MaLiang-Harness: A Programmable Path to Image and Video Generation

MaLiang-Harness. A Programmable Path to Image and Video Generation. By Haoyu Zhao and six others, from the National University of Singapore, Fudan University, and one other institution. Published on the twenty-eighth of September. Its code repository is worked out: about one thousand five hundred source files, with a readme.

Executable programs offer explicit control over how images and videos are constructed. A program can say exactly where a shape goes, what colour it is, and how it moves. But generating code that runs is only the beginning of visual creation. A program can execute correctly and still violate the requested composition, appearance, or motion. The authors name that discrepancy the Program-to-Visual gap: the distance between code that works and a picture that is right.

MaLiang-Harness is a unified framework that organises visual generation driven by multimodal language models into a persistent process of construction, inspection, and revision. Its central design decision is that three things share one common revision reference: the visual program as it evolves, the history of how it was constructed, and its verification. When all three point at the same revision, it is always clear which version of the program a check was run against.

Three mechanisms carry that out. The Persistent Executable Generation state, or PEG, preserves the programs and the task context. The Traceable Generation Process, or TGP, connects each edit to the rendered evidence. And Revision-aware Editing and Verification, or REV, supports restoration and checks the current revision before the task is completed.

Together these coordinate planning, execution, and visual feedback, across different rendering backends. The program is not simply written and run once. It is written, rendered, inspected, and revised, with a traceable record at each step. That record is what allows a bad edit to be undone, and what lets a check say precisely which version it approved. Without it, inspection and revision would be working from different pictures of the same program.

The evaluation runs eleven powerful closed-source multimodal models on MaLiang-IBench, and four on MaLiang-VBench, measuring generation success, visual quality, and computational cost. GPT six Astra reaches one hundred percent generation success on both benchmarks. Ninety-six percent of its image tasks, and seventy-six point nine percent of its video tasks, meet all of the quality thresholds.

The comparison also turns up a mismatch. Models with similar general capability scores differ substantially in their ability to satisfy visual requirements. A strong score on general benchmarks, in other words, is an unreliable predictor of how well a model will do here. Writing a program that renders the requested scene seems to call on something those benchmarks do not capture.

The framework is offered as a systematic basis for studying how multimodal models translate executable code into visual outcomes. It exposes two sides at once: how far programmable generation can already go, and how weakly general benchmarks predict which models will get there. Code that runs is where this work starts, not where it stops.