<h3>👁️ CROWN True First-Person Subjective POV &amp; Interactive Gaze | MiniMax H3</h3>
[English] 
Trained by NineCrowns (<code>[CROWN]</code>) on 100% real-world human GoPro/Ego4D head-mounted subjective POV and 175cm adult eye-level interactive date footage using an <code>8×NVIDIA B200</code> cluster (<code>Rank 64 / Alpha 64, Video-Only Blocks 10–49</code>). 
Designed for MiniMax H3 (Ref2VA / I2VA / T2VA) first-person vlogs, interactive dating sims, and subjective camera storytelling:
<ol>
 <li>Authentic Bottom-Border Ego-Centric Hand Geometry (<code>crown_pov</code>): Ensures the viewer's own first-person hands enter strictly upward from the bottom horizontal border (5 o'clock &amp; 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.</li>
 <li>Unbroken One-Take Eye-Level Tilt-Down &amp; 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).</li>
</ol>
[中文说明] 
由 NineCrowns (<code>[CROWN]</code>) 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 <code>8×NVIDIA B200</code> 集群上采用 <code>Video-Only (Blocks 10–49, Rank 64 / Alpha 64)</code> 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：
<ol>
 <li>正下方底边（5/7点钟方向）第一人称主观伸手几何锁（<code>crown_pov</code>）：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。</li>
 <li>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）。</li>
</ol>
<hr />
📊 Real Production Test Configuration / 真实测试参数配置
<ul>
 <li>Base Model (底模): <code>minimax_h3_ref2va_bf16.safetensors</code> + <code>qwen3vl_32b_minimax_h3_bf16.safetensors</code> + <code>minimax_h3_video_vae_fp32.safetensors</code></li>
 <li>Sampler &amp; Steps (采样器与步数): <code>25 steps</code>, <code>euler</code> + <code>beta</code>, <code>cfg: 1.0</code>, <code>MiniMaxH3SigmaShift (shift_video: 12.0, shift_audio: 3.0)</code></li>
 <li>LoRA Weight (单挂本 LoRA 推荐权重): <code>0.80 ~ 0.85</code> (Showcase video generated with ONLY this LoRA at <code>0.85</code>, zero third-party LoRAs)</li>
 <li>Trigger Word (触发词): <code>crown_pov</code></li>
</ul>

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

[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。

Files with SHA256: 5ae424a4a418ea13d57ef5cfda6357675da1c91f59055ba827fddf1c0cd2dfc4