OpenCoF

OpenCoF studies Chain-of-Frame (CoF) reasoning, where a video generation model reasons through the temporal evolution of its own generated frames rather than through static images or externally invoked tools. We construct OpenCoF-17K, a dataset of 17,312 videos across 11 task families built from four complementary curation pipelines, and use it to fine-tune Wan2.2-I2V-A14B into Wan-CoF, which improves over the baseline on four independent external benchmarks—MME-CoF, Gen-ViRe, VIPER, and RULER-Bench—purely from data supervision. Building on this, we further introduce two complementary reasoning-token designs, Visual Reasoning Tokens (vt) and Textual Reasoning Tokens (tt), yielding Wan-CoF_vt and Wan-CoF_tt, which improve further while specializing in different strengths across benchmarks.

OpenCoF-17K contains 17,312 samples across 11 task families. Each instance pairs an initial conditioning image and text prompt with a target reasoning video standardized to 480p, 15 fps, and 81 frames. It is curated through four complementary pipelines: instance-based and expert-guided rendering produce precise, rule-governed videos for tasks like chess, Sudoku, and geometry; procedural scene synthesis renders physics- and 3D-grounded motions with graphics engines; and repurposing of existing videos adds real-world diversity such as embodied manipulation.

Wan-CoF is obtained by LoRA fine-tuning Wan2.2-I2V-A14B on OpenCoF-17K. Building on this data-only stage, we explore two complementary reasoning-token designs: Visual Reasoning Tokens (vt), prepended to the visual latent sequence to capture low-level cues via self-attention, and Textual Reasoning Tokens (tt), prepended to the text-conditioning sequence to supply a high-level semantic prior via cross-attention. Training each separately yields Wan-CoF_vt and Wan-CoF_tt.

Wan-CoF improves the Wan2.2-I2V-A14B baseline on every headline metric across four external video reasoning benchmarks—MME-CoF, Gen-ViRe, VIPER, and RULER-Bench—despite being trained with no reasoning-specific machinery at all. The reasoning-token variants push further still and specialize in complementary directions: Wan-CoF_vt leads on Gen-ViRe and RULER-Bench, while Wan-CoF_tt leads on MME-CoF and VIPER.