Dataset Viewer
Auto-converted to Parquet Duplicate
run
stringclasses
2 values
stage
stringclasses
3 values
index
float64
0
20
⌀
ms
float64
32.4
2.4k
clip
text_encoder
null
871.4
clip
latent_frame
0
2,400.6
clip
latent_frame
1
935.8
clip
latent_frame
2
859.4
clip
latent_frame
3
909.2
clip
latent_frame
4
940.5
clip
latent_frame
5
980.8
clip
latent_frame
6
1,026.3
clip
latent_frame
7
1,097.8
clip
latent_frame
8
1,121.6
clip
latent_frame
9
1,202.3
clip
latent_frame
10
1,253.2
clip
latent_frame
11
1,325
clip
latent_frame
12
1,313.4
clip
latent_frame
13
1,391.9
clip
latent_frame
14
1,472.8
clip
latent_frame
15
1,543.4
clip
latent_frame
16
1,571.5
clip
latent_frame
17
1,628.5
clip
latent_frame
18
1,688.4
clip
latent_frame
19
1,787.5
clip
latent_frame
20
1,841.4
clip
decoder
null
1,115.8
smoke
latent_frame
0
1,036.8
smoke
latent_frame
1
32.4
smoke
decoder
null
429.2

Apple FastFrames-S on a 64GB M5 Pro Mac mini (MLX)

Headline: I got Apple's FastFrames-S open weights generating video on my 64GB M5 Pro Mac mini. The official code is CUDA-only, so I ran it through my MLX port, which is not yet public. One 832×480, 81-frame clip (5.06 s at 16 fps) took 30.43 s to generate, 375.7 ms per output frame, peaking at 22.93 GB of MLX memory.

This is an independent local measurement from November 18, 2026. I did not upload any weights.

A background inference app was paused for this run at the owner's request.

Prompt: A corgi surfing a wave at sunset (Apple's official example prompt), seed 0.

Seconds per latent frame on a 64GB M5 Pro Mac mini

Setup

  • Apple M5 Pro Mac mini, 64 GB unified memory, macOS 27.0.1 (26A434), AC power
  • Python 3.12.15, mlx 0.32.4, ffmpeg 9.0.2
  • Official FastFrames-S weights. No --bits quantization.

Smoke test

A tiny 8×8, 5-frame smoke test passed in 2.83 s (2.98 s wall).

The first smoke test on this machine failed before any frames were generated, with ValueError: float64 is not supported on the GPU. Metal does not support float64. I fixed that in the port and reran. The numbers below are from the successful rerun. See logs/first_failed_smoke.txt.

Corgi clip

Resolution 832×480
Frames 81 at 16 fps (5.06 s)
Seed / prompt 0, "A corgi surfing a wave at sunset"
Container H.264, 693 KB
Text encoder 871 ms
Latent frames 21 (first 2,401 ms, then 859 to 1,841 ms, rising)
Decoder 1,116 ms
Generation 30.43 s
Wall clock 35.9 s
Per output frame 375.7 ms (~2.7 fps, about 6× off real time)
Memory 22.93 GB MLX peak, 30.37 GB RSS peak

Each bar in the chart is one latent frame, about 4 video frames. The first latent frame is a 2.4 s warm-up, then they climb from about 0.9 s to 1.8 s as the video gets longer.

Caveats

  • Single run, single seed.
  • The MLX port was verified by 13 CPU parity tests vs the reference, but there is no side-by-side output comparison against the official CUDA reference yet.
  • --bits quantization was not used.
  • The MLX port is not yet public, so this dataset has the run data and notes, not the port code.

Files

  • clip.mp4, smoke.mp4
  • charts/fastframes_mac_chart.png
  • timings.csv, runs.csv, summary.json, ff_result.json
  • clip.ffprobe.json, smoke.ffprobe.json
  • logs/clip.log, logs/smoke.log, logs/first_failed_smoke.txt
  • environment.txt, environment.json
  • scripts/chart.py

Licenses and credit

Measurements and this card: CC BY 4.0, attribution Pratik Gandhi. The FastFrames weights and official code are under the Apple Sample Code License (apple-amlr). No model weights are redistributed.

Credit to Apple and the FastFrames authors.

Write-up on GitHub

How I ran this, what broke, and the chart: Apple FastFrames-S (MLX). It's part of Local AI field notes, where I log every model I run on my Mac mini.

Write-up on GitHub

How I ran this, what broke, and the scripts: Apple FastFrames-S (MLX). It's part of Local AI field notes, where I log every model I run on my Mac mini.

Downloads last month
139