acceptodds
Under review as a conference paper at ICLR 2027

ViMax: Agentic Video Generation

Abstract

Long-form video generation requires both systematic narrative planning and visual consistency that current short-clip methods cannot provide. Existing methods generate isolated sequences without narrative structure and lack mechanisms for maintaining character and environmental consistency across scenes. We present ViMax, an agentic video generation framework that addresses video creation through coordinated multi-agent collaboration where specialized components negotiate narrative decisions, visual continuity, and production quality. Our framework employs a hierarchical narrative engine with retrieval-augmented generation for global story coherence and a dependency-aware visual consistency mechanism that tracks character and environmental states across temporal boundaries, while VLM-guided agents continuously monitor and refine both narrative coherence and visual fidelity. The framework enables coordinated agent collaboration to generate extended narrative content from initial concepts to video sequences. This maintains both storytelling integrity and visual coherence across multi-scene timelines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
agentic video generation, multi-agent systems, long-form video generation, narrative planning, visual consistency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AiVROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents ViMax, an agentic framework for long-form multi-shot video generation. The system combines hierarchical narrative decomposition and retrieval-augmented planning with a camera-level visual dependency graph, transition-video-based spatial anchoring, and VLM-based best-of-n quality control. The paper also introduces ViMax-Bench, a 35-story benchmark targeting character, background, and multi-person consistency across 8–16-shot narratives. Experiments show improvements over shot-level, keyframe-based, multi-shot, and agentic baselines in automatic consistency metrics and human preference.

Critical strengths

The problem is timely and practically important; the proposed end-to-end workflow is comprehensive and clearly presented; the camera dependency graph provides an appealing mechanism for combining cross-shot reference propagation with parallel execution; human preference results are consistently favorable; and the paper includes extensive ablations, efficiency analysis, backbone comparisons, and hyperparameter studies.

Critical weaknesses
  1. The methodological novelty is somewhat unclear because several major components—multi-agent role decomposition, hierarchical planning, RAG, reference-conditioned generation, and VLM best-of-n—are established techniques.

  2. ViMax-Bench is relatively small (35 stories) and is designed around consistency challenges closely aligned with the proposed method. The main automatic metric is ViCLIP feature similarity, which may reward visual invariance even when story-driven state changes are desirable. Statistical uncertainty and correlation with human consistency judgments would strengthen the conclusions.

  3. The RAG ablation slightly improves the aggregate planning score (4.110 vs. 4.107), although it improves narrative faithfulness. Thus the evidence seems to support a narrower claim about source grounding/faithfulness rather than a general improvement in narrative coherence.

  4. The main contributions of this work appear to lie primarily in system engineering and the integration of existing techniques, rather than in methodological or algorithmic advances. As such, the work seems somewhat misaligned with the typical scope and expectations of ICLR.

Questions for authors

Please see the Critical Weaknesses.

Reviewer_MXubOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ViMax, a multi-agent framework for long-form video generation. It combines hierarchical narrative planning, RAG, a camera dependency graph, transition-video generation, and VLM-based candidate selection. The authors also introduce ViMax-Bench, containing 35 stories and 438 shots.

Critical strengths
  1. The paper addresses an important problem: maintaining narrative, character, and spatial consistency across multiple shots.

  2. The proposed pipeline is intuitive and integrates planning, generation, and quality control.

  3. The evaluation includes automatic metrics, human preference, ablations, and efficiency analysis.

Critical weaknesses
  1. Questionable automatic metrics. CC, IC, and GC are all based on ViCLIP similarity. These metrics may reward visually repetitive videos and do not directly measure identity preservation, object state, or 3D spatial consistency.

  2. Insufficient statistical analysis. The benchmark contains only 35 LLM-generated stories. The paper reports no repeated runs, error bars, confidence intervals, or significance tests. Human evaluation uses only three judgments per pair and omits MovieAgent and AniMaker.

  3. Limited technical novelty. The framework mainly combines established techniques, including hierarchical LLM planning, RAG, reference-conditioned generation, dependency graphs, and VLM-based best-of-(n) selection. The paper does not clearly identify a new learning algorithm, model architecture, or sufficiently novel coordination mechanism. In particular, the claimed multi-agent collaboration appears largely sequential, without a concrete negotiation or joint optimization procedure. The authors should better clarify what technical insight generalizes beyond this particular system implementation.

Questions for authors
  1. What is the central technical novelty beyond integrating existing planning, retrieval, generation, and selection components?

  2. How does ViMax perform on substantially longer videos, such as 30–60 shots?

Reviewer_DiHFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

ViMax is an agentic framework for long-form (multi-shot) video generation. It decomposes production into specialized LLM/VLM agents: screenwriting, shot planning, character/scene keyframe generation, video generation, and VLM-based best-of-n quality control. Two technical pillars support coherence: (1) hierarchical narrative planning (screenplay → events with "process chains" → scenes → shots) augmented with RAG over the source text to recover context lost in decomposition; (2) visual consistency via a camera dependency graph (topological generation with propagated visual anchors) plus transition videos between camera angles to anchor 3D/spatial layout. Validation is on a new 35-story ViMax-Bench (consistency-type × length) with ViCLIP-based CC/IC/GC metrics, a 15-annotator pairwise preference study (525 judgments), novel-to-video storyboard planning on NarrativeQA with an LLM judge, plus ablations, cost/latency Pareto analysis, LLM-backbone study, and best-of-n sensitivity.

Critical strengths

1. A coherent, end-to-end agentic pipeline with an explicit dependency structure. The camera-graph + transition-video combination is a concrete mechanism for a real problem (cross-shot spatial coherence under viewpoint change), and the paper is transparent about which components are ablated.

2. The best-of-n finding (n=2 optimal; more candidates hurt GC) is a useful, non-obvious empirical result that cautions against naive scaling of VLM-judge selection, and the cost/latency reporting is a welcome practical addition.

Critical weaknesses

1. The central consistency evidence is confounded by the method's own conditioning, so the headline quantitative gains do not establish that ViMax produces more coherent video—only that conditioned shots are more similar to their references.

2. The narrative-planning ablations undercut the claimed contribution: the aggregate metric improves when RAG is removed, and the faithfulness gains rest on an LLM judge that shares the planning backbone. This is a logical, not merely statistical, problem with the evaluation design.

Questions for authors
  1. How exactly is the camera dependency graph constructed (edge criteria, anchor representation, cycle handling), and what fraction of shots in ViMax-Bench actually have non-trivial dependencies?
  2. Can you report a consistency metric that is not computed on shots that were directly conditioned on earlier frames—e.g., human-annotated identity/spatial correctness on held-out shots, or a metric on shots generated without reference propagation at matched compute?
  3. Given that removing RAG improves Avg. planning score, on what basis do you claim RAG is beneficial? Can you show a faithfulness–specificity trade-off curve rather than a single operating point?
  4. How do you rule out self-preference bias when GPT-5.4 both plans and judges narrative quality? What happens with a judge from a different model family?
  5. Does ViMax dominate any baseline on the cost–quality–latency Pareto frontier at matched API budget, rather than only within its own scheduling variants?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.