UniMark Embeds User IDs Into AI Images Without Retraining Any Model
A training-free framework embeds multi-bit messages into autoregressive image generators like LlamaGen and VAR, resisting six attack types.
ProTakeaways
- UniMark is a training-free multi-bit watermark for autoregressive image generators like LlamaGen and VAR.
- Adaptive Semantic Grouping generates a different key-dependent green/red codebook split for every token position.
- Block-wise encoding with BCH codes reaches above 99% bit accuracy for 32-bit messages across three models.
- FID degradation is roughly half of IndexMark: 3.82 vs 3.75 unwatermarked baseline on LlamaGen.
- Forgery resistance under 50% partition exposure: under 2% for UniMark versus 68% for static-partition methods.
- Adds only 8 to 15 ms per image, about 1.5% overhead, with no model retraining required.
Watermarking for autoregressive image generators has been stuck at a frustrating level: current methods only answer yes-or-no questions like "was this made by our model?" They cannot embed a user ID, a timestamp, or a version number, which is exactly what content provenance requires in production. A new framework called UniMark tries to close that gap without any retraining, and it works across both major autoregressive paradigms.
The paper identifies three concrete failures in existing work. Most methods embed only zero-bit watermarks that support binary verification without encoding multi-bit messages, rely on static red-green codebook partitions that can be removed or forged once exposed, and are tailored to a specific architecture rather than generalizing across paradigms. UniMark attacks all three at once.
How the three pieces fit together
Autoregressive image models like LlamaGen and VAR turn images into discrete tokens drawn from a learned codebook, then generate them one at a time. Prior watermarks like IndexMark split that codebook into a fixed "green" set and "red" set, then nudge generation toward green tokens. UniMark keeps the green-token idea but rebuilds the machinery around it.
- Adaptive Semantic Grouping (ASG). Rather than one static partition, ASG generates a different green/red split for every token position by hashing the secret key together with the position index. Semantically similar codebook entries are grouped so that when a token has to be swapped, the visual replacement stays close to the original.