Anime Live Action Pipeline

Anime to Live-Action Pipeline

Overview

This skill provides the standard workflow for converting Japanese anime clips into cinematic live-action short films using the Deep Agents framework.

Pipeline Steps

Step 1: Analyze

Delegate to analyze-agent with the user's anime clip path and mode setting.

Pass:

• clip_path: Path to the input anime video
• mode: The execution mode (local / cloud / hybrid)

The agent returns structured data about scenes, characters, motion, and art style.

Step 2: Design Characters

Delegate to design-character-agent with the analysis result from Step 1.

Pass:

• The characters array from the analysis
• Any style_notes for guidance

The agent returns detailed live-action character sheets with appearance, wardrobe, and expression sets.

Step 3: Generate References

Delegate to generate-reference-agent with the character designs and style notes.

Pass:

• Character design data
• Style notes and reference prompts from the design step

The agent returns paths to generated reference images.

Step 4: Plan Shots

Delegate to plan-shots-agent with the scene analysis, style notes, and reference images.

Pass:

• Scene data from the analysis (scenes array)
• Style notes
• Reference image paths

The agent returns a detailed shot plan with camera, lighting, and composition specs.

Step 5: Generate Clips

Delegate to generate-clips-agent with the shot plan, references, and character designs.

Pass:

• Shot plan from Step 4
• Reference image paths
• Character design data
• Mode setting

The agent returns paths to generated video clips.

Step 6: Review Quality

Delegate to review-quality-agent with the generated clip paths and original shot plan.

Pass:

• Paths to generated clips
• The original shot plan for comparison

The agent returns a quality assessment with pass/fail, score, and feedback.

Quality Loop: If passed is False and iteration < max_iterations, increment iteration and return to Step 5 with the review feedback. If passed is True or iteration >= max_iterations, proceed to Step 7.

Step 7: Assemble Final

Delegate to assemble-final-agent with all approved clip paths.

Pass:

• All clip paths that passed quality review
• Desired transition type

The agent returns the final assembled video path.

Edge Cases

• Invalid input: If the user provides a non-video file or a path that doesn't exist, return an error immediately without running the pipeline.
• Mode switching: The mode can be changed between runs. Always use the mode the user specified.
• Partial results: If a step fails critically, return partial results with an error message rather than crashing.
• Quality loop exhaustion: If max_iterations is reached without passing quality review, proceed with the best available clips and note the quality limitation in the final output.

Mode-Specific Notes

• local: All processing runs on local infrastructure (LM Studio, ComfyUI). Slower but private.
• cloud: Uses cloud APIs (Gemini, Veo). Fastest and highest quality but requires API keys.
• hybrid (default): Analysis runs locally, video generation runs in cloud. Best balance.