acceptodds
Withdrawn from ICLR 2027

VisionCreator-P1: Closing the Loop for Long-Horizon Physical Visual Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-horizon physical visual generation requires modeling how a scene should evolve under physical laws, not just how each frame looks. However, open-loop models lack explicit physical feedback: small local violations accumulate across steps and progressively derail multi-step trajectories. To address this, we present **VisionCreator-P1**, a closed-loop physics-informed agent that casts multi-image trajectory synthesis as a process of verified state transitions. At inference, the agent alternates milestone planning, next-state proposal, and transition-level physics reflection, committing a state only after iterative verification against the given query. For training, we address the credit assignment bottleneck that destabilizes long-horizon agentic reinforcement learning by decomposing optimization into short, prefix-conditioned blocks. Each block branches from a shared verified prefix and competes under a physics-aware hierarchical reward, thereby isolating local transition errors from accumulated drift and enabling stable policy learning. Experiments show that VisionCreator-P1 achieves state-of-the-art performance on physical accuracy and temporal coherence, establishing closed-loop local transition verification as a core paradigm for physically grounded, long-horizon generation.

Primary area:
generative models
Keywords:
Visual Generation, Physics-informed AI, Agentic Reinforcement Learning, Trajectory Synthesis, Closed-loop Verification

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_5xzKOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a closed-loop, physics-informed agent framework for long-horizon physical visual generation. The authors employ reinforcement learning to train VisionCreator-P1, aiming to address the severe problem of Global Causal Drift. They decompose the optimization process into short sub-blocks to improve credit assignment during training. Additionally, they introduce two datasets for this task.

Critical strengths
  1. Provides datasets for future research.
Critical weaknesses
  1. The paper overstates its novelty, and I found few new insights in it. Using an agent to generate physical processes is a common idea, yet the paper introduces numerous elaborate terms to describe it. “Global Causal Drift,” for example, is essentially accumulated error, which is not new. I do not think the work offers substantially new ideas to the community. Moreover, this approach merely shifts the challenges of video generation to the agent rather than resolving them, placing the burden on the VLM without guaranteeing reliability.
  2. In the figure, the text isn't aligned, and the boxes aren't aligned either. The fonts also change. I think it needs more polishment.
Questions for authors
  1. How reliable is the VLM-based transition verifier? Although the paper includes held-out and human evaluations, it does not directly quantify the verifier’s false acceptance rate on physically invalid transitions.
  2. Table 2 does not include a comparison with the latest Seedance, H3, and Kling models. As such SOTA video models have temporal attention, I think they will better simulate the physical process.
  3. Could a recent VLM, such as GPT-5.6, outperform the RL-trained Qwen-VL model in a zero-shot setting?
Reviewer_9FvsOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

VisionCreator-P1 uses the Qwen3-VL-32B agent and Qwen-Image to solve the sparse physical trajectory generation problem. PRTC plans milestones and iteratively checks the transitions before accepting them. PFBR trains the agent using candidate actions sampled from a shared prefix. The paper reports improvements on PhysBench-M and VCR-Bench.

Critical strengths
  1. The model shows improvements in both physical and visual benchmarks.
  2. It has the potential to help video generation with physical plausibility.
Critical weaknesses
  1. The training data, reward, and verification loop change at the same time. The ablations show that the full system works better, but they don’t clearly separate these effects. I’d like to see comparisons under the same sampling and tool-call budget, especially with verification turned off.
  2. P1 is trained on specific tasks while the video baselines are frozen and generate in one pass. The comparison is unfair.
  3. The paper only shows physical improvement only on a self-designed benchmark.
Questions for authors

1.What does the verification loop add on its own? 2.How much of the gap to the video baselines comes from task-specific training? Do you have comparisons with task-adapted video baselines? 3.Do the physical improvements carry over to other existing benchmarks?

Reviewer_wS5SOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

In this paper, the authors propose an agent framework to generate multi-image trajectories as a process of verified state transitions. Given a prompt, this framework will draft a plan to generate images at different stages and use LLM-as-a-judge to evaluate generated images at each stage and revise if needed.

This work also shows the post-training stage by decomposing the full trajectory into block-level and applying visual rewards at substages. Then, the model is finetuned with a GRPO-style method.

Critical strengths

The method shows improvements over base video generation models on PhysBench-M.

PhysData-M and PhysBench-M will be released for training and evaluation purposes.

Critical weaknesses

In my opinion, the technical contributions in this paper are very limited. The idea of decomposing a user input into a plan, generate image at each stage, and using an evaluator to provide feedback is not rare. There are tons of papers doing agentic frameworks like this, and I don't see impressive technical novelty in this work.

The figures are AI-generated and very blurry with incorrect details, for ex in Figure 2, it should be text2img_tool not test2img_tool.

Whether this pipeline can generalize to other categories is not discussed.

I'm not sure about the motivation for generating some keyframes in the trajectory instead of full video synthesis. If the final goal is to generate full videos, the authors should further use keyframes as a conditioning for full video generation.

Questions for authors

See weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.