Heading 1

The Problem

Training video editing models using SFT differs from both text-to-video and video-to-video tasks. Each training example must align three inputs simultaneously: a text instruction, a source video, and a target video.

‍

The challenge lies in enforcing controlled change. The model must apply the edit specified by the prompt while preserving temporal structure across the entire sequence. Errors are often subtle, accumulate over time, and can introduce incorrect training signals even when individual frames appear acceptable.

‍

Current video models struggle to reliably perform this type of constrained transformation at scale. As a result, high-quality, training-grade SFT data becomes critical. Deccan leveraged its in-house video SFT capabilities to design and produce a dataset specifically structured to help models learn precise, prompt-aligned edits without violating structural consistency.

‍

‍

Comparison of Training Paradigms

‍

‍

Deccan’s Approach

Deccan combined a clear taxonomy of video edit types with a structured delivery process designed to enforce training correctness at scale.

‍

Edit Taxonomy & Constraints

• Manipulation (localized object addition or removal)
• Stylization (appearance changes applied consistently over time)
• Complex / VFX-style edits (motion- or interaction-heavy changes)

Each category imposed different constraints on what could change and what had to remain invariant. These constraints informed prompt design, evaluation criteria, and acceptance standards.

‍

Prompt-Conditioned Generation

Each sample paired a source video with a highly constrained text prompt specifying the edit. Prompts were written in natural language to reflect how end users describe video edits in real environments, ensuring that the resulting SFT dataset trains models to respond to real-world instructions rather than synthetic command formats. Prompt structure varied by edit category and was designed to explicitly bound the transformation.

‍

Generated outputs were treated as candidates, not final artifacts, and entered a delivery pipeline built to enforce these constraints.

‍

Multi-Stage Evaluation and Regeneration

Every generated video passed through multiple evaluation stages aligned to training correctness:

• Early rejection of structurally unsafe outputs
• Structured evaluation based on objective rubrics defined in accordance with client requirements, assessing temporal consistency, structural preservation, and prompt adherence
• Independent quality checks as a second layer of human evaluation to maintain consistent acceptance standards

This two-layer human review process ensured that only samples meeting defined criteria progressed into the final SFT dataset.

‍

Key Takeaways

• Training-grade Video SFT requires explicit control over what must remain unchanged
• Video editing SFT benefits from separating tasks into clear edit categories
• Prompt design directly shapes training outcomes
• Evaluation must account for temporal behavior across full videos
• With the right structure, Video SFT can scale without eroding quality

‍

Conclusion

This engagement demonstrates that prompt-conditioned Video SFT can be delivered at scale when grounded in clear edit definitions, constrained prompt design, and evaluation aligned to how video models learn over time. By combining taxonomy-driven task decomposition with disciplined execution, Deccan produced a large, consistent dataset suitable for training video editing models without relaxing standards as volume increased.