MiniMax H3 + ControlNet Union dedicated local inpaint workflow. A pure repaint-focused build on the native ReferenceToVideo pipeline for precise local editing of one source video.
Main features:
- Native R2V architecture: the single 24fps source video feeds ref_videos (<Video 1>) in both LOW and HIGH passes; first/last frame plus 3 reference images map to <Picture 1>-<Picture 4>
- AIO preprocessor dropdown: Pose / Depth / Canny / HED / MLSD or none (passthrough)
- ControlNet Union 2.0 patch (minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot) applied at depth strength 1.8
- SeC-4B point tracking produces the full-clip mask, GrowMask dilation and mask preview
- Aspect switch 16:9 / 9:16; two-stage sizing without hard crop: 960x544 LOW pass, learned latent x2 upscale to 1920x1088 HIGH pass
- OpenVDN DMD8 8NFE base + clean 4NFE refine (HIGH does not inherit the LOW Union2 patch)
- Native audio AV output, 24fps, approx 1080P safe save; MODE 1/2/3 master switch
Suggested workflow:
1. Load one source video and the reference images
2. Choose the AIO preprocessor (or none), pick the repaint region via SeC points and confirm in the mask preview
3. Run the graph and export the approx 1080P / 24fps AV save
Dependencies:
- minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors (H3 hybrid UNET)
- minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot.safetensors (Union 2.0 patch)
- SeC-4B-bf16.safetensors (point tracking)
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (CLIP), video/audio VAEs
- minimax_h3_latent_upscaler_3d_fp16.safetensors (learned latent x2)
- OpenVDN/vdn-minimax-h3 (DMD8 stage), KSK T8 custom nodes
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2108582772978110465?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV1pTph6EEUv/
RunningHub registration with invite code rh-v1111 (1000 RH coins): https://www.runninghub.ai/?inviteCode=rh-v1111
Description
MiniMax H3 + ControlNet Union dedicated local inpaint workflow. A pure repaint-focused build on the native ReferenceToVideo pipeline for precise local editing of one source video.
Main features:
- Native R2V architecture: the single 24fps source video feeds ref_videos (<Video 1>) in both LOW and HIGH passes; first/last frame plus 3 reference images map to <Picture 1>-<Picture 4>
- AIO preprocessor dropdown: Pose / Depth / Canny / HED / MLSD or none (passthrough)
- ControlNet Union 2.0 patch (minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot) applied at depth strength 1.8
- SeC-4B point tracking produces the full-clip mask, GrowMask dilation and mask preview
- Aspect switch 16:9 / 9:16; two-stage sizing without hard crop: 960x544 LOW pass, learned latent x2 upscale to 1920x1088 HIGH pass
- OpenVDN DMD8 8NFE base + clean 4NFE refine (HIGH does not inherit the LOW Union2 patch)
- Native audio AV output, 24fps, approx 1080P safe save; MODE 1/2/3 master switch
Suggested workflow:
1. Load one source video and the reference images
2. Choose the AIO preprocessor (or none), pick the repaint region via SeC points and confirm in the mask preview
3. Run the graph and export the approx 1080P / 24fps AV save
Dependencies:
- minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors (H3 hybrid UNET)
- minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot.safetensors (Union 2.0 patch)
- SeC-4B-bf16.safetensors (point tracking)
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (CLIP), video/audio VAEs
- minimax_h3_latent_upscaler_3d_fp16.safetensors (learned latent x2)
- OpenVDN/vdn-minimax-h3 (DMD8 stage), KSK T8 custom nodes
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2108582772978110465?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV1pTph6EEUv/
RunningHub registration with invite code rh-v1111 (1000 RH coins): https://www.runninghub.ai/?inviteCode=rh-v1111