M4 Pro 24GB: fold Turbo LoRA (PR #14) benchmark — 124f is the sweet spot on both 480P and 768P

AideYu

# M4 Pro 24GB: fold Turbo LoRA (PR #14) 4-spec benchmark — 124f is the sweet spot on both 480P and 768P Machine: Mac mini M4 Pro, 24GB unified memory, macOS 26.6.2, `--ssd-streaming`. Test: `MiniMax-H3-turbo` (fold Turbo v4), FL2VA first/last frames, seed 42, 6 steps, ET5T shot 1 (strong-motion drift, gravel dust). Wall-clock = process birth → `h3: wrote mp4`. | Resolution | Frames | Steps | Wall-clock | |---|---|---|---| | 832×480 | 124 | 6 | 19.4 min (01:12→01:31) | | 832×480 | 243 | 6 | 46.0 min (07:31→08:17) | | 1344×768 | 124 | 6 | 75.5 min (08:47→10:03) | | 1344×768 | 243 | 6 | 206.9 min (~3.4 h) (10:50→14:17) | Key observations: 1. **124 frames is the sweet spot for BOTH 480P and 768P.** Per-frame cost and wall-clock stay predictable at 124f; jumping to 243f is clearly superlinear on this memory budget. If you need ~10s, plan for a near-linear cost jump in time (and a healthy swap peak). 2. **Single-step linear extrapolation fails for long sequences.** Naively scaling 768P/124f (~11 min/step) by latent ×1.9 predicts ~21 min/step for 768P/243f; measured (incl. load) is ~34 min/step. ETA for folded long shots should not be extrapolated linearly — wait for step 1 and re-baseline. 3. **768P fold 6 steps runs fine on 24GB** — no OOM, no major swap (peak swap ~11.5 GB). Upscaling from 480P+super-res is still cheaper than native 768P/243f (~3.4 h) on this box, but native 768P is viable. Optional quality note: on this strong-motion clip, Laplacian variance was lower for fold (309.4) than 4-bit mlx-serve (604.0), yet human subjective judgement favored fold's natural sharpness — the mlx output reads as over-sharpened artificial halos. So Laplacian alone is not a quality proxy; "natural" vs "over-sharp" must be judged by eye.

Henninges

M4 Max 128 GB — folded Turbo LoRA (PR #14 tooling, larryvrh v4 adapter), Ref2VA image+audio: | Mode | Steps / reuse | Frames | Time (warm) | |------|---------------|--------|-------------| | Standard | 20 / 2 | 56 | 4:23 | | Turbo | 6 / 1 | 56 | 3:00 | End-to-end ~1.46× speedup — matches guyz' 1.42× on M5 Max. Fixed costs (text encode, VAE, FFmpeg) dominate at short clips, so the gain grows with clip length. Quality note: Turbo is excellent for iteration, but high-frequency detail (eyes, fingers) shows slightly more artifacts at 6 steps than Standard. For close-up faces / finals I stay on Standard. Confirms PR #14's "comparable quality" holds broadly, with the caveat that faces reveal the distillation.