinmusicresearch andML-assisted musiccreation.534

However, existing music datasets suffer from limited size, accessibility, and lack of audio resources.

A.1 Deriving the Optimum of the KL-Constrained Reward Maximization Objective In this appendix, we will derive Eq. 4. Analogously to Eq. 3, we optimize the following objective: max

Membership inference arises in these contexts as a potential auditing method for detecting unauthorized data usage.

ByteDance's new AI video tool Seedance 2.0 can generate videos based on just a few lines of text. ByteDance's new AI video tool Seedance 2.0 can generate videos based on just a few lines of text. Videos created by new Seedance 2.0 generator go viral, including one of Tom Cruise and Brad Pitt fighting Mon 16 Feb 2026 03.25 ESTLast modified on Mon 16 Feb 2026 03.29 EST ByteDance, the Chinese technology company behind TikTok, has said it will restrain its AI video-making tool, after threats of legal action from Disney and a backlash from other media businesses, according to reports. The AI video generator Seedance 2.0, released last week, has spooked Hollywood as users create realistic clips of movie stars and superheroes with just a short text prompt. On Friday, Walt Disney reportedly sent a cease-and-desist letter to ByteDance which accused it of supplying Seedance with a "pirated library" of the studio's characters, including those from Marvel and Star Wars, according to the US news outlet Axios. Disney's lawyers claimed that ByteDance committed a "virtual smash-and-grab" of their intellectual property, according to a report from the BBC.

Diffusion models have recently shown remarkable success in high-quality image generation. Sometimes, however, a pre-trained diffusion model exhibits partial misalignment in the sense that the model can generate good images, but it sometimes outputs undesirable images. If so, we simply need to prevent the generation of the bad images, and we call this task censoring. In this work, we present censored generation with a pre-trained diffusion model using a reward model trained on minimal human feedback. We show that censoring can be accomplished with extreme human feedback efficiency and that labels generated with a mere few minutes of human feedback are sufficient.