Creators test open-weight MiniMax H3 locally in ComfyUI

• MiniMax H3's local path can produce complete 15-second clips in ComfyUI: GlennHasABeard's fog test says the clip was generated locally and taken to 4K with Topaz.
• The strongest prompts specify a single path through time, from one scoop toward camera to a one-way paint stroke, as GlennHasABeard's moon test demonstrates.
• Visible mistakes remain part of the workflow: the pencil became stairs, while exact stripe counts stayed unreliable in the pencil clip and a follow-up.
• The local setup is a batch process, with one reported machine taking about 50 minutes per clip, according to GlennHasABeard's hardware reply.

MiniMax's open-source announcement describes H3 as a 4 to 15 second, 24 fps video model with native stereo audio, while the ComfyUI guide puts open-weight output at a 768-pixel short edge. The prompt guide includes a curious constraint for creators: H3's BasicGuider has no negative branch, so unwanted elements must be handled through positive descriptions. GlennHasABeard's clips turn that constraint into a series of practical tests involving fog, moon scoops, wax seals, daisies, ladybugs, and stair-shaped pencil shavings.

MiniMax's release gives H3 a 4 to 15 second range, 24 fps output, and a 768-pixel short edge for the local base model. Its hosted regeneration path reaches 2K, while the official ComfyUI model guide lists native T2V, I2V, and R2V workflows for local inference.

The creator's recurring recipe is simpler than the full model spec: locked camera, 15-second clip, local ComfyUI render, then Topaz for 4K output.

The clips use nearly static cameras and make the object carry the motion. Prompts describe a path, a hold, or a sequence of one-way movements instead of asking for a general effect.

• Fog rolls out of a pillow as a slow, thick river in GlennHasABeard's fog clip.
• A scoop pulls one cratered moon toward the camera, a detail GlennHasABeard's prompt says worked better as one directed stroke.
• A paintbrush travels only upward while revealing a meadow, according to GlennHasABeard's wall test.
• A pressure washer makes four left-to-right passes to reveal night sky, while GlennHasABeard's reply says the first two stripes still broke the requested count.
• An amber pour travels one way across a tray, letting the frost-like pattern grow along a defined path in GlennHasABeard's breakfast test.
• A tower of blocks avoids falling when each block slides one short step to the right, a sequence that took twelve seconds in GlennHasABeard's tower test.

Small practical objects make temporal drift easy to spot. The successful clips preserve the setup long enough for a familiar action to become a visual gag, while the misses expose where the model substitutes its own geometry.

• The pencil prompt asks for a spiral shaving with a thin red edge, but the resulting clip produces a straight staircase from the blade slot.
• A wax seal becomes a pond with an orange koi, and GlennHasABeard's seal test attributes the convincing lift to a two-second press and hold.
• A hole punch turns three paper circles into ladybugs that walk across the desk in GlennHasABeard's hole-punch clip.
• Five key strikes produce five daisies instead of letters in GlennHasABeard's typewriter clip.

ComfyUI's prompt guide says H3 uses a single conditioning path and recommends describing the desired positive content rather than relying on negative bans. Glenn's typewriter prompt still included “No letters or ink marks appear,” yet the clip produced the requested flowers.

Other supplied examples use H3 as one layer in a broader production chain. A techhalla reply names Midjourney and H3 Max Turbo for visuals, with Opus used for editing techhalla's workflow reply.

MayorKingAI says a flan drama came from a single text-to-video prompt MayorKingAI's reply, while the original post praises Hailuo H3's anime expressiveness and micro-reactions MayorKingAI's flan clip.

Venturetwins used Codex to orchestrate a stack in which Astra wrote lyrics, Fal made the media, Suno supplied the song, and MiniMax H3 generated the video, as venturetwins' workflow describes.

The local workflow carries a substantial render cost. The hardware screenshot attached to GlennHasABeard's reply lists an Intel i7-8700K, 32 GB of RAM, and an RTX 3060 with 12 GB of VRAM GlennHasABeard's hardware screenshot.

He reported roughly 50 minutes per clip on that setup GlennHasABeard's timing reply, then said he loads test prompts and lets them run overnight or while he is at work GlennHasABeard's batch-generation reply. ComfyUI's documentation also warns that downloading an H3 template can require tens of gigabytes, depending on the workflow.