<h2>⚡ CROWN VFX &amp; High-Speed Combat Boost | MiniMax H3 自训特效与高速打斗加强 LoRA</h2>
[English] 
This is our self-trained high-rank (<code>Rank 64</code>, <code>Alpha 64</code>, <code>Video-Only</code> targeted blocks 10–49) VFX &amp; high-speed martial-arts/weapon combat enhancement LoRA for MiniMax H3 (Ref2VA / I2VA / T2VA). 
Because it is trained exclusively on video motion/VFX blocks (<code>video_only: True</code>), it dramatically boosts high-speed bare-fist &amp; weapon choreography, multi-layer fluid/plasma trails, lightning particles, Sakuga impact frames, and screen-filling ultimate detonations while 100% preserving your character's facial identity and outfit consistency in multi-reference (<code>Ref2VA</code>) workflows.

[中文介绍] 
这是我们自己针对 MiniMax H3 (Ref2VA / I2VA / T2VA) 专门训练的高秩（<code>Rank 64 / Alpha 64</code>，仅针对视频动态与特效层 <code>Blocks 10–49 Video-Only</code>）特效加强与高速武打动作 LoRA。 
由于采用纯视频层（<code>Video-Only</code>）解耦训练，它在大幅强化赤手空拳高速体术肉搏、刀剑兵刃交锋轨迹、多层流体/等离子气浪、雷电粒子、黑闪冲击帧以及满屏终极大招（如百米星河冰凰、50米3D冰魄星河巨莲超新星大爆炸）的同时，在多参考图（<code>Ref2VA</code>）模式下100% 锁死男女主面部颜值与服装细节，绝不污染五官。

<hr />

<h3>📊 Real Test Records &amp; Exact LoRA Stacks / 真实测试记录与完整参数配置</h3>
All showcase videos uploaded to this page were generated using the exact models, steps, and LoRA weights listed below: 
本页面上传的三支实测宣传视频，其真实使用的底模、采样步数、以及混搭的所有 LoRA 名称与权重完整公开如下：

🎬 Showcase 1: 35s Full Combat &amp; Grand Ultimate (Wasteland Mowing + 100m Starlight Ice-Phoenix + 50m 3D Starlight Lotus Supernova Explosion) / 实测视频 1：35秒全片（荒原割草 + 百米九天星河冰凰 + 50米3D星河巨莲超新星大爆炸）
<ul>
 <li>Base Model (底模 UNet): <code>minimax_h3_ref2va_bf16.safetensors</code> (Full 66GB BF16)</li>
 <li>Text Encoder (文本编码器): <code>qwen3vl_32b_minimax_h3_bf16.safetensors</code></li>
 <li>VAE: <code>minimax_h3_video_vae_fp32.safetensors</code> + <code>minimax_h3_audio_vae_fp32.safetensors</code></li>
 <li>Steps &amp; Sampler (步数与采样参数): <code>25 steps</code>, <code>sampler: euler</code>, <code>scheduler: beta</code>, <code>denoise: 1.0</code>, <code>MiniMaxH3SigmaShift (shift_video: 12.0, shift_audio: 3.0)</code></li>
 <li>Full LoRA Stack &amp; Exact Weights (使用的全部 LoRA 与精确权重):
 <ol>
 <li><code>CROWN_H3_VFX_Combat_Boost_V2_r64.safetensors</code> (This Model / 本模型 · 我们自己训练的特效加强 LoRA): <code>1.35</code></li>
 <li><code>minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors</code>: <code>0.45</code></li>
 <li><code>Bunny_weapon_combatV1.safetensors</code>: <code>0.65</code></li>
 <li><code>H3_Combat_V2.safetensors</code>: <code>0.65</code></li>
 <li><code>motion_speed_slider_mmh3_v1_1.safetensors</code>: <code>1.60</code></li>
 <li><code>MMH3-ExposureSlider-V1.safetensors</code>: <code>0.75</code></li>
 </ol>
 </li>
</ul>

🥊 Showcase 2: 15s Male &amp; Female Lead Bare-Fist High-Speed Duel (3D Camera &amp; Dual-Color Fist Plasma Clash) / 实测视频 2：第一段 · 男女主 15 秒赤手空拳高速徒手格斗对决（3D 运镜 + 双色拳风流体对撞）
<ul>
 <li>Base Model (底模 UNet): <code>minimax_h3_ref2va_bf16.safetensors</code> (Full 66GB BF16)</li>
 <li>Text Encoder (文本编码器): <code>qwen3vl_32b_minimax_h3_bf16.safetensors</code></li>
 <li>VAE: <code>minimax_h3_video_vae_fp32.safetensors</code> + <code>minimax_h3_audio_vae_fp32.safetensors</code></li>
 <li>Steps &amp; Sampler (步数与采样参数): <code>24 steps</code> (<code>ManualSigmas</code> + <code>euler</code>), <code>MiniMaxH3SigmaShift (shift_video: 12.0, shift_audio: 3.0)</code></li>
 <li>Full LoRA Stack &amp; Exact Weights (使用的全部 LoRA 与精确权重):
 <ol>
 <li><code>CROWN_H3_VFX_Combat_Boost_V2_r64.safetensors</code> (This Model / 本模型 · 我们自己训练的特效加强 LoRA): <code>1.25</code></li>
 <li><code>minimax_h3_singularity_v2_lora_avg_rank_351_bf16.safetensors</code>: <code>0.55</code></li>
 <li><code>pika_h3_flex_ref2va_comfy_bf16.safetensors</code>: <code>0.55</code></li>
 </ol>
 </li>
</ul>

🔥 Showcase 3: 15s Solo Bare-Fist Martial Arts &amp; Black-Flash Shockwave (Standalone Single-LoRA Test) / 实测视频 3：单人 15 秒赤手空拳高速体术连招与黑闪爆破（仅单挂本 LoRA 独立实测）
<ul>
 <li>Base Model (底模 UNet): <code>minimax_h3_ref2va_pruned_bf16safetensors.safetensors</code> (37.5GB Pruned BF16)</li>
 <li>Text Encoder (文本编码器): <code>qwen3vl_32b_h3_ultra_uncensored_heretic_int8_convrot.safetensors</code></li>
 <li>VAE: <code>minimax_h3_video_vae_int8_convrot.safetensors</code> + <code>minimax_h3_audio_vae_fp32.safetensors</code></li>
 <li>Steps &amp; Sampler (步数与采样参数): <code>20 steps</code>, <code>sampler: euler</code>, <code>scheduler: beta</code>, <code>MiniMaxH3SigmaShift (shift_video: 12.0, shift_audio: 3.0)</code></li>
 <li>Full LoRA Stack &amp; Exact Weights (使用的 LoRA 与精确权重):
 <ol>
 <li><code>CROWN_H3_VFX_Combat_Boost_V2_r64.safetensors</code> (This Model Only / 仅单挂我们自己训练的特效加强 LoRA，无任何其他 LoRA): <code>1.50</code></li>
 </ol>
 </li>
</ul>

<hr />

<h3>🛠️ Recommended Weight &amp; Trigger Word / 推荐权重与触发词</h3>
<ul>
 <li>Trigger Word (触发词): <code>sd25_vfx_combat</code></li>
 <li>Recommended Weight (推荐权重):
 <ul>
 <li>Stacked with other combat/speed LoRAs (与其他动作/加速 LoRA 混搭时): <code>1.10 ~ 1.35</code></li>
 <li>Standalone solo usage for extreme bare-fist / VFX (单独挂载测试极限体术与流体爆炸时): <code>1.25 ~ 1.50</code></li>
 </ul>
 </li>
</ul>

CROWN VFX & High-Speed Combat Boost | MiniMax H3

v2.0 Video-Only Rank 64 release. Preserves 100% of character facial identity in Ref2VA while boosting high-speed combat choreography and fluid/particle VFX.

Files with SHA256: 799a7dd52d4d23729c85c84e7fd0a7f1fd4fe0be667aa1dca18ebfc6c3fa31d7