Unified multimodal models can't weave text and images on their own: UniWeaver adds planning and visual checks, lifting BAGEL's overall score from 27.7 to 58.9
Synopsis
Unified multimodal models can understand and generate images but generally lack native support for coordinating interleaved text-image output, and turning a coarse request into coherent multimodal content requires explicit planning. UniWeaver is a model-agnostic structured harness that decomposes a request into content units, assigns what text and images each convey, then generates images and verifies them against observable checklists, regenerating only failed units. On the 120-task UniWeaver-Bench, the BAGEL, OmniGen2 and TokenFlow backbones rise from overall scores of 27.7, 27.9 and 39.2 to 58.9, 52.7 and 60.5 without updating model parameters.
Interpretation
Interleaved text-image output no longer depends entirely on a unified multimodal model's native ability; a pluggable outer harness handles planning, generation and visual verification. The prior approach was a minimal step-wise harness that generated successive text-image pairs conditioned on the request and preceding steps, with no explicit planning and no visual verification. On UniWeaver-Bench, the same backbones gain substantially when wrapped by the framework: BAGEL from 27.7 to 58.9, OmniGen2 from 27.9 to 52.7, and TokenFlow from 39.2 to 60.5, without parameter updates.
The framework first splits a coarse request into content units and decides which information goes to text and which to images. Planning runs in four stages: fix a title, a task endpoint and each unit's visual subject and content goal; write reader-facing prose and image prompts separately under a shared visual style; select reusable references and route them only to relevant units; and list per-unit checks judgeable directly from the image. After generation each image is checked item by item, failed units are regenerated alone with feedback appended to the original instruction for at most two extra attempts while passing units stay unchanged; with BAGEL as backbone, LoRA fine-tuning only the understanding decoder while freezing image-generation experts on the 10,000 trajectories of UniWeaver-10K reaches an overall score of 87.1.
Perspective
Evaluation is confined to the 120 held-out tasks of UniWeaver-Bench, covering 20 task families in four categories with six tasks per family and 40 easy, 40 medium and 40 hard tasks, each giving only a natural-language goal; the framework is validated on seven unified multimodal backbones, and supervised fine-tuning is done only on BAGEL.
The authors note that output quality remains constrained by the underlying unified multimodal models, whose text and images may not yet meet professional-grade standards for demanding applications; the study covers only a defined set of interleaved generation tasks, leaving broader task scenarios and application settings to be explored, and the framework does not independently guarantee factual accuracy, fairness or appropriateness of generated content.
