Wan Animate 2 — Apple Silicon Lab
A field report on local Wan Animate 2 experiments with Apple M5, ComfyUI, and MPS.
32GBのMacBook Airでローカル動画生成を検証。動画を書き出せたことと、キャラクターの一貫性に残る課題を記録しました。
Overview
A local AI video field report that keeps execution success and visual quality separate. The specialist site holds the experiment records, limitations, and public workflow templates.
動画を書き出せたことと、使える品質であることを分けて記録した実験です。
Tested environment
Apple M5 MacBook Air with 32 GiB unified memory, running ComfyUI / MPS. The test envelope was 512 × 512, 17 frames, 8 fps on this single machine.
Technical outcome
V2 and V3 completed loading, sampling, VAE decoding, and video saving: Technical PASS. Approximate observed runtimes were 02:55:37 for V2 and 02:23:41 for V3. Each is an individual run, not an average or a performance benchmark.
Visual limitations
Technical PASS ≠ Visual SUCCESS. V2 retained identity relatively well through F11, then collapsed from F12 onward, with severe breakdown at F17. V3 reduced subject oversizing but worsened identity, with face/hood collapse from F1 and reduced motion readability.
V4 passed historical workflow static/UI validation but remains UNRUN. The public template received static checks only; there is no V4 generated result or evidence of improvement.
Key observations
- Post-run swap was 17.76 GiB (V2) and 17.69 GiB (V3). These are system observations after execution, not peak memory or exclusive model usage.
- Six identified model files total 31.758971 GiB of storage; this is not a runtime RAM requirement.
- Two short completed runs do not establish production readiness, stable identity, or broad Apple Silicon performance.
Read the full field report
Explore the field report for the evidence, caveats, and experiment lineage. Browse the GitHub source for the public records and templates.
詳しい検証記録は専門サイトでご覧ください。非公開のキャラクター画像・動画は使用していません。