https://youtu.be/3Ak4mkrhOrY 
 
This workflow packages the MiniMax H3 martial arts combat combo that gives the most reliable fight results right now. It starts from the author's dual-stage graph with 3D latent upscale and loads three purpose-built LoRAs on top: wushu action, spatial physics, and camera motion. The result is fight choreography with readable strikes, believable object behavior, and controlled camera movement, without rebuilding the loader, encoder, VAE, sampler, or video export chain by hand. 
 
The connected chain uses MiniMax-H3-int8.safetensors as the main UNET, the Qwen3-VL 32B MiniMax H3 text encoder (int8 convrot), the video VAE in fp16, and the audio VAE in fp32, so the final export carries both video and audio. Four LoRAs are active: wushu action FL2VA V7 (2000 step, no adaLN) at 0.4, spatial physics at 0.3, camera motion v1 3000 pruned at 0.3, and the FL2V turbo 8-step LoRA at 0.8, which lets the first stage run short. The photorealism LoRA and the legacy AnimateDiff combine node are both bypassed; the active output is the H3_Ref2VA combine node exporting H.264 at 24 fps with CRF 19. 
 
The graph runs 16:9 at 1344x768. The reference-to-video stage is set to 124 frames, about 5.2 seconds at 24 fps. MiniMax H3 frame counts must follow the 17n+5 pattern (56, 90, 107, 124, 175, 243), so a 10 second clip uses 243 frames. The reference image enters through the ReferenceToVideo node; the preset includes multiple reference slots and you enable any extra one with Ctrl+B. After the turbo first stage, the second stage upscales through the 3D latent upscaler to about 1.2 megapixels before final combining. 
 
Main features: 
 
- Dual-stage MiniMax H3 graph: short turbo first stage, then 3D latent upscale to about 1.2 MP 
- Wushu action LoRA V7 at 0.4 for fight choreography 
- Spatial physics LoRA at 0.3 for collisions, stacking and falling objects, the main source of hit impact 
- Camera motion LoRA v1 3000 pruned at 0.3 for controlled camera movement 
- Turbo 8-step FL2V LoRA at 0.8 keeps the first stage fast 
- Qwen3-VL 32B text encoder with int8 convrot weights 
- Video and audio VAEs connected for full media output 
- 1344x768 16:9 at 24 fps, frame count locked to the 17n+5 rule 
- Multi-slot reference input; enable extra slots with Ctrl+B 
- Photorealism LoRA and legacy AnimateDiff branch staged but bypassed 
- H.264 export with CRF 19 from the active H3_Ref2VA combine node 
 
Suggested workflow: 
 
Start with one reference image and a short prompt: wushu trigger first, camera motion trigger second, then one concrete fight move at 124 frames. Check strike readability and object physics before adding anything else. When describing techniques, write less but be specific; words from the author's vocabulary bank trigger the LoRAs more reliably than long improvised sentences. If the choreography drifts, make the technique description shorter and more concrete instead of longer. For a longer clip, switch to 243 frames and keep the same prompt structure. If the turbo result feels too soft for a hero shot, disable the turbo LoRA and raise the second-stage step count. Keep the spatial physics LoRA in the chain even in simple scenes, since it is what makes impacts land. MiniMax H3 prompts should not include CFG values. 
 
RunningHub Workflow 
 
Try the workflow online right now - no installation required. 
Workflow: https://www.runninghub.ai/zh-cn/post/2096152360778584065?inviteCode=rh-v1111 
 
If the results meet your expectations, you can later deploy it locally for customization. 
 
Fan Benefits: Register to get 1000 points + daily login 100 points - enjoy 4090 performance and 48 GB super power! 
 
Bilibili Updates (Mainland China &amp; Asia-Pacific) 
 
If you're in the Asia-Pacific region, you can watch the video below to see the workflow demonstration and creative breakdown. 
Bilibili Video: https://www.bilibili.com/video/BV1rBty6AENs/ 
 
Support Me on Ko-fi 
 
If you find my content helpful and want to support future creations, you can buy me a coffee. 
Every bit of support helps me keep creating. 
Ko-fi: https://ko-fi.com/aiksk 
 
Business Contact 
 
For collaboration or inquiries, please contact aiksk95 on WeChat. 
 
打开下方链接即可在线体验，无需安装。 
工作流：https://www.runninghub.ai/zh-cn/post/2096152360778584065?inviteCode=rh-v1111 
 
如果觉得效果理想，你也可以在本地进行自定义部署。 
 
粉丝福利：注册即可领取 1000 积分，每日登录再领 100 积分，体验 4090 和 48GB 大显存性能。 
 
B站视频（中国大陆及亚太地区） 
 
如果你在中国大陆或亚太地区，可以通过下面的视频查看工作流的实测效果与创作思路。 
B站视频：https://www.bilibili.com/video/BV1rBty6AENs/ 
 
我会在夸克网盘持续更新模型资源： 
https://pan.quark.cn/s/9ec095b95838

Files with SHA256: f690b7c08370b8d95a4356765df864b81dc5c22ee3a98bbbc0aae3aab4471a09