[H3剑心比赛] rtx5070 minimax 视频生成
-
[H3剑心比赛] RTX PRO 6000 96GB 单卡|MiniMax H3 3D / 2D
参赛帖 ID: 本帖3D CGI 成片: https://www.youtube.com/watch?v=feXrBkh6K-8
传统 2D 成片: https://www.youtube.com/watch?v=QMGr8VBMum4
硬件配置
代码块GPU: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
VRAM: 96GB (97,887 MiB) 运行中占用 ~32GB(NInfer 30GB + ComfyUI 2GB)
主机: astro-5090 (.14, 192.168.1.14)
系统: Ubuntu 22.04 LTS, PyTorch 2.14.0+cu130, Python 3.11.9
ComfyUI: 0.37.0, comfy-kitchen 0.2.35, comfy-aimdo 0.5.5
加速: INT8 / SOL Attention / H3 turbo 4-step LoRA (BF16)
工作流
Goohai MiniMax H3 Integration 节点工作流(Ref2VA 模式)关键节点:
MiniMaxH3IntegrationGH - H3 整合节点(all_reference 模式)
MiniMaxH3IntegrationAdapterGH - 适配器(提取 positive/av_latent 等)
UNETLoader - 加载 minimax_h3_fl2va_pruned_bf16
LoraLoaderModelOnly - H3 turbo 4-step LoRA (768p)
MiniMaxH3DualClockT8GH - 双时钟采样 (video shift 12.0 / audio shift 3.0)
SamplerCustomAdvanced - 自定义采样器 (4 steps)
MiniMaxH3AVDecodeT8GH - 音视频解码
CreateVideo - 创建视频(接 audio 线)
SaveVideo - 保存为 h264/mp4
参数:aspect: 16:9
megapixels: 1.0 → 输出 1376×768(接近比赛要求 1344×768)
duration: 6 秒 × 2 段拼接 = 13.2 秒
steps: 4 (turbo)
seed: 42 (固定)
运行时间
任务 Total Time 峰值显存 ffprobe 验证
3D 成品 ~260s ~32 GB 1376×768 · 24 fps · 316 帧 · 13.18 s
2D 成品 ~255s ~32 GB 1376×768 · 24 fps · 316 帧 · 13.18 s
Total Time 从任务启动到 MP4 导出完成,含模型加载、 文本编码、参考图编码、采样、CUDA 解码和封装。画面对照
3D CGI 风格
时间点 描述
0.5s 剑心突进,低角度镜头
4.0s 双剑碰撞,火焰冲击波
8.0s 环绕镜头,能量爆发
12.0s 剑心落地,志志雄指剑
2D Traditional 风格
时间点 描述
0.5s 赛璐璐风格剑心突进
4.0s 蓝白剑光碰撞特效
8.0s 火焰背景环绕镜头
12.0s 双角色对峙
命令行输出
bash清显存
POST /free {"unload_models":true,"free_memory":true}
提交任务 (每个 6s 片段)
curl -s http://127.0.0.1:8188/prompt
-H "Content-Type: application/json"
-d '{"prompt": <workflow_json>, "client_id": "competition_3045_v2"}'输出结果 (所有 1376×768, 24fps, AAC 音轨)
3D_seg1_00001_.mp4 6.58s 3.6MB
PASS
3D_seg2_00001_.mp4 6.58s 3.9MB
PASS
2D_seg1_00001_.mp4 6.58s 3.3MB
PASS
2D_seg2_00001_.mp4 6.58s 3.5MB
PASS拼接 ffmpeg
ffmpeg -y -i 3D_seg1.mp4 -i 3D_seg2.mp4
-filter_complex "[0:v][0:a][1:v][1:a]concat=n=2:v=1:a=1"
-map "[outv]" -map "[outa]" 3D_v2_final.mp4最终成品验证
ffprobe -select_streams v:0 -show_entries stream=width,height,r_frame_rate 3D_v2_final.mp4
输出: 1376,768,24/1
提示词 (Prompt)
代码块(Style: 3D CGI animation) Masterpiece, high quality, cinematic lighting.
A fierce battle sequence featuring two characters based on reference images —
a red-clad swordsman with orange ponytail and scar on cheek, wearing red gi and white hakama, wielding a reverse-blade katana;
and a fierce warrior wrapped in white bandages with flowing ribbon appendages, dark blue robes.
The camera begins with a dynamic low-angle tracking shot.
The swordsman in red dashes forward with a long reverse-blade sword, leaving a glowing trail,
while the bandaged warrior executes a sweeping counter-attack with flaming energy effects.
They clash in the center of the frame, creating a shockwave of energy and sparks.
The camera rapidly orbits around them as the fiery battle intensifies,
concluding with the red-clad swordsman landing gracefully while the bandaged warrior points his weapon forward.
Dramatic background with embers floating in the air, high contrast cinematic lighting.
在 Terry 的原比赛主题回帖
参赛帖 ID:本帖
3D YouTube:https://www.youtube.com/watch?v=feXrBkh6K-8
2D YouTube:https://www.youtube.com/watch?v=QMGr8VBMum4
硬件、工作流 JSON、加速配置、Total Time、完整 CLI 日志及截图:见附件硬件证据:RTX PRO 6000 96GB · 97887 MiB VRAM · ComfyUI 0.37.0 · Goohai GH 15 nodes
生成时间:2026-11-14 10:15-10:30 UTC+8
竞赛:lcz.me/topic/3045 MiniMax H3 对決
作品位置:/data/ComfyUI/output/competition_3045_v2/ -
-
-
,
M Mountain HO 引用了 此主题
-
MiniMax H3 视频生成工作流整理(DSH 自动操作 ComfyUI)
背景说明
由于不直接操作 ComfyUI,而是用 DSH 直接叫它操作 ComfyUI 自动生成:
- 好处:不用管那么多细节
- 坏处:控制权不在自己手上
一、GH 节点工作流(.17 / .18 用,.14 相同)
完整工作流 JSON
{ "1": { "class_type": "MiniMaxH3IntegrationGH", "inputs": { "main_mode": "all_reference", "clip_name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors", "video_vae_name": "minimax_h3_video_vae_fp16.safetensors", "audio_vae_name": "minimax_h3_audio_vae_fp32.safetensors", "aspect": "16:9", "megapixels": 1.0, "duration_seconds": 6, "prompt": "(Style: 3D CGI animation) Masterpiece, high quality, cinematic lighting. A fierce battle sequence featuring two characters - a red-clad swordsman with orange ponytail and scar on cheek, wearing red gi and white hakama, wielding a reverse-blade katana; and a fierce warrior wrapped in white bandages with flowing ribbon appendages, dark blue robes. The camera begins with a dynamic low-angle tracking shot. The swordsman in red dashes forward with a long sword, while the bandaged warrior executes a sweeping counter-attack with flaming energy effects. They clash in the center of the frame, creating a shockwave of energy and sparks. The camera rapidly orbits around them as the fiery battle intensifies, concluding with the red-clad swordsman landing gracefully while the bandaged warrior points his weapon forward. Dramatic background with embers floating, high contrast lighting.", "task_type": "ref2va", "audio_mode": "native", "strict_prompt_tags": false, "audio_denoise_strength": 0.0, "drive_audio_ordinal": 0, "ref_image_size": "match", "first_frame": "", "last_frame": "", "hybrid_audio": "", "ref_image_1": "zhizhixiong_front.jpeg", "ref_image_2": "jianxin_front.jpeg", "ref_image_3": "", "ref_image_4": "", "ref_image_5": "", "ref_image_6": "", "ref_image_7": "", "ref_image_8": "", "ref_image_9": "", "ref_video_1": "", "ref_video_2": "", "ref_video_3": "", "ref_audio_1": "", "ref_audio_2": "", "ref_audio_3": "" } }, "2": { "class_type": "MiniMaxH3IntegrationAdapterGH", "inputs": { "integration": ["1", 0] } }, "3": { "class_type": "UNETLoader", "inputs": { "unet_name": "minimax_h3_fl2va_pruned_bf16.safetensors", "weight_dtype": "default" } }, "4": { "class_type": "LoraLoaderModelOnly", "inputs": { "model": ["3", 0], "lora_name": "minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors", "strength_model": 1.0 } }, "5": { "class_type": "MiniMaxH3DualClockT8GH", "inputs": { "model": ["4", 0], "av_latent": ["2", 1], "steps": 4, "shift_video": 12.0, "shift_audio": 3.0 } }, "6": { "class_type": "BasicGuider", "inputs": { "model": ["5", 0], "conditioning": ["2", 0] } }, "7": { "class_type": "SamplerCustomAdvanced", "inputs": { "noise": ["8", 0], "guider": ["6", 0], "sampler": ["5", 1], "sigmas": ["5", 2], "latent_image": ["2", 1] } }, "8": { "class_type": "RandomNoise", "inputs": { "noise_seed": 42 } }, "9": { "class_type": "MiniMaxH3AVDecodeT8GH", "inputs": { "av_latent": ["7", 0], "video_vae": ["2", 2], "audio_vae": ["2", 3] } }, "10": { "class_type": "CreateVideo", "inputs": { "images": ["9", 0], "audio": ["9", 1], "fps": 24.0, "pixel_format": "sRGB" } }, "11": { "class_type": "SaveVideo", "inputs": { "video": ["10", 0], "filename_prefix": "competition/3D_seg", "format": "auto", "codec": "h264" } } }各机差异(只改两处)
机 unet_name(节点 3)megapixels实测结果 .14 minimax_h3_fl2va_pruned_bf16.safetensors1.0 1376×768 
.17 minimax_h3_fl2va_pruned_bf16.safetensors1.0 1376×768
~5min.18 minimax_h3_fl2va_pruned_int8_convrot.safetensors1.0 1376×768
101s.13 —(GH 节点在 ROCm 卡死,不用) — 
注意:
megapixels是 number(不是字符串);duration_seconds只收整数 2–30,6s → 124 帧(snap 17k+5)。
二、原生节点工作流(.13 ROCm 用,实测可行)
完整工作流 JSON
{ "1": { "class_type": "UNETLoader", "inputs": { "unet_name": "minimax_h3_fl2va_pruned_int8_convrot.safetensors", "weight_dtype": "default" } }, "2": { "class_type": "LoraLoaderModelOnly", "inputs": { "model": ["1", 0], "lora_name": "minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors", "strength_model": 1.0 } }, "3": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors", "type": "minimax" } }, "4": { "class_type": "VAELoader", "inputs": { "vae_name": "minimax_h3_video_vae_fp16.safetensors" } }, "5": { "class_type": "VAELoader", "inputs": { "vae_name": "minimax_h3_audio_vae_fp32.safetensors" } }, "6": { "class_type": "MiniMaxH3ReferenceToVideo", "inputs": { "clip": ["3", 0], "vae": ["4", 0], "audio_vae": ["5", 0], "prompt": "(同一段比赛提示词)", "width": 512, "height": 288, "length": 124, "ref_image_size": "match", "ref_images.ref_image_0": ["40", 0], "ref_images.ref_image_1": ["41", 0] } }, "7": { "class_type": "BasicGuider", "inputs": { "model": ["2", 0], "conditioning": ["6", 0] } }, "8": { "class_type": "BasicScheduler", "inputs": { "model": ["2", 0], "scheduler": "simple", "steps": 4, "denoise": 1.0 } }, "9": { "class_type": "KSamplerSelect", "inputs": { "sampler_name": "res_multistep" } }, "10": { "class_type": "RandomNoise", "inputs": { "noise_seed": 42 } }, "11": { "class_type": "SamplerCustomAdvanced", "inputs": { "noise": ["10", 0], "guider": ["7", 0], "sampler": ["9", 0], "sigmas": ["8", 0], "latent_image": ["6", 1] } }, "12": { "class_type": "VAEDecode", "inputs": { "samples": ["11", 0], "vae": ["4", 0] } }, "13": { "class_type": "VAEDecodeAudio", "inputs": { "samples": ["11", 0], "vae": ["5", 0] } }, "14": { "class_type": "CreateVideo", "inputs": { "images": ["12", 0], "audio": ["13", 0], "fps": 24.0, "pixel_format": "sRGB" } }, "15": { "class_type": "SaveVideo", "inputs": { "video": ["14", 0], "filename_prefix": "native13/test", "format": "auto", "codec": "h264" } }, "40": { "class_type": "LoadImage", "inputs": { "image": "jianxin_front.jpeg" } }, "41": { "class_type": "LoadImage", "inputs": { "image": "zhizhixiong_front.jpeg" } } }
️ 原生节点最大陷阱ref 图必须用扁平点号 key
"ref_images.ref_image_0"。写成嵌套
{"ref_images":{"ref_image_0":...}}会静默丢弃参考图(出片同没 ref 逐像素一样),零警告。.13 实测
- 512×288 / 124 帧 = 108 秒出片
- 1344×768 会卡在采样 0/4(ROCm 上限)
三、两套怎么选
场景 用哪套 NVIDIA(.14/.16/.18) GH 节点(快、支持多参考图、音频接线简单) .17 GB10 GH 节点(bf16,加载慢但可行) .13 AMD ROCm 原生节点,且只限 ≤512×288 .19 20GB 两套都放不进 38G bf16,需 Q4 GGUF 文件位置
- GH:
@/home/astro/docs/.dsh-scratch/h3-jianxin/submit_h3.py- 可用
H3_SERVER/H3_MP/H3_UNET环境变量切机
- 可用
- 原生:
@/home/astro/docs/.dsh-scratch/h3-jianxin/submit_native13.py
-
-
AI有整理成markdown,但是我唔知點上傳markdown格式, 我只能COPY 貼上去
-