<div data-youtube-video="https://youtu.be/ZP3rffpcIXg"><iframe src="https://www.youtube.com/embed/ZP3rffpcIXg" width="853" height="480" allowfullscreen></iframe></div> 
 
MiniMax H3 Ref2VA multi-reference + keyframe story control, with Motion Lab action repair - a single ComfyUI workflow that turns multiple reference pictures into one coherent cinematic video. 
 
This workflow is built around the MiniMax H3 Ref2VA model (minimax_h3_ref2va_int8_convrot.safetensors) and drives it through a multi-pass SamplerCustomAdvanced pipeline with res_multistep sampling, a Qwen3-VL 32B text encoder (qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors), separate video VAE (minimax_h3_video_vae_fp16) and audio VAE (minimax_h3_audio_vae_fp32), plus a compatible 3D latent upscaler pass at 1.5x before the final save. 
 
Main features: 
- Multi-reference fusion: bind subject, outfit, environment and prop to &lt;Picture N&gt; slots in the prompt so identity stays consistent across shots. 
- Keyframe story control: anchor specific frames to narrative beats inside one continuous generation. 
- Motion Lab action repair: fixes motion artifacts on fast-action sequences. 
- Two-stage output: an 8-second PASS1 preview plus the final AllInOne video, with a 3D HiRes 1.5x upscale pass before saving. 
 
Suggested workflow: 
1. Load your reference pictures (subject / outfit / environment / prop) into the image slots. 
2. Write subject_definitions for each &lt;Picture N&gt;, then one summary line describing the full sequence. 
3. Run PASS1 to check composition, then run the final pass with the 1.5x 3D upscale enabled. 
 
Notes: 5 output-reachable nodes are intentionally bypassed in this build; their effects are disabled by design. Requires MiniMax H3 Ref2VA int8 weights, the Qwen3-VL 32B encoder and both video/audio VAEs. 
 
RunningHub (one-click run): https://www.runninghub.ai/zh-cn/post/2095021906982813697?inviteCode=rh-v1111 
Bilibili tutorial: https://www.bilibili.com/video/BV1L4tn6WEVb/ 
 
Models &amp; resources: 
MiniMaxAI/MiniMax-H3-Turbo-Lora: https://huggingface.co/MiniMaxAI/MiniMax-H3-Turbo-Lora 
lightx2v/Minimax-h3-Turbo: https://huggingface.co/lightx2v/Minimax-h3-Turbo 
fal/MiniMax-H3-Realism-People-LoRA: https://huggingface.co/fal/MiniMax-H3-Realism-People-LoRA 
Model pack (Quark): https://pan.quark.cn/s/9ec095b95838 
 
RunningHub signup with invite code rh-v1111 gets 1000 RH coins: https://www.runninghub.ai/?inviteCode=rh-v1111

MiniMax H3 Ref2VA Multi-Reference Story Control

<div data-youtube-video="https://youtu.be/ZP3rffpcIXg"></div> 
 
MiniMax H3 Ref2VA multi-reference + keyframe story control, with Motion Lab action repair - a single ComfyUI workflow that turns multiple reference pictures into one coherent cinematic video. 
 
This workflow is built around the MiniMax H3 Ref2VA model (minimax_h3_ref2va_int8_convrot.safetensors) and drives it through a multi-pass SamplerCustomAdvanced pipeline with res_multistep sampling, a Qwen3-VL 32B text encoder (qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors), separate video VAE (minimax_h3_video_vae_fp16) and audio VAE (minimax_h3_audio_vae_fp32), plus a compatible 3D latent upscaler pass at 1.5x before the final save. 
 
Main features: 
- Multi-reference fusion: bind subject, outfit, environment and prop to &lt;Picture N&gt; slots in the prompt so identity stays consistent across shots. 
- Keyframe story control: anchor specific frames to narrative beats inside one continuous generation. 
- Motion Lab action repair: fixes motion artifacts on fast-action sequences. 
- Two-stage output: an 8-second PASS1 preview plus the final AllInOne video, with a 3D HiRes 1.5x upscale pass before saving. 
 
Suggested workflow: 
1. Load your reference pictures (subject / outfit / environment / prop) into the image slots. 
2. Write subject_definitions for each &lt;Picture N&gt;, then one summary line describing the full sequence. 
3. Run PASS1 to check composition, then run the final pass with the 1.5x 3D upscale enabled. 
 
Notes: 5 output-reachable nodes are intentionally bypassed in this build; their effects are disabled by design. Requires MiniMax H3 Ref2VA int8 weights, the Qwen3-VL 32B encoder and both video/audio VAEs. 
 
RunningHub (one-click run): https://www.runninghub.ai/zh-cn/post/2095021906982813697?inviteCode=rh-v1111 
Bilibili tutorial: https://www.bilibili.com/video/BV1L4tn6WEVb/ 
 
Models &amp; resources: 
MiniMaxAI/MiniMax-H3-Turbo-Lora: https://huggingface.co/MiniMaxAI/MiniMax-H3-Turbo-Lora 
lightx2v/Minimax-h3-Turbo: https://huggingface.co/lightx2v/Minimax-h3-Turbo 
fal/MiniMax-H3-Realism-People-LoRA: https://huggingface.co/fal/MiniMax-H3-Realism-People-LoRA 
Model pack (Quark): https://pan.quark.cn/s/9ec095b95838 
 
RunningHub signup with invite code rh-v1111 gets 1000 RH coins: https://www.runninghub.ai/?inviteCode=rh-v1111

Files with SHA256: 5be1d49fe54a0b05665928373ada4cc305c3b1a425107ab349a86b08268be5b6