MiniMax-H3 camera motion LoRA: "camera motion" trigger, strength 0.8-1.0 and reliable moves

This community LoRA, trained on newly collected footage, adds camera moves that MiniMax-H3 handles poorly. Files: camera_motion_h3_lora_v1_1000_pruned.safetensors and camera_motion_h3_lora_v1_3000_pruned.safetensors. Put the file in ComfyUI/models/loras/, load it with LoraLoader at strength_model 0.8 (gentle) to 1.0 (clear moves); above 1.2 the image may become unstable. Use a MiniMax-H3 video workflow with the pruned INT8 ConvRot base. Prompting: the trigger "camera motion" is required and must come first. Follow the training caption structure: camera motion, [subject], [scene], [camera move], [lighting/style], [quality]. The repository's camera_motion_提示词库.md holds 55 prompts copied from the training data across nine move types; the author suggests reusing them, swapping only subject and scene to keep the move, or swapping only the move phrase (for example slow push-in to slow orbit). Example: "camera motion, two men in a dynamic martial arts fight ..., slow pull-back combined with gentle orbit around the fighters revealing the full scale of the environment, cinematic gritty texture, natural volumetric lighting, high production quality". Control: add slow for gentler moves or dynamic handheld / aggressive push-in for stronger ones, and direction words such as push-in from left, orbit right or pan left (that table row is partly garbled in the card). Author ratings: push-in (slow push-in, dolly in), pull-back (dolly out) and handheld tracking are most stable; orbit, aerial (drone shot, flyover), crane or tilt, push-then-pull and close-up or macro are stable; pan is weak because of few samples, so combine it with other move words. Licence: the LoRA card declares Apache-2.0; the MiniMax-H3 base model is under the MiniMax H3 Community License. Based on the Jojocodex/minimax-h3-Camera-Motion-lora card pinned to revision fac88d6196ca84f3cee014ec79c71dc73572ab0a and observed on 2026-10-27; the Chinese text was read directly. Nothing was run here; the test videos and the 55-prompt library file were not inspected. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/Jojocodex/minimax-h3-Camera-Motion-lora/raw/fac88d6196ca84f3cee014ec79c71dc73572ab0a/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 367:484, 553:660, 804:946, 1088:1168, 6730:6907; exact excerpt SHA256 288c5f44cd9f639c71f20cad18a3c63d840d0a537dd826cdb9529998b3ee5cb0,6aecbdbf70bd86d1feb381225fa82d0d19f4eb3c66f1f772f651b57342510c3e,23baee39dcd8d3e6dcbd0041f4cb42c04198b806580093f78d4974b6320fa24a,6bf9daa7f628d"
    }
  ]
}