The Adapter/LoRA is experimental and may generate artifacts. The attached image in the corners of the video frame is the input provided for video generation. The results are generated using 4–6 steps with MiniMax-H3 [Fast] + MiniMax-H3-I2V-Anime-Motion-LoRA. Below are example prompts used to generate the videos with respect to the image input.

prompts:
  - "Anime-Motion: Back view of a short brown-haired anime boy carrying a guitar case on his back, standing on a quiet hilltop road beside a railroad crossing sign, gazing at a vast pastel sky glowing with sunrise clouds; flower petals drifting and swirling in the wind, clouds flowing slowly, hair and jacket swaying gently in the breeze, soft shimmering light, nostalgic wistful atmosphere, seamless loop."

  - "Anime-Motion: A black-haired anime swordswoman with a dark scarf and scout-regiment jacket strikes a dynamic mid-air combat pose with dual blades, steel cables arcing around her against a swirling crimson energy aura on a pitch-black background; hair and scarf fluttering in the wind, red flames pulsing and spiraling, wires tensing and swaying, blade edges glinting, embers drifting, intense dramatic action atmosphere, seamless loop."

MiniMax-H3-I2V-Anime-Motion-LoRA

MiniMax-H3-I2V-Anime-Motion-LoRA is an experimental anime motion LoRA adapter trained on the general-purpose multimodal model MiniMax H3. This LoRA is trained on a curated collection of stop-motion-based video clips, each up to 15 seconds long, designed for seamless looping motion videos with subtle emotional expressions and minimalistic movements in characters and backgrounds.

All checkpoints are available here: https://huggingface.co/buckets/prithivMLmods/minimax-h3-i2v-anime-motion-lora. It is recommended to use checkpoint 1000/1400.

Model Details

• Base Model: MiniMax H3
• Adapter Name: minimax-h3-i2v-anime-motion-lora
• Created by: prithivMLmods
• Model Type: LoRA / Adapter

Training Specifications

Parameter	Configuration
Resolution	Default 480x832 (dynamically adjusted based on packed video frames)
Save Precision	BF16
Learning Rate	1e-4
Optimizer	AdamW
Network Dimension (Rank)	16
Total Steps	1400
Trigger Words	Anime-Motion

Training Dataset

The LoRA was trained on a curated collection of stop-motion-based video clips, with clips of up to 15 seconds in length. The training collection focuses on:

• Seamless loop motion videos
• Subtle emotional expressions
• Minimalistic character movements
• Subtle background movements
• Natural and consistent motion patterns

License

This adapter follows the MiniMax H3 Community License in accordance with the base model terms.