https://youtu.be/3Ak4mkrhOrY 
 
This is the multi-keyframe version of the MiniMax H3 martial arts combat workflow. It keeps the exact same dual-stage graph and three-LoRA combat stack as the base version, and adds keyframe anchoring on top: you can drop reference images at specific frames in the middle of a clip and use them as story anchors. That gives you much stronger control over what happens at a given moment in the fight, which is the main upgrade over the single-reference version. 
 
The connected model chain is identical to the base workflow: MiniMax-H3-int8.safetensors as the main UNET, the Qwen3-VL 32B MiniMax H3 text encoder, the video VAE in fp16, and the audio VAE in fp32, with the same four active LoRAs: wushu action FL2VA V7 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 for the fast first stage. The photorealism LoRA and the legacy AnimateDiff combine node stay bypassed; the active output is the H3_Ref2VA combine node exporting H.264 at 24 fps. 
 
The keyframe difference lives in the reference handling chain. This version adds image scaling stages (lanczos to 1344x768) so each anchor image is prepared before it enters the graph, and the AddGuide area lets you assign an image to a specific frame position. Each anchor is a single unique frame used as a visual anchor, not a sequence; the frame position must follow the 17n+5 pattern; and every anchor must sit inside the total render length, so in a 124 frame clip no anchor can be placed past frame 124. The base reference input still runs at 1344x768 16:9, and the second stage still upscales through the 3D latent upscaler to about 1.2 megapixels before final combining. 
 
Main features: 
 
- Same dual-stage MiniMax H3 graph as the base wushu workflow, with 3D latent upscale to about 1.2 MP 
- Multi-keyframe anchoring: reference images placed at specific frames inside the clip 
- AddGuide area for story anchor frames with 17n+5 frame position rules 
- Lanczos pre-scaling stages (1344x768) for each anchor image before it enters the graph 
- Wushu action LoRA V7 at 0.4, spatial physics LoRA at 0.3, camera motion LoRA at 0.3 
- Qwen3-VL 32B text encoder, video and audio VAEs connected for full media output 
- 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 from the base single-reference test until the choreography reads correctly, then add anchors one at a time. Place each anchor on a valid 17n+5 frame inside your total length, use a single unique frame per anchor, and keep the prompt description of that beat consistent with the image. 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 an anchor over-constrains the motion, move it to a less critical moment or use fewer anchors in one clip. For a 10 second sequence, 243 frames gives room for several anchors while staying inside the frame count rule. Review identity drift at each anchored beat before pushing more prompt detail. 
 
RunningHub Workflow 
 
Try the workflow online right now - no installation required. 
Workflow: https://www.runninghub.ai/zh-cn/post/2095907246101291010?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/2095907246101291010?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: e70d894526a2891a28d9dcde8718f668bb55ccf02fb21a3a7f523699dcacdb27