GREENWAKAME LABRESEARCH / FIELD REPORTS

FIELD REPORT 01 / LOCAL AI VIDEO

WANANIMATE 2ON APPLE SILICON

M5・32GBでローカル生成は完走。
キャラクターの同一性は未解決。

TECHNICAL
PASS
VISUAL IDENTITY
NOT SOLVED

APPLE M5 / 32GBMACBOOK AIR · 512 × 512 · 17F

Read the result Artifact index

THE CENTRAL FINDING

TECHNICAL PASS ≠
VISUAL SUCCESS.

Load → sampling → VAE decode → saved video: completed.

動画を書き出せたことと、品質が使えることは別の判定です。今回の完走は、実用性やキャラクターの同一性を保証しません。

INT8の保存形式だけでは、実際の演算がネイティブINT8だとは確認できません。

01 / TESTED SYSTEM

One machine.
A narrow test envelope.

この1台と固定した環境での実行記録です。Apple Silicon全体への対応を示す検証ではありません。

CHIP / MEMORY
Apple M5 / 32 GiB
MacBook Air · arm64 · Mac17,3
EXECUTION
MPS / shared RAM
Built true · available true on host
SOFTWARE
ComfyUI 0.37.0
Python 3.12.14 · macOS 27.0.1 / 26A434
VIDEO ENVELOPE
17 frames / 8 fps
512 × 512 · batch 1 · 2.125 seconds
Open the frozen software record

torch 2.15.0.dev20260929 · torchvision 0.30.0.dev20260929 · torchaudio 2.11.0 · ffmpeg / ffprobe 9.0.2.

ComfyUI HEAD 73c9bad4d21e7addbe1d13bc92eee0f1431b017d

The host MPS result is authoritative. A sandbox reported unavailable; it was not treated as a host failure. The clean-checkout state and future nightly-wheel availability are unverified.

ENVIRONMENT PINS · TXT ↓

02 / EXPERIMENT LINEAGE

A driver. Two runs.
One deliberate stop.

成功したdriverから2回の生成へ進み、次の実験で停止しました。この順序は、品質が改善し続けたことを意味しません。

  1. 01

    VACE

    VACE-512

    DRIVER READYBaseline record ↓
  2. 02

    ANIMATE V2

    ANIMATE2-V2

    TECHNICAL PASSVISUAL PARTIALLate collapse ↓
  3. 03

    ANIMATE V3

    ANIMATE2-V3

    TECHNICAL PASSVISUAL WORSE / DAlignment test ↓
  4. 04

    ANIMATE V4

    ANIMATE2-V4-UNRUN

    STATIC / UI VALIDATEDUNRUNThe stopping point ↓

V4 validation refers to the historical private workflow. The public templates were only statically checked.

Read the earlier VACE and engineering chronology

VACE 1.3B baseline → reference identity comparisons → OpenPose crouch / jump / landing controls → 384 to 512 → successful VACE driver → HyperFrames integration proof → Animate feasibility investigation → INT8 installation → sampling mapping → v2 → v3 → unrun v4 → project pause.

HyperFrames is project context, not evidence of Animate visual quality.

DOWNLOAD NOTES · EXPERIMENT LOG · MD ↓

03 / VISUAL DIAGNOSTICS

The identity problem
survived completion.

元の非公開動画を目視で確認した結果です。ここでは区間と症状を示し、キャラクター画像や自動評価スコアは公開しません。

ANIMATE2-V2 / BASELINE

It held.
Then it collapsed.

TECHNICAL PASSVISUAL PARTIAL

F1–F11: relatively good identity.
F12 onward: collapse.
F17: severe breakdown.

短い生成は完走したものの、F12以降で同一性が崩れました。書き出しの成功だけでは、時間方向の品質は判断できません。

  1. F1–F11Relatively good identity
  2. F12+Collapse begins
  3. F17Severe breakdown
Frame markers locate the reported observations. They are not per-frame similarity measurements.
OBSERVED RUNTIME02:55:37
POST-RUN SWAP17.76 GiB
REFERENCE / DRIVERPadded / RGB
VIEW NOTES — Hypothesis & fixed conditions
HYPOTHESIS
The INT8 Animate 2 base path can technically complete a short MPS generation and retain reference identity.
BASELINE / CHANGED
Base-path feasibility / identity baseline. Driver was staged byte-identically from output into input; no re-encoding.
FIXED CONDITIONS
512 × 512 · 17 frames · batch 1 · 8 fps · 2.125 seconds; seed 12345; 20 steps; CFG 1; shift 5; simple / DPM++ 2M SDE (midpoint, eta 0, s_noise 1, CPU noise); denoise 1; no LoRA; pose strength 1; pose start 0; reference strength 1. pose end 1; original padded reference.
TECHNICAL / VISUAL
PASS: technical generation completed. F1-F11 relatively good identity; collapse from F12 onward; severe collapse at F17.
MEMORY / LIMITATION
Recorded before / after indicators appear below. Post-run is not peak. One observed run, approximate runtime; private assets and prompts are required for exact historical reproduction.

Output fingerprint 2a5195ed39e3e6522bdfd603bc0a92d5ca5f5cd2cbe07e098a3c8179a3791140 (160,876 bytes; video excluded).

DOWNLOAD WORKFLOW · JSON ↓DOWNLOAD JSON · EXPERIMENTS ↓DOWNLOAD NOTES · IDENTITY · MD ↓

ANIMATE2-V3 / REFERENCE ALIGNMENT

Better framing.
Worse identity.

TECHNICAL PASSVISUAL WORSE / D

Subject oversizing reduced.
Face / hood collapse from F1.

右端に縦のartifactも生じ、正常な同一性は回復せず、動きの読み取りも悪化しました。固定seedの1回の結果から、小さい参照画像が一般に不利だとは言えません。

Reference geometry only. Abstract preparation diagram; no character image.
OBSERVED RUNTIME02:23:41
POST-RUN SWAP17.69 GiB
CHANGED VARIABLESpatial reference
VIEW NOTES — Hypothesis & fixed conditions
HYPOTHESIS
Spatial alignment of the reference to driver F1 may reduce subject oversizing and preserve identity.
BASELINE / CHANGED
ANIMATE2-V2; replace the reference with the aligned 174×342 preparation. Reference SHA changes; seed and RGB driver remain fixed.
FIXED CONDITIONS
512 × 512 · 17 frames · batch 1 · 8 fps · 2.125 seconds; seed 12345; 20 steps; CFG 1; shift 5; simple / DPM++ 2M SDE (midpoint, eta 0, s_noise 1, CPU noise); denoise 1; no LoRA; pose strength 1; pose start 0; reference strength 1. pose end 1.
TECHNICAL / VISUAL
PASS: technical generation completed. D / worse: subject oversizing reduced, but face/hood collapse from F1; right-side vertical artifact; no normal identity recovery; reduced motion readability.
MEMORY / LIMITATION
Recorded before / after indicators appear below. Post-run is not peak. One observed run, approximate runtime; private assets and prompts are required for exact historical reproduction.

Output fingerprint 45cf2edab69190c0bd403f80d65e105ad2b072a919a015704b939d3be8557b5f (141,731 bytes; video excluded).

DOWNLOAD WORKFLOW · JSON ↓DOWNLOAD JSON · EXPERIMENTS ↓DOWNLOAD NOTES · IDENTITY · MD ↓
VACE-512 — the successful driver baseline

The earlier VACE 1.3B experiment produced the 512×512 RGB driver used by both Animate runs: 17 frames, 8 fps, 2.125 seconds. It is a successful baseline within this project, not a universal identity-quality claim. Runtime and memory were not frozen for this record.

Hypothesis: matching the 512-pixel control and reference could produce a short usable driver. Changed condition: 384→512. Settings: seed 12345, 20 steps, CFG 6, uni_pc / simple, shift approximately 8, denoise 1, strength 1.

Driver SHA-256 d27024c4787635be5c881d94c5ad54ca5b99670b3482075c5b015694c2aed7a4

DOWNLOAD WORKFLOW · VACE · JSON ↓

04 / MEMORY & RUNTIME

The run finished.
The cost was substantial.

どちらも1回の観測値です。17フレームの生成に2時間以上を要し、終了後には大きなswapが残りました。平均性能を示すベンチマークではありません。

V2 RUNTIME02:55:37~10,537 seconds
V3 RUNTIME02:23:41~8,621 seconds
SIX MODEL WEIGHTS31.758971GiB · storage, not runtime RAM

V2 / SWAP

0 → 17.76 GiB

V3 / 0 → 17.69 GiB

POST-RUN ≠ PEAK

終了後の測定値であり、生成中の最大値ではありません。swapはディスク上の領域で、各指標を足して必要RAMにはできません。

ANIMATE2-V2

0–32 GiB

AVAILABLE

BEFORE13.00 GiB
AFTER6.42 GiB

COMPRESSED

BEFORE3.03 GiB
AFTER15.78 GiB

SWAP

BEFORE0.00 GiB
AFTER17.76 GiB

ANIMATE2-V3

0–32 GiB

AVAILABLE

BEFORE13.09 GiB
AFTER6.30 GiB

COMPRESSED

BEFORE2.88 GiB
AFTER15.57 GiB

SWAP

BEFORE0.00 GiB
AFTER17.69 GiB

Common visual scale: 0–32 GiB. Bars compare recorded before / after values; they do not estimate peak or exclusive physical usage.

One physical memory pool.

MPSとCPU offloadは同じunified memoryを共有します。offloadしても、別のRAMが増えるわけではありません。

Different system measures.

RSSの合計は排他的な物理使用量ではなく、memory_pressureのfree率もActivity Monitorの色とは別の指標です。他のプロセスの影響も受けます。

DOWNLOAD NOTES · MEMORY · MD

05 / ENGINEERING FINDINGS

Keep the failures.
Keep the knowledge.

01

TECHNICAL FEASIBILITY

A short run
is possible.

M5・32GBで、512×512・17フレームの生成は完走しました。ただし約2時間24分〜2時間56分と大きなswapを伴い、快適な運用まで確認できたわけではありません。

02

VISUAL IDENTITY

A saved video
is not enough.

生成完了と品質は別問題です。v2ではF12以降で同一性が崩れ、最後のF17では深刻な崩壊が見られました。

03

Alignment did not fix identity.

v3は人物の大きさを改善しても、同一性は悪化しました。サイズ・配置・余白の効果を個別に切り分けた検証ではありません。

04

UI and backend disagreed.

backendの検証は通っても、UIでは素材を読めませんでした。byte-identicalなinputへの配置で通りましたが、原因は未確定です。

05

Swap was a practical constraint.

両runとも終了後のswapが17 GiBを超え、長い実行時間と圧縮メモリを伴いました。peakを測定した値ではありません。

06

The next result does not exist.

pose_end=0.7は確認済みの設定であり、生成結果ではありません。効果は将来の実験で検証する仮説として残っています。

VIEW NOTES — Sampling & LoadVideo

Closest approximation

20 steps · CFG 1 · shift 5 · simple · DPM++ 2M SDE · midpoint · eta 0 · s_noise 1 · CPU noise · denoise 1 · seed 12345 · add_noise true.

Upstream guide=0 conditional-only semantics map to CFG=1 here. The shift transform is equivalent; the compared 20-step sigma sequence matched. Solver equivalence was not established: first-sigma correction and integer vs floating timestep differences remain.

DOWNLOAD NOTES · SAMPLING · MD ↓

Byte-identical staging

The historical VACE output was staged into input without re-encoding. Source and staged driver share bytes and SHA. Backend PASS did not establish UI availability or explain the root cause.

Staging retains embedded workflow / prompt metadata. It does not sanitize video for release.

DOWNLOAD NOTES · LOADVIDEO · MD ↓

06 / REPRODUCTION & ARTIFACTS

A route back.
Not an exact replay.

モデル・環境・workflow templateを再構築するための記録です。完全再現は保証せず、非公開の素材やpromptは自分の入力に置き換えます。

  1. 01

    Recover the environment.

    Use the recorded ComfyUI commit and Python pins. Preserve the existing source and environment; nightly availability and a clean checkout are not guaranteed.

  2. 02

    Retrieve and verify the weights.

    Use the six-record model manifest for revision URLs, exact filenames, categories, bytes and SHA-256. Verify before loading, then place each file beneath models / its category.

  3. 03

    Prepare private inputs.

    Provide reference.png, driver.mp4, or pose-control.mp4 as required. v3 needs reference-aligned.png. Historical media and prompts remain private.

  4. 04

    Adapt the reference template.

    Replace CHARACTER_DESCRIPTION and any negative-prompt placeholder. Check host memory feasibility before a separately decided future test.

The tested files are Comfy-Org distributions packaged for ComfyUI; they are not asserted byte-identical to original upstream Wan checkpoints. Public templates are not drop-in exact reproductions.

07 / LIMITS OF THE RECORD

What this test
cannot tell you.

確認できたのは短い生成の完走と、観測した失敗です。崩壊の原因や、別条件での品質・性能は未確認として分けて残します。

  • No long-video or multi-seed evidence.
  • No confirmed cause for v2 collapse or v3 degradation.
  • No peak memory, exclusive physical usage or per-stage timing.
  • No official exact upstream sampling reproduction.
  • No proof of native INT8 kernel arithmetic.
  • No guarantee of future nightly-package restoration.
  • No public rights decision for private inputs or generated media.
DOWNLOAD NOTES · LIMITATIONS · MD

ANIMATE2-V4-UNRUN / NEXT HYPOTHESIS

1.0 0.7
Still unrun.

End pose conditioning earlier in denoising. Test whether late identity stability changes.

設定とUIの確認までは完了しましたが、メモリ条件の評価Cで生成を中断。その後、ハードウェア・実行時間・ディスクの制約でプロジェクトを停止しました。

VALIDATEDUNRUN
BASELINE
ANIMATE2-V2
CHANGED VARIABLE
pose_end_percent
VALIDATION
Historical static / UI PASS
OUTPUT / RUNTIME / VISUAL
None / unknown / untested

0.7はdenoisingの進行割合です。動画の70%の時点や、フレームの区切りを指定する値ではありません。

Historical private workflow: static / UI PASS.
Public template: static checks only.

DOWNLOAD WORKFLOW · V4 / UNRUN