CivArchive
    CROWN True First-Person Subjective POV & Interactive Gaze | MiniMax H3 - v2.0 - VideoOnly r64
    NSFW

    👁️ CROWN True First-Person Subjective POV & Interactive Gaze | MiniMax H3

    [English]
    Trained by NineCrowns ([CROWN]) on 100% real-world human GoPro/Ego4D head-mounted subjective POV and 175cm adult eye-level interactive date footage using an 8×NVIDIA B200 cluster (Rank 64 / Alpha 64, Video-Only Blocks 10–49).
    Designed for MiniMax H3 (Ref2VA / I2VA / T2VA) first-person vlogs, interactive dating sims, and subjective camera storytelling:

    1. Authentic Bottom-Border Ego-Centric Hand Geometry (crown_pov): Ensures the viewer's own first-person hands enter strictly upward from the bottom horizontal border (5 o'clock & 7 o'clock directly beneath the camera lens) when pouring drinks, holding glasses, or interacting on a tabletop—eliminating the common H3 flaw where a disembodied arm sticks in horizontally from the side wall or spawns a third hand on the character.
    2. Unbroken One-Take Eye-Level Tilt-Down & Tilt-Up Continuity: Supports smooth, continuous single-take subjective head tilts (starting at 175cm adult eye level gazing at the heroine across a table ➔ tilting down 25° to pour champagne while the heroine remains seated across the table watching ➔ tilting back up to eye level as she lifts the freshly poured glass to toast the lens and speak).

    [中文说明]
    由 NineCrowns ([CROWN]) 基于 100% 纯真人 GoPro/Ego4D 头戴第一人称主观视角与 175cm 成人平视面对面互动实拍素材,在 8×NVIDIA B200 集群上采用 Video-Only (Blocks 10–49, Rank 64 / Alpha 64) 架构精炼而成。专为 MiniMax H3 (Ref2VA / I2VA / T2VA) 第一视角 Vlog、沉浸式恋爱互动与主观运镜打造:

    1. 正下方底边(5/7点钟方向)第一人称主观伸手几何锁(crown_pov):确保第一视角自己的手严格从镜头正下方下边框往上伸出执行倒酒、抓握、递物等交互,彻底根除 H3 原生容易出现的“第三只手长在女主身上”或“手臂从右侧面墙壁水平横捅进来”的解剖学穿帮。
    2. 一镜到底成人平视俯仰连贯运镜:支持 15 秒无切镜自然低头与抬头(平视餐桌对面的女主 ➔ 镜头自然下移对准桌面倒满香槟、女主全程坐在桌对面不消失 ➔ 镜头自然上移回平视、女主伸右手端起刚倒满的酒杯向镜头敬酒说台词)。

    📊 Real Production Test Configuration / 真实测试参数配置
    • Base Model (底模): minimax_h3_ref2va_bf16.safetensors + qwen3vl_32b_minimax_h3_bf16.safetensors + minimax_h3_video_vae_fp32.safetensors
    • Sampler & Steps (采样器与步数): 25 steps, euler + beta, cfg: 1.0, MiniMaxH3SigmaShift (shift_video: 12.0, shift_audio: 3.0)
    • LoRA Weight (单挂本 LoRA 推荐权重): 0.80 ~ 0.85 (Showcase video generated with ONLY this LoRA at 0.85, zero third-party LoRAs)
    • Trigger Word (触发词): crown_pov

    Description

    [EN] v2.0 Video-Only Rank 64 (Blocks 10–49) release. Recommended weight: 0.80–0.85. Enables true bottom-up ego-centric hand POV and unbroken one-take eye-level tilt-down/tilt-up interactive camera movement.
    [中文] v2.0 纯视频层 Rank 64(Blocks 10–49)正式版,推荐权重 0.80–0.85。实现正下方底边伸手第一人称主观交互与一镜到底平视低头倒酒/抬头敬酒连贯运镜。

    FAQ

    LORA
    MiniMax H3

    Details

    Downloads
    524
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/9/2026
    Updated
    11/10/2026
    Deleted
    -
    Trigger Words:
    crown_pov

    Files

    CROWN_H3_True_POV_Interactive_V2_r64.safetensors