Agentic Image-to-Video Optimization: Make Adherence Reviewable

Reconstructed and revised on 2026-09-30 from the surviving Chinese manuscript and checked sources. This is not a verbatim recovery of the previous English article.

Research turns generation into feedback

The original paper combines multimodal prompt refinement through semantic and artifact questions with Bayesian search over seeds and CFG settings. It studies adherence in image-to-video generation. Commercial services may not expose those controls. VQOS has not reproduced the method; author preference results are not project success rates.

Turn actions into questions

Specify subject, action, direction, order and protected elements. A hypothetical instruction is “pick up the cup with the right hand, keep it visible and do not cut the camera.” Check the correct hand, object persistence and completion separately.

A reference image establishes appearance but cannot independently prescribe future movement. Resolve contradictions between reference and instruction first.

Budget search and define a stop

Limit retries or total expenditure and record each change with its reason. Changing references, prompts and all parameters together makes attribution difficult. If a highly scored candidate still violates a critical requirement, revisit the task or method before generating more.

Keep a human acceptance review

Watch motion, occlusion, product details and background, then check text and sound. Automated questions may omit branding, true dimensions or project-specific constraints.

Review VQOS services and specify pilot scope, retry charges and acceptance questions in your brief. Closed-loop research can inform workflow design; it does not promise unattended review or reliable delivery.