Reusable ~14 KB sliders for MiniMax H3, built from pairs of prompts... - Vibeus
Reusable ~14 KB sliders for MiniMax H3, built from pairs of prompts rather than a training run. MatlowAI’s ComfyUI-Dials can steer movement, mood and acting style, combine several dials or layer them with a LoRA, and influence the generated audio.
Each pair describes the same subject with one quality changed—stillness versus energetic motion, for example. The node averages those differences into a steering vector and adds it to H3’s processed text conditioning. It adjusts concepts the model already understands; it does not teach a new face or an unfamiliar style.
The demonstration combines a detail LoRA with a motion dial to bring movement back, then explores darker moods and a two-pass Pika H3 Flex workflow that retains the first pass’s audio. The complete experiment video is attached.
Created for the Open Source Higgsfield challenge. The interface and calibration of LoRA exports are still being refined; MOSS/Krea support is planned. A project demo, source code and ready-made dials are linked from the original post. Code and workflows use AGPL-3.0. Dial weights remain under the MiniMax H3 Community License, including its territory and use restrictions.