acceptodds
Under review as a conference paper at ICLR 2027

VideoRISE: Recursive Improvement through Self-Evaluation for Black-Box Video Inversion

Abstract

Recovering a prompt that reproduces a target video is difficult when a video generator is accessible only through queries: gradients and model internals are unavailable, each generation is stochastic, and semantic, appearance, motion, and temporal errors can improve or regress independently. We introduce VideoRISE (Recursive Improvement through Self-Evaluation), a budgeted control layer for black-box video prompt inversion. VideoRISE executes candidate prompts, evaluates target–output videos with a fixed multi-metric monitor, accepts a revision only after verification, and retains the best checkpoint together with failed-attempt evidence. Its reflective VLM proposer diagnoses differences in entities, action, style, camera, and lighting, then emits structured natural-language edits. On a 60-video, six-category protocol with Wan 2.2 Lightning T2V and I2V backends, the method improves the strongest one-shot VLM baseline from 0.607 to 0.683 in overall inversion score under 100 generations. We treat the score as a search and selection proxy rather than a perceptual ground truth, and report query, VLM-call, wall-clock, and estimated-cost accounting. Downstream prompt editing is shown qualitatively; broader multi-task use remains future work.

Primary area:
generative models
Keywords:
generative models, multimodal large language models, video generation, black-box inversion

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_PCMFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes VideoRISE, a "control layer" for black-box video prompt inversion

Critical strengths

(1) The paper studies a practical problem: recovering an editable prompt for a target video when the generator is only accessible as a black box

(2) The framework is modular and attempts to report generator calls, VLM calls, runtime, and monetary cost, which is useful for evaluating query-intensive methods

Critical weaknesses

(1) Why does Section 4 claim to report pp-values, Wilcoxon tests, 95% bootstrap intervals, sweeps over B∈10,25,50,100B \in {10, 25, 50, 100} and K∈2,3,5K \in {2, 3, 5}, tier escalation frequencies, and trajectory/failure figures when none of them exist in the paper? Maybe I miss some parts?

(2) Scores computed from different metric subsets are not necessarily comparable, yet the exact reevaluation and checkpoint-update policy is not given. Table 4 even presents “Fixed evaluation” as an ablation, further confusing what “fixed” means.

(3) The source of the 60 target videos is unclear, even though it matters whether they are real videos or outputs of the same Wan

(4) Videorise is essentially an existing VLM prompt optimizer wrapped with best-so-far retention, rollback, and patience-based early stopping. These are standard search practices. Moreover, the main search-only baseline removes checkpoint recovery, making it unnecessarily weak: any reasonable optimizer should return its best observed candidate rather than its final mutation.

Questions for authors

See above, and one more question: why the figure 3/4/5 are so blurred and the framework figure 1/2 look strange? Can author elaborate more on this issue?

Reviewer_1dpqOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VideoRISE addresses black-box video prompt inversion: recovering a readable prompt whose generated video resembles a target clip, without accessing generator internals. It separates prompt proposal from execution and selection. A reflective VLM proposes structured edits, while a controller evaluates generated videos, maintains a best checkpoint and Pareto alternatives, records failed attempts, and enforces stopping and query budgets.

The evaluation uses 60 self-collected videos and Wan 2.2 Lightning T2V and I2V backends. Table 2 reports an aggregate utility improvement from 0.661 for GEPA-style search to 0.701 for VideoRISE. The paper also reports controller ablations, human evaluation, and reduced query, runtime, and estimated monetary costs. Downstream editing is demonstrated qualitatively.

Critical strengths
  • A practical and clearly scoped system formulation. Separating a reflective proposer from execution, verification, and recovery is useful for costly black-box generation. The explicit workspace links prompts, generated artifacts, scores, and failed revisions, making the optimization process easier to inspect and recover. The paper appropriately limits its claims to prompt inversion and treats downstream editing as a qualitative application.

  • Evaluation considers both reconstruction quality and resource usage. The paper reports component ablations and results on both T2V and I2V backends, alongside human judgments and query/time/cost accounting.

Critical weaknesses
  • Video demonstrations in Figure 3,4 and 5 are very blur.
  • 60 self-collected videos as evaluation benchmark is not enough.
  • Relevant details are deferred to supplementary material, but there is not supp in the submitted pdf.
  • Controller gains are not clearly isolated from basic best-so-far selection. It is unclear whether search-only returns its best evaluated candidate or its latest state. A matched baseline with simple best-so-far retention is needed.
  • The stated 100-generation budget also needs reconciliation with the one-shot baseline’s single query in Table 5.
  • The evaluation protocol remains underspecified. The differing scores across Tables 1–3 (0.683 versus 0.701/0.682) also require explanation.
Questions for authors
  • Controller attribution: What does search-only execution return at termination? How are the one-shot baseline’s reported one-query usage and the stated 100-generation comparison reconciled?

  • Evaluation clarification: What settings explain the different results in Tables 1–3? Please specify the metric formulas, normalization and aggregation weights, target/output temporal alignment, and I2V conditioning inputs. What comparison reference and aggregation procedure produce each human-preference value?

Reviewer_1eViOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies prompt inversion for black-box video generators. VideoRISE combines VLM-guided prompt revisions with multi-metric evaluation, candidate retention, checkpoint recovery, and outcome memory, reporting improvements on 60 videos across two Wan backends.

Critical strengths
  1. The problem is practically relevant, given the cost and stochasticity of video generation. Separating prompt proposal from candidate acceptance and preserving the best observed result are sensible design choices.
  2. The experiments examine several controller components while keeping the proposer fixed, including selection, recovery, memory, Pareto alternatives, and structured edits. Reporting query counts, runtime, and estimated cost also provides useful context for the quality improvements.
Critical weaknesses
  1. The search-only baseline removes selection, recovery, and outcome-history mechanisms, whereas GEPA already incorporates evaluated candidate retention and Pareto-based selection. Without a comparison against a complete video adaptation of GEPA, the reported improvement does not clearly establish the proposed controller’s contribution beyond existing optimization mechanisms.

  2. The main results are difficult to interpret under a consistent evaluation protocol: Tables 1 and 2 report different aggregate scores (0.683 versus 0.701), and their 100-generation budgets are not clearly reconciled with the single one-shot query reported in Table 5. These ambiguities directly affect whether the claimed gains reflect better inversion or differences in sampling, selection, and computational budget.

Questions for authors

Clarification requests

  1. What explains the different VideoRISE scores in Tables 1 and 2 (0.683 versus 0.701)? Please clarify the corresponding backends, seed aggregation, and output-selection procedures, including whether the one-shot baselines use repeated sampling under the 100-generation budget or the single generation reported in Table 5.

  2. How are the five evaluation metrics computed, normalized, and weighted, and how is Fidelity in Table 2 defined? Are final outputs from cheap, full, and adaptive evaluation assessed using the same complete metric suite?

  3. Which mechanisms are retained in the GEPA-style search baseline in Table 2, and precisely what changes when VideoRISE is added? This would help establish which components account for the reported improvement from 0.661 to 0.701.

  4. Which method pairings and comparison counts underlie the human preference scores in Table 2? Does VideoRISE’s score of 0.672 represent a direct comparison against GEPA-style search, or an aggregate across different opponents?

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

This paper presents VideoRISE, a black-box video prompt inversion framework that iteratively generates videos, evaluates their similarity to a target video, and refines prompts using VLM feedback. Its main components include multi-metric evaluation, quality-gated selection, checkpoint recovery, and outcome memory. Experiments on 60 videos using two Wan 2.2 backends report improvements over one-shot prompting and GEPA-style search baselines.

Critical strengths

The framework provides a practical, modular implementation of iterative video prompt optimization, with explicit budget management, rollback mechanisms, and cost accounting. The ablation studies suggest benefits from these components.

Critical weaknesses

Main weakness is limited novelty relative to existing work. The core generate–evaluate–refine loop closely resembles several prior methods. ARPO (Ren et al., 2025) already performs target-conditioned prompt inversion through iterative generation, feedback, and greedy selection in the image domain. More importantly, VISTA (Long et al., CVPR 2026) implements iterative video prompt refinement using multi-dimensional critiques and best-candidate selection, although its objective differs from inversion. GEPA (Agrawal et al., ICLR 2026) already incorporates reflective feedback, candidate retention, and Pareto-based selection. The proposed quality gates, checkpoints, and memory therefore appear largely to be adaptations or combinations of established techniques. The paper does not sufficiently establish what constitutes a fundamentally new algorithmic contribution.

Questions for authors

What is the fundamental methodological novelty beyond ARPO, VISTA, and GEPA? Can the authors compare against a faithful implementation of GEPA and an inversion-adapted VISTA under identical evaluation and generation budgets?

Can the authors evaluate on VI-Bench and provide matched-cost comparisons, including best-of-N selection, to demonstrate that the gains come from the proposed controller rather than additional search and metric-guided candidate selection?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.