AI-KSK/minimax-h3-refmod-workflows-327

MiniMax H3 RefMod: Six ComfyUI Workflows | AI-KSK Video 327

A single, reusable-reference workflow collection for MiniMax H3 in ComfyUI. This repository holds six original, unmodified JSON workflows and their six original 3:4 cover images. The first four graphs create reusable RefMod assets; the last two use saved assets to generate H3 video and audio. It is one episode and one collection, not six independent repositories.

Full video tutorial: YouTube — AI-KSK MiniMax H3 RefMod (six workflows)

Deploy the prepared AI-KSK image: H3｜RefMod免训练超级参考 on 优云智算. Disclosure: this is a creator referral link and may support AI-KSK. Provider fees and instance/model availability should be checked on the deployment page.

This README describes the JSON settings and the supplied episode transcript. It is not a claim that these exact files were independently executed or benchmarked. The cover artwork demonstrates what each workflow is for; it is not a gallery of proven model outputs.

Workflow catalog

#	Graph	Original JSON	Cover	Real role
01	Character	JSON		Create character/Rin_Identity from the connected folder-loader route
02	Motion	JSON		Create motion/Action_01 from the connected video loader
03	Audio	JSON		Create voice/Rin_Voice_01 via LoadAudio and H3 audio VAE
04	Scene	JSON		Create scene/Dojo_OneLocation from two images
05	Generation A	JSON		Load four saved references and inject them via Apply
06	Generation B	JSON		Encode references with numbered text-reference mapping

All six covers are 1086 × 1448 PNGs. Images and JSONs were preserved byte for byte. Model weights, media sources, saved .safetensors RefMods, video outputs and any third-party license grant are not part of the release.

Quick start

1. Prepare a ComfyUI build with MiniMax H3 support, ComfyUI-MiniMaxH3Mod and Video Helper Suite where VHS_ nodes are used. The selected H3 models, VAEs, text encoder and LoRA must also exist and be compatible.
2. Download the six JSONs from workflows/ and load them in ComfyUI, or use the AI-KSK prepared cloud image linked above. A prepared image does not remove the need to verify dependencies and provide your own inputs.
3. Supply your own rights-cleared media: character images in a ComfyUI input subfolder for 01; a motion video for 02; a voice file for 03; and two consistent scene views for 04. Original example sources are not bundled.
4. Run 01–04. Confirm new saved RefMods are present under ComfyUI/models/refmods/character/, motion/, voice/, and scene/. Refresh loaded node options if newly created files are missing.
5. Run 05 with the four selected references, or prepare 06 by adding its missing second-character asset and filling its empty prompt. Inspect the reference map before using <Video n> and <Picture n> labels. Review the actual generated output rather than the marketing covers.

Exact reference-creation settings

Field	01: Character	02: Motion	03: Audio	04: Scene
Node	MiniMaxH3RefModExtract	same	MiniMaxH3RefModAudioExtract	MiniMaxH3RefModExtract
Saved name	Rin_Identity	Action_01	Rin_Voice_01	Dojo_OneLocation
Subfolder	character	motion	voice	scene
Concept type	identity	pose_motion	voice	background
Mode / limit	Full Reference, 768 resolution	Full Reference, 512 resolution	8-second max	Full Reference, 1024 resolution
Source actually wired	MiniMaxH3RefModFolderLoader	VHS_LoadVideo	LoadAudio	2× LoadImage
Detail	32 items, 240 max frames, 1024 max edge	video load cap 121 frames; 56 extractor latent frames	max_tokens=0, truncate	2 images; 16 extractor latent frames

The character and scene extractors also select latent_frames=16, pool_h=pool_w=16. The motion extractor selects pool_h=pool_w=16, motion_only=false (not enabled) and save=true. The audio extractor uses minimax_h3_audio_vae_fp32.safetensors and save=true. The character loader can be replaced with a direct image route if you wire it; the motion JSON currently contains no connected folder batch reader, notwithstanding its descriptive filename and the optional process explained in the episode.

Exact generation graph settings

Configuration	05: Apply	06: Text Encode
Connected RefMod loader	Rin_Identity, Action_01, Dojo_OneLocation, Rin_Voice_01	Rin_Identity, Action_01, Dojo_OneLocation, character/war
Conditioning path	MiniMaxH3ReferenceToVideo → MiniMaxH3RefModApply → guider	MiniMaxH3RefModTextEncode → guider; no Apply
Image size	832 × 480	832 × 480
Frame count	243	73
Video output fps	24	24
Sampler	Euler, beta scheduler, 4 steps, denoise 1	Euler, beta scheduler, 4 steps, denoise 1
Seed	9527, fixed	9527, fixed
Text prompt	Fully written, one Rin action shot with Mandarin dialogue	Empty (""): author before running
RefMod controls	Apply retention 1; loader strengths/copies 1	reference_fps=24, max_total_tokens=0; loader strengths/copies 1
Voice ref selected?	Yes	No (audio VAE used for output decoding only)

Both renderers select minimax_h3_ref2va_pruned_int8_convrot.safetensors, qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, minimax_h3_video_vae_int8_convrot.safetensors, minimax_h3_audio_vae_fp32.safetensors and minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors (LoRA strength 1). The final MP4 is assembled using VHS_VideoCombine at 24 fps, H.264, yuv420p, CRF 19. A's 243 frames amount to approximately 10.125 seconds and B's 73 frames to approximately 3.04 seconds by configuration, not an observed playback or timing result.

B needs additional setup: character/war is not generated by any supplied creator JSON and is not bundled as an asset. Also, the separate inspector's loader is a secondary inspection path, not additional conditioning; the connected loader feeding the Text Encode node is the relevant one. RefMod Text Encode already attaches refs; do not duplicate injection by adding Apply with the same bundle.

Dependencies and model-source guidance

• RefMod node author: Luisacaotica/ComfyUI-MiniMaxH3Mod.
• H3 ComfyUI model location reference: Comfy-Org/MiniMax-H3; original weights: MiniMaxAI/MiniMax-H3.
• VHS_LoadVideo/VHS_VideoCombine: ComfyUI Video Helper Suite.

The Comfy-Org page lists minimax_h3_video_vae_fp16.safetensors; the shared workflows instead request minimax_h3_video_vae_int8_convrot.safetensors. The selected Turbo 4-step LoRA also needs independent provenance/version verification. Do not invent download URLs or silently rename model files. The original video/image/audio inputs and character/war are missing from this collection by design. Confirm all relevant terms before running third-party weights; this repository does not assert one universal license for different authors' components.

When to choose which renderer

05 Apply is a practical start for a scripted, single-character action shot with identity, motion, environment and voice references already saved. The plain H3 prompt conditioning is built first; pre-encoded refs are appended afterwards.

06 Text Encode is better suited to constructing a prompt around a reference_map and identifying specific saved visual references as <Video n>/<Picture n> after inspection. It requires the missing character/war ref and a new prompt, and does not currently use the voice RefMod. Neither graph is a controlled speed/quality benchmark, since their source sets, prompts and frame counts differ.

Project scope and disclosure

This collection is authored by AI-KSK as accompanying workflow/tutorial content. It distributes the provided JSONs and illustrative covers only. No pre-generated MP4, input media, external checkpoint or dependency bundle is included. Respect rights to source portraits, voice recordings, footage and the applicable upstream model/node licenses. For the full walk-through: YouTube · Bilibili. Additional image projects are indexed in the AI-KSK mirror directory, but this episode's image above is the primary deployment link.

Individual workflow downloads on Civitai

• 01 — MiniMax H3 RefMod Character Reference
• 02 — MiniMax H3 RefMod Motion Reference
• 03 — MiniMax H3 RefMod Audio Reference
• 04 — MiniMax H3 RefMod Two-View Scene
• 05 — MiniMax H3 Multi-Reference Apply
• 06 — MiniMax H3 Multi-Reference TextEncode