<div data-youtube-video="https://youtu.be/t7pv8YAzpzY"><iframe src="https://www.youtube.com/embed/t7pv8YAzpzY" width="853" height="480" allowfullscreen></iframe></div> 
 
MiniMax H3 action director workflow with 9 combat presets. One graph covers text-to-video, first/last-frame and dual-character reference-to-video, with preset LoRA combos that choreograph punches, blades and gun-fu between two characters. 
 
Main features: 
- Three modes behind control switches: text-to-video (20 steps), first/last frame (20 steps), reference-to-video with character A/B images (25 steps) 
- 9 combat presets (A0-A8): A0 basic fists, A1 high-dynamic repair, A2 two-person attack/defense (default), A3 comprehensive martial arts, A4 blade weapons, A5 gun-fu, A6 continuous gunfight, A7 close-quarters gun-fu, A8 general action repair 
- Auto LoRA management: H3_Combat_V2 + H3_Fight_Flow_V1 applied automatically per preset (0.65/0.25 refine weights); Bunny_weapon_combatV1, GunFu, Motion_Repair_V2 available for manual tuning 
- Optional second pass: 896x512 base pass refined through the MiniMax H3 3D latent upscaler to fixed 1344x768 (low-noise 0.30 denoise); otherwise the official ResolutionSelector picks aspect ratio and megapixels with automatic 32-pixel alignment 
- Duration input in seconds auto-converts to the H3-valid 17n+5 frame grid at 24fps 
- Native audio output (VAEDecodeAudio + CreateVideo/SaveVideo) 
 
Suggested workflow: 
1. Pick the mode with the left control switches, then upload the matching inputs (prompt only, first/last frames, or character A/B images) 
2. Choose the combat preset (start with A2) and enter the video duration in seconds 
3. Queue the graph; the preset LoRA combo, sampling and optional 1344x768 refine pass run automatically 
 
Dependencies: 
- minimax_h3_fl2va_pruned_int8_convrot.safetensors (H3 base UNET) 
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (CLIP) 
- minimax_h3_video_vae_int8_convrot.safetensors + minimax_h3_audio_vae_fp32.safetensors 
- minimax_h3_latent_upscaler_3d_fp16.safetensors (3D latent refine, xmarre node) 
- minimax_h3_ref_lora_layer20-49adaln_proj_rank_256_bf16.safetensors (reference LoRA) 
- H3_Combat_V2, H3_Fight_Flow_V1, Bunny_weapon_combatV1, GunFu, Motion_Repair_V2 LoRAs 
 
Bilibili tutorial video: https://www.bilibili.com/video/BV1oFJ16BEH5/ 
RunningHub registration with invite code rh-v1111 (1000 RH coins): https://www.runninghub.ai/?inviteCode=rh-v1111 RunningHub (this episode, Chinese page): https://www.runninghub.ai/post/2110642838903279617?inviteCode=rh-v1111

MiniMax H3 Action Director 9 Combat Presets

<div data-youtube-video="https://youtu.be/t7pv8YAzpzY"></div> 
 
MiniMax H3 action director workflow with 9 combat presets. One graph covers text-to-video, first/last-frame and dual-character reference-to-video, with preset LoRA combos that choreograph punches, blades and gun-fu between two characters. 
 
Main features: 
- Three modes behind control switches: text-to-video (20 steps), first/last frame (20 steps), reference-to-video with character A/B images (25 steps) 
- 9 combat presets (A0-A8): A0 basic fists, A1 high-dynamic repair, A2 two-person attack/defense (default), A3 comprehensive martial arts, A4 blade weapons, A5 gun-fu, A6 continuous gunfight, A7 close-quarters gun-fu, A8 general action repair 
- Auto LoRA management: H3_Combat_V2 + H3_Fight_Flow_V1 applied automatically per preset (0.65/0.25 refine weights); Bunny_weapon_combatV1, GunFu, Motion_Repair_V2 available for manual tuning 
- Optional second pass: 896x512 base pass refined through the MiniMax H3 3D latent upscaler to fixed 1344x768 (low-noise 0.30 denoise); otherwise the official ResolutionSelector picks aspect ratio and megapixels with automatic 32-pixel alignment 
- Duration input in seconds auto-converts to the H3-valid 17n+5 frame grid at 24fps 
- Native audio output (VAEDecodeAudio + CreateVideo/SaveVideo) 
 
Suggested workflow: 
1. Pick the mode with the left control switches, then upload the matching inputs (prompt only, first/last frames, or character A/B images) 
2. Choose the combat preset (start with A2) and enter the video duration in seconds 
3. Queue the graph; the preset LoRA combo, sampling and optional 1344x768 refine pass run automatically 
 
Dependencies: 
- minimax_h3_fl2va_pruned_int8_convrot.safetensors (H3 base UNET) 
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (CLIP) 
- minimax_h3_video_vae_int8_convrot.safetensors + minimax_h3_audio_vae_fp32.safetensors 
- minimax_h3_latent_upscaler_3d_fp16.safetensors (3D latent refine, xmarre node) 
- minimax_h3_ref_lora_layer20-49adaln_proj_rank_256_bf16.safetensors (reference LoRA) 
- H3_Combat_V2, H3_Fight_Flow_V1, Bunny_weapon_combatV1, GunFu, Motion_Repair_V2 LoRAs 
 
Bilibili tutorial video: https://www.bilibili.com/video/BV1oFJ16BEH5/ 
RunningHub registration with invite code rh-v1111 (1000 RH coins): https://www.runninghub.ai/?inviteCode=rh-v1111 RunningHub (this episode, Chinese page): https://www.runninghub.ai/post/2110642838903279617?inviteCode=rh-v1111

Files with SHA256: 2939f88be9c4bf857c864dc2e4ab14ab813a15f8f4a03adc3f69b09590daf615