MiniMax H3 Person Remover LoRA V1 Released: Open Source Clean Plate Generation for Video
AI creator akatz_ai released the Person Remover V1 LoRA for MiniMax H3, enabling clean plate extraction and background inpainting for character swap workflows.
On October 8, 2026, AI video creator akatz (@akatz_ai) officially released the open-source Hugging Face repository for 'MiniMax H3 Person Remover V1 LoRA', a specialized fine-tuned adapter designed for the MiniMax H3 video generation foundation model.
Image source: akatz (@akatz_ai) / Hugging Face
The newly released LoRA focuses on masking and removing human subjects across video frames, cleanly inpainting and restoring the occluded background (clean plate). It can be directly integrated into video compositing and generation workflows to prepare backgrounds for character swapping or dataset curation.
Dedicated Person Removal LoRA and Clean Plate Extraction Pipeline
In character swap and video generation pipelines, extracting an artifact-free background video without the original subject is a crucial prerequisite.
MiniMax H3 Person Remover V1 is explicitly tuned for pure person and subject removal rather than direct character replacement.
- Automated Clean Plate Synthesis: The model erases moving human subjects across video frames while reconstructing consistent spatial structures and background textures.
- Data Pipeline for Character Swap V2: Creator akatz noted that this LoRA serves as a core tool in generating high-quality training samples and clean background datasets for the upcoming 'Character Swap V2' model.
For workflows aiming to swap a character with another subject, this removal LoRA is intended to be paired with a dedicated Swap LoRA rather than performing direct replacements standalone.
Mitigating Identity Drift in Videos Longer than 5 Seconds
When processing videos exceeding 5 seconds with MiniMax H3, the baseline model can experience visual identity drift across frames.
To address this challenge, akatz outlined a context window workflow trick:
- Overlapping 5s Context Windows: Dividing long videos into 5-second context segments and blending overlapping boundary frames helps preserve temporal and background consistency across extended sequences.
- Workflow Roadmap: An updated workflow designed to streamline character swaps and clean plate processing for videos longer than 5 seconds is scheduled for upcoming release.
Sources
- Hugging Face Repository: akatz-ai/MiniMax-H3-Person-Remover-LoRA
- akatz Official X (@akatz_ai): MiniMax H3 Person Remover V1 Release Announcement