RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2106048569508659202?inviteCode=rh-v1111

Bilibili tutorial video: https://www.bilibili.com/video/BV1VjaD6jEU3/

# MiniMax H3 — Standard RGB and Depth Reference Editing

## Purpose and audience

Inspect two source-specific examples using hybrid RGB/depth references and manually injected source-audio latents. Intended for users who can inspect matching ComfyUI dependencies and actual output media. This is a workflow resource, not a tested model release.

## Asset references

`minimax_h3_hybrid_fl2va_ref2va_zs05_b25-49_int8.safetensors`, `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`, `minimax_h3_video_vae_int8_convrot.safetensors`, `minimax_h3_audio_vae_fp32.safetensors`, `video_depth_anything_vits.pth`.

The main VDN composers reference OpenVDN/vdn-minimax-h3 and stage_dmd_8nfe. These are configuration references, not measured NFE, timing, memory, or compatibility results. Resolve matching SeC, VideoDepthAnything, H3 conditioning/VDN, audiovisual latent, mask, and save-node implementations. The SeC loader’s bfloat16 selection does not establish a checkpoint filename. Weights and selected input media are not bundled.

## Saved setup

81 nodes/113 links; both example pipelines are saved mode 0. Dimension targets differ: about0.9MP and0.48MP. Both inject encoded source audio; the second also connects native video-reference audio. Different media, prompts, and references mean this is not a controlled A/B test.

## How to prepare

Select an example and replace its source video, reference image, and points. Follow the grown/softened mask and depth composite into reference conditioning. Check loaded frame/fps metadata and effective dimensions. Distinguish generated-result saves from hybrid-reference visualization saves and review the soundtrack.

## Known limitations

Both roles on each AudioMix node receive the same source audio, not a decoded generated track. Peak-limit settings are not measured audio results or bit-perfect guarantees. Shared output prefixes require file identification. Unmasked RGB exists in the input reference, not a final original-background pasteback.

The RGB/depth blend is an input reference. Prompts asking for unchanged motion, identity or scenery do not prove preservation. Refresh points and inspect actual results, audio and filenames after a small run. No new inference, performance, visual/audio, or license verification is included.

## Files and version note

- 03: [Workflow JSON]() · [Selected cover]()

Historical VHS input-preview metadata is removed from distribution copies. Point-editor state, including embedded source imagery, is retained pending content/rights review. All other data is unchanged; originals are preserved. Selected covers are not independently verified inference samples.

This descriptive copy is not a remote release, model upgrade or passed benchmark.

## Tutorial and rights

[Source-provided tutorial](https://www.youtube.com/watch?v=UespHFT9cjc), not live-verified. The workflows are supplied through AI-KSK; confirm permissions for code, weights, nodes, source footage, identities, reference/editor images, covers and audio separately. No blanket commercial license is asserted.