Skip to content
AI Primer
workflow

PurzBeats shows MiniMax H3 video looping in Comfy Cloud

PurzBeats demonstrates looping MiniMax H3 videos in Comfy Cloud using _pxlpshr’s node pack. Follow-up instructions cover 48 fps interpolation and a VRAM reservation setting intended to prevent memory errors on an RTX 4090.

4 min read
PurzBeats shows MiniMax H3 video looping in Comfy Cloud
PurzBeats shows MiniMax H3 video looping in Comfy Cloud

TL;DR

The node pack samples a clip’s end alongside its beginning, with a second sampling view to clean up audio at the join. Its prompt guide treats loops as persistent states, while a ControlNet workflow can restyle an existing looping depth pass.

Loop generation and decode

PurzBeats introduced _pxlpshr’s pack in the Cloud demo thread. Four nodes divide the work, according to the README:

  • MiniMax H3 Loop: patches sampling so the latent’s end wraps around to its start.
  • MiniMax H3 Loop Decode: decodes with wrapped context and outputs exactly one loop of frames and audio.
  • MiniMax H3 Loop Control Video: fits an existing seamless clip to the generated loop for Fun ControlNet conditioning.
  • Loop Interpolate: generates intermediate frames and supplies the updated output frame rate.

The default views=2 also samples a half-rotated version of the loop, taking audio around the join from that second view. The author puts its sampling-time cost at approximately 1.8×.

48 fps across the seam

H3 generates at 24 fps. Increasing the frame rate in Create Video alone speeds up playback, according to the usage notes.

Loop Interpolate fills the last-to-first gap as well as the rest of the clip, the pack’s smartest detail. The first three supplied example workflows use FILM 2× interpolation for 48 fps output, matching the result PurzBeats described in his reply.

Continuous motion prompts

The author’s prompting guide specifies six conditions for a clean join:

  1. Locked camera: no pan or zoom.
  2. One clear subject: continuous or cyclic motion such as breathing, rotating or swaying.
  3. Steady surroundings: consistent lighting, environment and ambient sound.
  4. No accumulation: spills, crumbs and footprints would reset at the seam; any wet or scattered end state is present from the first frame.
  5. No timeline: no timestamps, “then,” “finally,” or isolated events.
  6. Subtle turbulence: sand, rain, smoke and spray remain secondary because their patterns can change at the join.

A separate H3 transformation from GlennHasABeard uses the prompt “the dust lifts off as a swirl of tiny orange moths.” He rendered the 15-second clip locally in ComfyUI, then used Topaz to upscale it to 4K.

A leaf dissolves into orange moths

DJ avatars and visual experiments

Onstage, PurzBeats demonstrated Sentinel with Comfy and RekordBox. For another performance, he explained how the DJ avatar and screen content were sequenced:

  • Comfy generated the screen content and DJ avatar.
  • Sentinel and Ableton Live sequenced the assets.
  • Ableton location markers supplied lighting and “song pressure” cues.

The Comfy-made DJ avatar performance

Other posts in the evidence include separate visual and performance experiments:

Local setup and RTX 4090 memory

Comfy’s Cloud support page describes curated, preinstalled custom packs and says users cannot directly install arbitrary nodes. The local installation documented in the linked repository requires:

  • ComfyUI v0.42.0 or newer.
  • The repository cloned into ComfyUI/custom_nodes, followed by a restart.
  • ComfyUI-Frame-Interpolation for Loop Interpolate.
  • ComfyUI-VideoHelperSuite for the Control example’s Load Video Path node.

Example workflows appear in ComfyUI’s template browser. For an RTX 4090, PurzBeats relayed the --reserve-vram 6 setting as a way to avoid memory errors.

The author’s RTX 4090 timing estimates, measured at 1280×704 with length=107, are:

  • Standard: 25 steps, approximately 21 minutes per loop.
  • Turbo: 8 steps, approximately 7 minutes.
  • FastH3: 8 steps, approximately 7 minutes.
  • Control: 8 steps, approximately 4 minutes.

Those measurements use views=2, except Control’s views=1. The first three workflows default to a larger 1344×768 render and a roughly 6.4-second loop, which takes longer; the author also reports possible fast-motion judder with Turbo and FastH3.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 6 threads
TL;DR2 posts
Loop generation and decode2 posts
48 fps across the seam1 post
Continuous motion prompts1 post
DJ avatars and visual experiments5 posts
Local setup and RTX 4090 memory2 posts
Share on X