English Version：<h1 id="minimax-h3-soft-and-natural-motion-lora">MiniMax H3 — Soft &amp; Natural Motion LoRA</h1>Softness in Every Motion, Natural Rhythm in Every MoveA motion-focused LoRA for MiniMax H3, designed to improve the softness, elasticity, and natural physical movement of the female chest during video generation.The goal is to achieve a more realistic and natural visual appearance, with softer deformation, more convincing elasticity, and natural secondary motion.<h3 id="features">Features</h3>1. Softness &amp; ElasticityProduces a noticeably softer appearance with natural deformation and recovery, aiming for a visual result that closely resembles real soft-tissue behavior.2. Natural Physical MotionImproves the natural physical movement during walking, turning, body movement, and other dynamic actions, creating more convincing inertia and follow-through.3. Reduced Oily / Artificial AppearanceHelps reduce the overly smooth, glossy, or oily appearance that can sometimes occur in AI-generated video, resulting in a more natural overall look.<h3 id="recommended-weight">Recommended Weight</h3>0.9No special trigger word is required.Simply describe the character and actions normally. You can also describe the natural movement of the chest in the prompt to further enhance the effect.The prompts used in the example videos are included for reference.<h3 id="test-environment">Test Environment</h3>1. <code>minimax_h3_f2va</code> + <code>minimax_h3_f12v_turbo_8step_v1.0_comfyui_bf16</code>2. <code>minimax_h3_hybrid_fi2va_ref2va_b25-49</code> + <code>minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16</code>Supports Text-to-Video, Image-to-Video, and Reference-to-Video.All example videos in this release were generated using Reference-to-Video.The <code>minimax_h3_ref2va</code> base model is not recommended, as testing showed that its overall results were inferior to the combinations listed above.中文版本：<h1 id="minimax-h3-lora">MiniMax H3 — 柔软且自然的乳房 LoRA</h1>绵柔入境，灵动有韵这是一款针对 MiniMax H3 的动态表现增强 LoRA，主要用于改善女性胸部在视频生成过程中的柔软性、弹性以及自然的物理运动表现。目标是让动态画面中的形变、回弹以及自然产生的二次运动更加真实，使整体视觉效果更加接近真实的柔软组织运动。<h3>特点</h3>1. 柔软性与弹性带来更加明显的柔软视觉效果，使形变和恢复更加自然，整体表现更加接近真实柔软组织的运动状态。2. 自然的物理晃动改善人物在走路、转身、身体移动以及其他动态动作中的自然运动表现，使运动更加符合身体惯性，并产生更加自然的跟随与回弹效果。3. 较好的去油腻效果降低 AI 视频中有时会出现的过度光滑、油亮以及塑料感，使整体视觉效果更加自然。<h3>推荐权重</h3>0.9无需特殊触发词。正常描述人物和动作即可。也可以在提示词中描述胸部的自然晃动，以进一步增强效果。示例视频中使用的相关提示词均可作为参考。<h3>测试环境</h3>1. <code>minimax_h3_f2va</code> + <code>minimax_h3_f12v_turbo_8step_v1.0_comfyui_bf16</code>2. <code>minimax_h3_hybrid_fi2va_ref2va_b25-49</code> + <code>minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16</code>支持 文生视频、图生视频、参考生视频。本次发布的所有示例视频均使用**参考生视频（Reference-to-Video）**生成。不建议使用 <code>minimax_h3_ref2va</code> 作为底模，实际测试中整体效果不如上述组合。

SoftBreastMotion

SoftBreastMotion.safetensors

v1.0


# 

## Model Introduction

This model is trained using the [Civision/model training](https://modelscope.ai/civision/modelTraining) environment and computing resources provided by ModelScope.

* Model Type: LoRA
* Base Model: [MiniMax/MiniMax-H3](https://modelscope.ai/models/MiniMax/MiniMax-H3)
* Training Code: [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio)
* Data amount：14
* Total training steps：1900
* License: Apache-2.0

## Inference Code

Install [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio):

```bash
pip install diffsynth
```

Inference:

```python
import torch
from diffsynth.pipelines.minimax_h3_audio_video import MiniMaxH3Pipeline, ModelConfig
from diffsynth.utils.data.audio_video import write_video_audio

vram_config = {
    "offload_dtype": "disk",
    "offload_device": "disk",
    "onload_dtype": "disk",
    "onload_device": "disk",
    "preparing_dtype": torch.bfloat16,
    "preparing_device": "cuda",
    "computation_dtype": torch.bfloat16,
    "computation_device": "cuda",
}
pipe = MiniMaxH3Pipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id="MiniMax/MiniMax-H3", origin_file_pattern="FL2VA/text_encoder/model*.safetensors", **vram_config),
        ModelConfig(model_id="MiniMax/MiniMax-H3", origin_file_pattern="FL2VA/transformer/model*.safetensors", **vram_config),
        ModelConfig(model_id="MiniMax/MiniMax-H3", origin_file_pattern="FL2VA/video_vae/source/model.safetensors", **vram_config),
        ModelConfig(model_id="MiniMax/MiniMax-H3", origin_file_pattern="FL2VA/audio_vae/model.safetensors", **vram_config),
    ],
    processor_config=ModelConfig(model_id="MiniMax/MiniMax-H3", origin_file_pattern="FL2VA/processor/"),
    vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 2,
)
pipe.load_lora(pipe.dit, ModelConfig(model_id="mugensiki/526123541", origin_file_pattern="526123541_c1-st1900.safetensors"))

prompt = "A cat"
video, audio = pipe(
    prompt=prompt,
    height=480, width=832, num_frames=124, num_inference_steps=50, seed=0,
)
write_video_audio(
    video=video, audio=audio,
    output_path="t2va.mp4", fps=24, audio_sample_rate=32000,
)
```


526123541

526123541_c1-st1600.safetensors

c1-st1600

c1-st100

c1-st200

c1-st300

c1-st400

c1-st500

c1-st600

c1-st700

c1-st800

c1-st900

c1-st1000

c1-st1100

c1-st1200

c1-st1300

c1-st1400

c1-st1500

c1-st1700

c1-st1800

c1-st1900

Files with SHA256: 2ed88a1499ccb9101035892c6f5e79b71db2ba1c25899e9d518168bb760ee262