Home About Projects Blog Graph Resume Contact 中文
Back to list
2026年10月14日

Local Video Generation (FreeVideo)

Running MiniMax H3 on an 8GB laptop GPU: eleven minutes for a ten-second clip, 6.85GB peak VRAM — and the real bottleneck turned out to be system RAM.

#ComfyUI#MiniMax H3#Local Inference#VRAM

I wanted to see with my own eyes how far “generate video on my own laptop” actually is, so I tried the open-source project FreeVideo.

FreeVideo is a local inference engine for MiniMax H3, powered by Video DeltaNet (VDN), shipped as a ComfyUI plugin with a one-click launcher. It picks its own acceleration path based on the hardware, so there is nothing to configure by hand.

Setup

  • GPU: RTX 4060 Laptop, 8GB VRAM
  • Task: ref2va (reference image to video), with a photo of a person as the reference
  • Prompt: generate a video of the person in the reference dancing, any background

What came out

  • A 10.125 second clip (243 frames at 24fps), 960×544
  • 11 sampling steps: 8 steps at 480×288, a learned latent upscale to 960×544, then 3 independent refinement steps
  • About 11 minutes
  • 6.70GB peak VRAM (1.2GB baseline), inside the 8GB budget the whole way

What it does to fit in 8GB

This was the part I actually wanted to see. The VRAM and GPU samples it produces show the key moves:

  • int8 linear layers plus a sage2 attention kernel
  • Streamed weights: weights are not all resident in VRAM, they are fed in as computation proceeds, with about 14.4GB of pinned host memory as a buffer
  • AdaLN caching, disk cache included
  • Tiled decoding: 27 resident blocks alongside streamed output

The cost shows up elsewhere: 93% mean GPU utilisation, 84°C peak temperature and 138W peak power. And system memory dropped to 2.2GB available at its lowest — on this machine the tight resource is not VRAM, it is RAM. To push resolution or duration further, more memory is a more practical upgrade than a new GPU.

What it left behind

Every generation leaves a full workflow.json, per-stage VRAM and GPU samples, and parameter logs alongside the video. The numbers above are read out of video.engine.memory.json and video.engine.gpu.json rather than recalled from memory.

One unexpected bonus: dragging a FreeVideo output back onto the ComfyUI canvas restores the prompt, the seed and every parameter. The workflow that produced this is reproducible, not a one-off roll of the dice.