MiniMax H3 Video Workflows: Trace Control, Audio, and the Two-Pass Output

A video can supply structural control without being connected to a native video-reference input. Audio can drive a latent while a different audio path supplies the final file. And a workflow labeled 1080P can be saved with portrait dimension controls. Those are practical distinctions in this MiniMax H3 collection. The source headline says four workflows, but the supplied folder and source-link list contain five JSON files. They contain 457 node records in total, including repeated branches inside some files. No model execution, soundtrack inspection, timing test, or visual-result acceptance was performed for this guide. Choose the file by its actual route Code Purpose Nodes Important saved state 01 ControlNet with native-audio conditioning 120 Structure-only mode; Scribble preprocessing 02 General generation with manually injected audio latent 44 One reference image; portrait size controls; LOW and HIGH saves 03 General generation with native generated audio 43 Two active reference images; video input bypassed 04 Segmentation-assisted local-edit experiment 184 Middle clone active; inpaint-only mode 05 ControlNet plus manually injected audio latent 66 Structure-only mode; selected audio comes from video These are five supplied files, not five identical templates with different titles. Conversely, cloned branches inside a file should not automatically become separate published projects. Resolve assets before trusting labels Files 01, 02, and 05 select minimax_h3_fl2va_pruned_int8_convrot.safetensors. File 03 selects minimax_h3_fl_ref_hybrid_scheme_a_25-49_bf16.safetensors; file 04 selects minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors. Labels such as “H3 Base direct” are not a substitute for the actual loader selection. Common references include qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, minimax_h3_video_vae_fp16.safetensors, minimax_h3_audio_vae_fp32.safetensors, and minimax_h3_latent_upscaler_3d_fp16.safetensors. There is also a ControlNet naming discrepancy. Several active loaders select minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors, while titles or download metadata name a version containing _2.0. File 04's bypassed outer clones select the _2.0 filename, but its active middle clone does not. Resolve this discrepancy; neither the title nor a filename alone proves the weight contents or compatibility. The VDN composers reference OpenVDN/vdn-minimax-h3 and a stage_dmd_8nfe selection. Do not invent a separate distilled-LoRA filename: the inspected graphs do not provide one. Dependency references also do not prove installed files, verified hashes, licensing, or GPU support. Understand the two-pass spine The usual path is conditioning and an audiovisual latent, then a LOW execution plan and sampler, a learned latent upscale, reconciliation with HIGH conditioning, and a refinement plan and sampler. Audio-audit, audiovisual-decode, trim, and save nodes follow. The first plan's label and composer stage describe an eight-NFE intent; the refinement plan stores four refinement steps. This is not a twelve-step setting on one ordinary sampler, nor a measured count of model evaluations in a successful run. The saved random-noise seeds are 9527 and 9528 for the two passes. Reconciliation requests first-pass audio handling with second-pass audio strength zero. An audit node is configured to check the second-pass audio behavior. Its presence is a requested check, not evidence that the check passed or that a voice, soundtrack, or identity was preserved. 01 — Structural control is not native video-reference conditioning The active branch has MODE 70 set to 1. Two expressions select structural control for modes 1 or 3, and source-video/mask inputs for modes 2 or 3. Thus the saved selection is structure only, not simultaneous structure and inpainting. Source video 1 feeds a Scribble preprocessor at 512, then a resize and the control switch before MiniMaxH3FunControlNetApply. That Apply node's named controls store strength 1, start 0, and end 1. The preprocessor choice is observable; it does not independently establish how the custom ControlNet implementation interprets or detects every control type. The video also feeds the point/SeC segmentation route. Its grown mask reaches a preview and the optional mask switch. A connected mask preview means that turning off the mask input to the model is not sufficient evidence that all segmentation work disappears. The active LOW/HIGH native-conditioning nodes have unlinked video-reference ports, despite source notes claiming that source frames reach those ports. Describe the connected structural route, not the note's stronger claim about native video references. There is a mostly bypassed duplicate branch. Some of its stored named-port assignments need review: an image loader feeds a video-reference port, and a LOW-height switch is mapped to audio_mode at node 208. Do not simply enable every clone and assume a clean alternate configuration. This package preserves the source connections rather than making an untested repair. For a first check, replace the source video, inspect preprocessing and the enabled branch, then verify effective dimensions and duration. The requested duration is five seconds; planner outputs, not stale conditioner length widgets, supply the frame budget and trim controls. The final audio path comes from generated-audio decoding, not the source video's soundtrack. 02 — Audio conditioning and output sound follow different paths This file selects an uploaded audio file and a single reference image. The audio window receives the loader's reported audio length; its old five-second widget is not the effective duration input. The upload is configured with a roughly ten-second selection, but the actual file and resulting window still need inspection. Windowed audio passes through separation. Only the vocals output reaches audio-VAE encoding and a zero latent-noise mask. The LOW audiovisual template is split, and its video latent is concatenated with that encoded audio before planning and sampling. The final output audio does not come from the decoder's generated-audio output. The selected original audio feeds both output trims. This describes routing, not bit-perfect passthrough or proven synchronization. The saved dimensions are portrait Resolution selector 47 is saved as 3:4 with 544×736 controls. Those linked values override the old 960/544 primitive defaults and titles. The learned upscale is configured as ×2, suggesting approximately 1088×1472 dimension controls downstream, not a fixed landscape 1920×1088 setup. Confirm actual output dimensions in the matching implementation. There are two save routes The LOW sampler also feeds a decode/trim/save chain, alongside the HIGH refined chain. Both saves are active and use the same prefix. Inspect the resulting names to distinguish LOW from HIGH; do not promise only one final file or assume how numbering handles that shared prefix. Prepare the authorized reference image and audio, check the window and resolution selector, and inspect both output chains. Lip synchronization and audio-latent alignment require an actual test, not a historical note describing a previous fix. 03 — Native audio, with a different current input selection This graph is documented as a general native-audio route. Its current save state enables reference images 79 and 80. The video loader, slice, video-component extractor, first/last frames, and third reference image are bypassed. That is different from the note saying the basic example loads one visual-reference video. There is also a bypassed LoadAudio node with a None selection connected to a LOW reference-audio port. Do not rewrite this as “no audio-input node exists”; check how the installed implementation handles the saved bypass state. The LOW/HIGH conditioners request AUTO/native audio behavior. After sampling, the selected audiovisual latent is decoded, and generated audio goes through trimming to the final saver. The source video's audio output is not the selected soundtrack route. The saved format switch is landscape: LOW controls are 960×544 and the learned upscale is configured ×2. Source notes describe 1920×1088 rather than a hard-cropped 1920×1080 result. The ten-second duration control feeds a custom planner; its actual frame and trim outputs were not computed or executed in this review. Edit the prompt and the two active reference inputs, or deliberately configure another route in a separate working copy. Do not assume optional first/last images are active merely because their nodes are present. 04 — Inspect the active middle clone before editing masks This file contains three cloned main branches. The outer two are saved bypassed; the middle is active and sets MODE 175 to 2, selecting source-video and mask inputs while gating off structural control. The active source loader skips 24 frames and requests a four-second scene through the planner. Its AIO preprocessor stores Canny at 768, despite stale Depth titles and a DepthAnything note. Since the selected mode gates off the structural route, advertising it as a current Depth-control example would be doubly misleading. The active segmentation route selects SeC-4B-bf16.safetensors, uses positive and negative points, grows the mask, and supplies it alongside source video to Union. Refresh the point annotations against the actual loaded frames: saved editor canvas values and coordinates are not proof of correct runtime alignment. The native reference-video ports remain unlinked. The separate picture loaders feeding the active reference-conditioning nodes are bypassed, so their presence does not establish five usable character references. The LOW model path receives Union processing, while HIGH uses the selected base-loader path directly before its own VDN composer. Do not describe the HIGH stage as automatically inheriting the same patched ControlNet model. Generated audio feeds the active final trim/save path. Extra bypassed saves and an unconnected SaveImage node are not evidence of additional finished outputs. 05 — ControlNet plus manual audio has its own defaults MODE 70 is 1, so this file currently uses structural control. The separate-audio toggle is false: video audio, not the separately loaded audio file, is selected for both the audio-window path and output trimming. The audio window requests five seconds. The video loader has no corresponding five-second frame cap, so do not assume this trims every frame sent to preprocessing and segmentation. Check visual and audio alignment in a real run. The structural path uses DepthAnythingV2 at 512, then resizing and the Union control input. The source/mask model inputs are gated off, but the mask-preview branch remains connected. Refresh the saved points and inspect any segmentation computation rather than assuming it is disabled. Like file 02, separated vocals are encoded and masked before entering the sampling latent. Unlike file 02, this graph sends the complete LOW template directly to the concatenation node without a separate split node. A source note describes intended temporal fitting; it does not verify the custom implementation or a successful correction. The native reference-conditioning nodes have no connected image, video, or audio references, even though the saved prompt talks about references. Structural video control and manual audio-latent injection are separate paths. The final sound is the selected original audio through trimming, not the decoder's unused generated-audio output. Check the route before judging the result For each file, verify model selections, enabled branches, prompt and media inputs, linked size/duration controls, and actual save destinations. Use authorized images, voices, and footage. Run a small test before a full experiment, then inspect continuity, masks, dimensions, timing, audio, and unexpected changes. Distribution copies omit historical VHS video-preview metadata. Point-editor state, including embedded editor imagery where present, is retained rather than altered without implementation evidence; review it and its rights before any publication. Neither editor images nor cached UI text are fresh inference samples. The source-provided tutorial supplies context; its link was not live-verified here. The work remains a source-based guide, not an installation certification, measured 1080P result, voice-preservation guarantee, or license grant. Getting the workflows Published by this channel. The graphs are free; weights, LoRAs and ControlNet files are separate and carry their own terms. Hugging Face: AI-KSK/minimax-h3-cn2-video-control-315 — five JSONs, covers, SHA256 manifest. Start here. ModelScope (魔搭): Sunjian520/minimax-h3-cn2-video-control-315 Civitai: 01 CN2.0 control + inpaint · 02 audio-driven avatar · 03 general video · 04 inpaint-only · 05 combined Run without a local GPU: 01 · 02 · 03 · 04 · 05 Tutorial: YouTube · Bilibili