Check my GPU →

A 5-second video on a 16 GB card: Wan 2.2 5B, 14B and HunyuanVideo 1.5 timed

Three local video models, one 5-second clip each, on an RTX 5060 Ti 16 GB with 32 GB of RAM: time, peak VRAM and RAM, and why the small Wan was the fastest.

TimedBy Tomasz · aboutUpdated 2026-11-08

A 5-second clip is the usual test for local video. I made one with three models on my PC: RTX 5060 Ti 16 GB, 32 GB of RAM, ComfyUI portable. Each was a text-to-video run with the settings from the official template, timed after a warm-up run.

The results

Model and fileSizeFramesTimePeak VRAMPeak RAM
Wan 2.2 5B, 16-bit1280×7041218.6 min11.0 GB-
HunyuanVideo 1.5 480p, FP8848×48012111.0 min14.3 GB25.2 GB
Wan 2.2 14B, FP8832×4808119.1 min13.3 GB28.6 GB
Wan 2.2 14B, GGUF Q8_0832×4808125.3 min13.1 GB32.9 GB
Wan 2.2 14B, GGUF Q5_K_M832×4808125.7 min14.0 GB33.1 GB

20 steps for every run. Time = sampling plus VAE decode. Wan 2.2 makes 24 fps clips with 121 frames for the 5B and 16 fps with 81 frames for the 14B, so both are about 5 seconds. I did not log RAM for the first Wan 5B run.

The clips

Every clip came from the same prompt: a red fox walking through fresh snow in a pine forest, slow camera pan, soft morning light, cinematic. They are the files ComfyUI saved, only made smaller for the page. They load when you press play.

Wan 2.2 14B, FP8 832×480 · 81 frames · 19.1 min
Wan 2.2 14B, GGUF Q5_K_M 832×480 · 81 frames · 25.7 min
Wan 2.2 5B, 16-bit 1280×704, shown smaller · 121 frames · 8.6 min
A frame of the HunyuanVideo 1.5 test clip: a red fox in snow
HunyuanVideo 1.5 480p, FP8 one frame · 848×480 · 121 frames · 11.0 min

Each model read soft morning light, cinematic its own way. The 14B put a low sun behind the trees. The 5B went for orange trunks and a small fox far away. HunyuanVideo gave me a close-up in plain daylight.

The two Wan 2.2 14B clips, FP8 and GGUF Q5_K_M, came out as almost the same scene. At 832×480 I cannot see anything the GGUF run bought me for its extra 6.6 minutes.

What surprised me

The small model was the fastest, at the highest resolution. Wan 2.2 5B made a 1280×704 clip in under 9 minutes. The 14B needed 19 minutes for a clip with less than half the pixels. Counted per pixel and frame, the 5B got through about 7.5 times more per second.

On 32 GB of RAM, GGUF was slower than FP8. Wan 2.2 14B uses two 14B models (high noise and low noise) and keeps both in RAM. With the GGUF files RAM went over 32 GB, Windows started using the page file, and the clip took 25 minutes instead of 19. With 64 GB of RAM this would probably look different. I can only say what happens on 32.

HunyuanVideo 1.5 sits in the middle. 11 minutes for 121 frames at 480p, with the CFG-distilled file. I used VAE Decode (Tiled) from the start, so the decode at the end did not run out of memory.

Which one I would use on a 16 GB card

  • To try ideas: Wan 2.2 5B. It is the only one where waiting for a clip feels reasonable, and it fits with room to spare.
  • For the best motion I can get: Wan 2.2 14B with the FP8 files, if you can leave the PC working for 20 minutes per clip. Use FP8, not GGUF, unless you have more than 32 GB of RAM.
  • HunyuanVideo 1.5 is worth a try if Wan does not give you the look you want. It is not faster.

I am not judging which clip looks best here, only time and memory. The workflows for all three are free: ComfyUI workflows. Why video needs so much more memory than images: video VRAM explained. FP8 against GGUF for the 14B in more detail: Wan 2.2 14B on 16 GB.