H3 Fun ControlNet Brings Control-Video Guidance for MiniMax-H3 to ComfyUI

wyzborrero · reddit · 2026-08-27

A developer open-sourced ComfyUI-H3-FunControl, the first way to drive MiniMax-H3 video generation with a control video in ComfyUI: depth, canny, pose, HED and MLSD all work (the underlying model is a union adapter). Control sources can be 3D renders, estimators over reference footage, hand-drawn frames or ComfyUI preprocessors; generation follows the structure frame by frame.

Why it didn't exist: Alibaba PAI's MiniMax-H3-Fun-Controlnet-Union ships full-width AdaLN, while the pruned H3 checkpoints most people actually run are curve-form — incompatible. Using the official ControlNet meant the 34GB unpruned base and adherence still failed. Kijai re-derived the ControlNet into the curve-form basis, after which a control block is literally comfy's DiTBlock and the AdaLN is an unmodified AdalnProj; this node is the glue.

Practical notes (the README is largely a list of failure modes): chained control towers sum — depth 1.0 + pose 1.0 behaves like a single control at 2.0 and saturates, treat strength as a budget; pose held the subject in the author's test while depth let camera distance drift; endpercent matters more than it looks — held at 1.0 the control asserts itself during texture formation and suppresses prompted texture; with Sol-Attn, morton must be off or control silently fails, producing a clean video that ignored its input. Apache-2.0, three annotated example workflows, tested only on an RTX 5090.

Original post →

More from Multimodal

Multimodal channel →