A MiniMax H3 LoRA for luminous satin pantyhose and hosiery appearance in continuous video and image scenes. The supplied tests cover live-action walking, office, mall, cafe close-up, jogging, and an anime station scene.
Trigger evidence: use "tutujiguang pantyhose"; the training captions also directly use "tutujiguang". Start with LoRA weight 1.0. Requires the MiniMax H3 base model. This checkpoint is a diffusion-transformer LoRA only, with no text-encoder LoRA; BF16, rank 16, alpha 16, training checkpoint 3300 steps.
图图极光 MiniMax H3 LoRA，用于在连续视频和图像场景中表现明亮细腻的缎面光泽连裤袜/丝袜材质。已用实测样例覆盖步行街、办公室、商场中庭、咖啡馆腿部近景、公园慢跑和动漫车站广场。
触发词证据：建议使用 "tutujiguang pantyhose"；训练字幕中也直接使用了 "tutujiguang"。建议从 LoRA 权重 1.0 开始。需要 MiniMax H3 底模。本 checkpoint 仅包含 diffusion transformer LoRA，不含文本编码器 LoRA；BF16，rank 16，alpha 16，3300 步。
Training/evaluation note: the checkpoint was saved and selected by training steps; an auditable effective epoch count was not recorded.
训练/评测说明：本模型按训练步数保存与筛选，未记录可核验的有效 Epoch。
Description
Base model: Minimax H3 Open (SeaArt selectable base corresponding to the supplied MiniMax H3 training setup).
Trigger evidence: "tutujiguang" and "tutujiguang pantyhose". Start with Strength 1.0.
Transformer-only LoRA; no text-encoder weights. BF16, rank 16, alpha 16, checkpoint at 3300 steps.
基础模型：Minimax H3 Open（海艺当前可选的视频基础模型）。
触发词证据："tutujiguang" 与 "tutujiguang pantyhose"；建议 Strength 从 1.0 开始。
仅含 diffusion transformer LoRA，不含文本编码器权重；BF16，rank 16，alpha 16，3300 步。