AI AGENT SKILLS

xeon-smartupscale_v2

一个面向 Dev Tools 场景的 Agent 技能。原始说明:Experiment-faithful 1080p upscaler for Intel Xeon CPUs. Decodes frames, resizes them in memory to 480x270 with lanczos, then runs Real-ESRGAN general-x4v3 x4...

SKILL.md

SKILL.md


name: xeon-smartupscale_v2
description: Experiment-faithful 1080p upscaler for Intel Xeon CPUs. Decodes frames, resizes them in memory to 480x270 with lanczos, then runs Real-ESRGAN general-x4v3 x4 in OpenVINO CPU mode with a BF16 hint.
metadata:
openclaw:
requires:
bins: [bash, python3, curl, tar, unzip]


Goal

Reproduce the exact 1080p experiment pipeline:

  1. Decode source frames
  2. In-memory lanczos -> 480x270
  3. Real-ESRGAN general-x4v3 x4 -> 1920x1080
  4. Encode with H.264 medium, crf 15

Usage

bash <skill_dir>/install.sh
bash <skill_dir>/smartupscale.sh <input.mp4>

Optional custom output path:

bash <skill_dir>/smartupscale.sh <input.mp4> -o <output.mp4>

Packaging Notes

  • model/realesr-general-x4v3_480x270.xml and .bin are vendored for the fixed 1080p path.
  • No dynamic export path is included. This package stays on the fixed experiment model.
  • For ClawHub publishing, the raw .bin weights can be omitted from the package and restored from model/realesr-general-x4v3_480x270.bin.b64 or chunked model/realesr-general-x4v3_480x270.bin.b64.part-* files during install.sh.

Output

  • Final video: 1920x1080, H.264 medium, crf 15
  • Audio: copied from source when possible
  • Lanczos pre-scale is done in memory before inference; no encoded intermediate stage is produced.