Towards Reason-Informed Video Editing in Unified Models with Self-Reflective Learning

1. Problem Statement

Current unified video generation models, despite possessing powerful internal Vision-Language Models (VLMs) for understanding, struggle with Reason-Informed Video Editing (RVE).

• The Gap: Existing models excel at "literal editing" (e.g., "remove the boat") where the target state is explicitly defined. However, they fail at reason-informed transformations where the outcome must be inferred from physical laws, causal dynamics, temporal progression, or world knowledge (e.g., "Imagine the scene an hour after the boat departed" or "What will the egg look like after frying?").
• Root Causes:
  1. Data Deficiency: Existing datasets lack the reasoning depth required to train models on physical plausibility and causal logic.
  2. Reasoning-Generation Disconnect: In unified architectures, the VLM's reasoning capabilities are not effectively transferred to the generator. Standard Supervised Fine-Tuning (SFT) optimizes for pixel-level reconstruction, providing no signal regarding whether an edit satisfies the underlying reasoning constraints.

2. Methodology: ReViSE Framework

The authors propose ReViSE (Reasoning-Visual Self-Reflective Editing), a self-reflective learning framework that bridges the gap between VLM understanding and video generation.

A. Self-Reflective Learning Loop

Unlike prior methods that rely on external reward models (which are costly and non-differentiable), ReViSE leverages the model's frozen internal VLM as a self-reflective critic.

1. Generation: The generator ($G_\phi$) produces an edited video based on a source video and a reasoning instruction.
2. Clean Estimation: To make the noisy latent output interpretable by the VLM, the system performs a one-step clean video estimation ($\hat{x}_0$) from the noisy latent $z_t$. This is gated to occur only at low-noise timesteps ($t \leq \sigma_{max}$) to ensure semantic coherence.
3. Intrinsic Critique: The frozen VLM evaluates the clean estimate against the instruction across four dimensions:
   • Edit Accuracy
   • Preservation Consistency
   • Generation Naturalness
   • Generation Realism
     The VLM is prompted to provide a Chain-of-Thought reasoning followed by a "Yes" or "No" verdict.
4. Differentiable Reward: Instead of using a hard binary "Yes/No" token (which is non-differentiable), ReViSE uses the token-level probability $P(\text{"Yes"})$ as a soft, differentiable reward signal. This allows gradients to flow directly from the VLM's logits back through the frozen VAE decoder to the generator parameters.

B. Training Objective (Unified Semantic Optimization - USO)

The total loss function combines standard flow-matching loss with a reasoning-aware loss:
$$\mathcal{L}_{USO} = \mathcal{L}_{FM} + \lambda \cdot \mathcal{L}_{reason}$$

• $\mathcal{L}_{FM}$: Standard flow-matching loss for generative fidelity.
• $\mathcal{L}_{reason}$: A binary cross-entropy loss based on the logit difference between "Yes" and "No" tokens. Crucially, this is a relative loss that subtracts a baseline (comparing the source to itself) to isolate the improvement gained by the edit, preventing bias.
• Key Advantage: The VLM and VAE remain frozen; only the generator is updated. This creates a stable, intrinsic reward signal without external critics or policy-gradient estimators.

3. Key Contributions

1. RVE-Bench and RVE-Dataset

• RVE-Bench: The first benchmark dedicated to reason-informed video editing, containing ~1,000 test triplets. It comprises two subsets:
  • RAVE (Reasoning-Aware Video Editing): Requires targeted modifications inferred from physical/temporal/causal reasoning.
  • ICVG (In-Context V2V Generation): Requires generating plausible future frames based on narrative or causal intent.
• RVE-Dataset: A large-scale corpus of 56K reasoning-enriched triplets constructed via a scalable, automated pipeline using GPT-4o to rewrite instructions with implicit reasoning cues.

2. ReViSE Framework

• Introduces a self-reflective, differentiable training loop that uses the model's own VLM as a reward signal.
• Eliminates the need for external reward models, labeled preference data, or additional inference costs at deployment.
• Successfully bridges the disconnect between high-level semantic reasoning and low-level pixel generation.

4. Experimental Results

The authors evaluated ReViSE on RVE-Bench and compared it against state-of-the-art baselines (Omni-Video, InsV2V, InsViE, VACE).

• Quantitative Performance:
  • On the RAVE subset, ReViSE achieved an Overall score of 4.6689, outperforming the base Omni-Video model by 49% and the SFT-trained counterpart by 10%.
  • It showed massive improvements in Edit Accuracy (EA) (+27% over Omni-Video) and Preservation Consistency (PC) (+82% over Omni-Video), proving it can edit correctly without destroying the scene.
  • On ICVG, it also outperformed baselines in EA and Overall scores.
• Generalization:
  • ReViSE maintained strong performance on conventional editing benchmarks (Ditto-1M, VIE-Bench), achieving a 36.7% gain in Overall score over previous bests on Ditto-1M, demonstrating it does not sacrifice literal editing capabilities.
• Human Evaluation:
  • In a user study with 15 experts, ReViSE achieved the highest win rates across Edit Accuracy, Content Preservation, and Visual Quality compared to all baselines.
• Ablation Studies:
  • Confirmed that the Unified Semantic Optimization (USO) objective is superior to standard SFT or Reward-Weighted Regression (RWR).
  • Identified the optimal noise gating threshold ($\sigma_{max} = 0.5$) to balance signal reliability and training coverage.

5. Significance

This work addresses a critical bottleneck in generative AI: the inability of current models to perform edits that require common sense, physics, and causal reasoning.

• Paradigm Shift: It moves video editing from "pixel manipulation" to "reasoning-guided transformation."
• Efficiency: By using the internal VLM as a differentiable critic, it avoids the computational overhead and data requirements of external reward models or Reinforcement Learning (RL) pipelines.
• Scalability: The automated construction of the RVE-Dataset and Benchmark provides a foundation for future research into complex, logic-driven video generation tasks.

In summary, ReViSE demonstrates that unified models can be trained to "think" about the physical and causal consequences of an edit, significantly advancing the state-of-the-art in instruction-guided video editing.