LLM Video Generator

Generate videos via ZhipuAI CogVideoX-3. Each API call produces ~5s of video.
For longer videos, chain multiple calls using last-frame continuation, then concatenate.

Scripts

All scripts use /opt/anaconda3/bin/python3. Resolve <skill-dir> to this skill's directory.

Workflow

Step 1: Assess Request & Clarify

Clear request → proceed to Step 2. A request is clear when:

• - Video content/scene is described with enough detail
• Style or visual tone is specified or implied
• Duration is stated (default: 5s if not specified)

Vague request → propose a plan first:

CODEBLOCK0

Iterate with the user until confirmed.

Step 2: Estimate Time & Notify User

Before starting generation, calculate and report the estimated time:

Time estimation formula:

• - Base: 1 minute per second of video (e.g., 20s video ≈ 20 minutes)
• High-definition (4K or 60fps): add +30% (e.g., 20s 4K video ≈ 26 minutes)
• Additional overhead: ~2 minutes for frame extraction, concatenation, and compression
• Segments: ceil(target_duration / 5)

MUST send this message to the user before starting generation:

CODEBLOCK1

Example for a 30s 1080P video:

• - 6 segments, base time = 30 minutes, +2 min overhead → ~32 minutes
• Message: "预计总耗时：约 32 分钟"

Example for a 20s 4K video:

• - 4 segments, base time = 20 * 1.3 = 26 min, +2 min → ~28 minutes

Step 3: Plan Generation Segments

Each API call produces ~5 seconds. Calculate segments: INLINECODE5

For multi-segment videos, plan how the content evolves across segments. Write a prompt for each segment describing what happens in that 5-second window, maintaining visual continuity.

Step 4: Execute Generation with Progress Reports

CRITICAL: After each segment completes, IMMEDIATELY send a progress message to the user before starting the next segment. Do not wait until all segments are done.

Progress message format (send via message tool or inline reply after each segment):

CODEBLOCK2

Generation process:

Segment 1 — Text-to-Video:

CODEBLOCK3

→ Send progress message to user

Segments 2+ — Image-to-Video (last-frame continuation):

For each subsequent segment:

1. 1. Extract last frame from the previous segment's video:

1. 2. Generate next segment using the last frame as input:

1. 3. → Send progress message to user

Repeat for all segments.

Alternative — Frames-to-Video mode:

If you have both a starting and ending image for a segment:
CODEBLOCK6

Step 5: Concatenate Segments

After all segments are generated, combine them:

CODEBLOCK7

If the final file exceeds 25MB (Feishu upload limit), compress with ffmpeg:
CODEBLOCK8

Step 6: Deliver

• - Share the final video file with the user
• For Feishu delivery: use feishu-send-file skill to send the .mp4 file
• Final report:

CODEBLOCK9

Prompt Tips

• - Use English prompts for best quality (translate Chinese descriptions)
• Be specific: scene, camera angle, lighting, motion, atmosphere
• Include style keywords: cinematic, realistic, cartoon, watercolor, etc.
• For continuation segments, describe the action progression, not the full scene from scratch
• Keep each segment prompt concise (1-3 sentences)

Parameters Reference

Error Handling

• - Missing ZHIPUAPIKEY: Ask user to set environment variable
• Missing zai-sdk: pip install zai-sdk (under anaconda)
• Missing ffmpeg: Required for frame extraction and concatenation
• Task timeout: Increase --max-wait or retry; check task status manually via API
• Task failed: Simplify the prompt and retry
• File too large for Feishu: Compress with ffmpeg (reduce resolution or increase CRF)