⚡ CROWN VFX & High-Speed Combat Boost | MiniMax H3 自训特效与高速打斗加强 LoRA
[English]
This is our self-trained high-rank (Rank 64, Alpha 64, Video-Only targeted blocks 10–49) VFX & high-speed martial-arts/weapon combat enhancement LoRA for MiniMax H3 (Ref2VA / I2VA / T2VA).
Because it is trained exclusively on video motion/VFX blocks (video_only: True), it dramatically boosts high-speed bare-fist & 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 (Ref2VA) 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(Rank 64 / Alpha 64,仅针对视频动态与特效层 Blocks 10–49 Video-Only)特效加强与高速武打动作 LoRA。
由于采用纯视频层(Video-Only)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(Ref2VA)模式下100% 锁死男女主面部颜值与服装细节,绝不污染五官。
📊 Real Test Records & Exact LoRA Stacks / 真实测试记录与完整参数配置
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:
- Base Model (底模 UNet):
minimax_h3_ref2va_bf16.safetensors(Full 66GB BF16) - Text Encoder (文本编码器):
qwen3vl_32b_minimax_h3_bf16.safetensors - VAE:
minimax_h3_video_vae_fp32.safetensors+minimax_h3_audio_vae_fp32.safetensors - Steps & Sampler (步数与采样参数):
25 steps,sampler: euler,scheduler: beta,denoise: 1.0,MiniMaxH3SigmaShift (shift_video: 12.0, shift_audio: 3.0) - Full LoRA Stack & Exact Weights (使用的全部 LoRA 与精确权重):
CROWN_H3_VFX_Combat_Boost_V2_r64.safetensors(This Model / 本模型 · 我们自己训练的特效加强 LoRA):1.35minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors:0.45Bunny_weapon_combatV1.safetensors:0.65H3_Combat_V2.safetensors:0.65motion_speed_slider_mmh3_v1_1.safetensors:1.60MMH3-ExposureSlider-V1.safetensors:0.75
- Base Model (底模 UNet):
minimax_h3_ref2va_bf16.safetensors(Full 66GB BF16) - Text Encoder (文本编码器):
qwen3vl_32b_minimax_h3_bf16.safetensors - VAE:
minimax_h3_video_vae_fp32.safetensors+minimax_h3_audio_vae_fp32.safetensors - Steps & Sampler (步数与采样参数):
24 steps(ManualSigmas+euler),MiniMaxH3SigmaShift (shift_video: 12.0, shift_audio: 3.0) - Full LoRA Stack & Exact Weights (使用的全部 LoRA 与精确权重):
CROWN_H3_VFX_Combat_Boost_V2_r64.safetensors(This Model / 本模型 · 我们自己训练的特效加强 LoRA):1.25minimax_h3_singularity_v2_lora_avg_rank_351_bf16.safetensors:0.55pika_h3_flex_ref2va_comfy_bf16.safetensors:0.55
- Base Model (底模 UNet):
minimax_h3_ref2va_pruned_bf16safetensors.safetensors(37.5GB Pruned BF16) - Text Encoder (文本编码器):
qwen3vl_32b_h3_ultra_uncensored_heretic_int8_convrot.safetensors - VAE:
minimax_h3_video_vae_int8_convrot.safetensors+minimax_h3_audio_vae_fp32.safetensors - Steps & Sampler (步数与采样参数):
20 steps,sampler: euler,scheduler: beta,MiniMaxH3SigmaShift (shift_video: 12.0, shift_audio: 3.0) - Full LoRA Stack & Exact Weights (使用的 LoRA 与精确权重):
CROWN_H3_VFX_Combat_Boost_V2_r64.safetensors(This Model Only / 仅单挂我们自己训练的特效加强 LoRA,无任何其他 LoRA):1.50
🛠️ Recommended Weight & Trigger Word / 推荐权重与触发词
- Trigger Word (触发词):
sd25_vfx_combat - Recommended Weight (推荐权重):
- Stacked with other combat/speed LoRAs (与其他动作/加速 LoRA 混搭时):
1.10 ~ 1.35 - Standalone solo usage for extreme bare-fist / VFX (单独挂载测试极限体术与流体爆炸时):
1.25 ~ 1.50
- Stacked with other combat/speed LoRAs (与其他动作/加速 LoRA 混搭时):
Description
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.