https://youtu.be/H5L5sO7dRu8

Creator: AI-KSK · Collection: Video 327 / six-workflow MiniMax H3 RefMod episode

Workflow file: 05_多参考生成_A_Apply.json · Cover file: 多参考生成 A：Apply 注入-5.png

Main features:

This is the first *generation* graph in the six-workflow set. It combines a saved character, motion, scene and audio/voice reference for one planned action sequence in MiniMax H3. The primary architectural feature is the order: the native MiniMaxH3ReferenceToVideo builds prompt conditioning and an initial latent; MiniMaxH3RefModApply then injects the already saved references into that conditioning. The H3 text encoder does not independently see the original images stored inside each RefMod during its initial prompt-encoding stage.

Exact connected references and settings

- Slot 1: character/Rin_Identity; slot 2: motion/Action_01; slot 3: scene/Dojo_OneLocation; slot 4: voice/Rin_Voice_01.
- Every selected slot has strength 1 and copies 1; remaining slots are (none).
- MiniMaxH3RefModApply: override=false, retention=1, curve_direction=constant, scramble_seed=-1, max_total_tokens=0.
- Generation: 832×480, length=243, fixed seed 9527, Euler, beta scheduler, 4 steps, denoise 1.
- MP4 output: 24 fps, H.264, yuv420p, CRF 19, prefix AIKSK_REF05_APPLY_.
- Both generators use an H3 ref2va pruned int8-convrot checkpoint, Qwen3-VL-32B NVFP4-AWQ text encoder, H3 video/audio VAEs and a selected Turbo 4-step model-only LoRA at strength 1; the checkpoint/LoRA files are not included.

The filled prompt instructs Rin to practice guarded punches and a roundhouse kick alone in the dojo, then speak a Mandarin line: “看清楚了吗？再来一次。” It requests no music and no additional characters. The stated behavior is a prompt, not a verified output. 243 frames at 24 fps are nominally 10.125 seconds; no measured runtime, frame fidelity or benchmark accompanies this upload.

Suggested workflow:

Run workflows 01–04 to make the four named references, or create your own and update the loader selections. Confirm every asset and exact checkpoint filename is available, refresh dropdowns if needed, and review the prompt and sound conditions. Queue generation, then inspect the actual video for reference mixing, identity drift, action continuity and unwanted background audio. The cover illustrates the four-input scheme and a martial-arts protagonist; it is not a rendered output.

Limits

This graph does not give two fighters independently controlled skeletons. Increasing reference copies is not a free speed or quality improvement, and the fixed seed alone does not guarantee identical outputs across different software versions.

Dependencies, tutorial and deployment

This project uses ComfyUI, MiniMax H3 weights and the upstream RefMod node pack; workflows involving VHS_ nodes require Video Helper Suite. See Comfy-Org's H3 model inventory for official distribution guidance. The exact int8-convrot video VAE and Turbo 4-step LoRA named in the creator's JSON need separate availability/provenance checks. We distribute this workflow JSON and its original cover, not external checkpoints or third-party media. Respect applicable rights and licenses.

Deploy AI-KSK's H3｜RefMod免训练超级参考 image on 优云智算. Referral disclosure: the deployment link includes an AI-KSK referral code and may support the creator. Check the cloud provider's actual billing and environment before deploying.

Watch the complete six-workflow YouTube tutorial. This project is one module in that episode. Description and settings were prepared by inspecting the source JSON and transcript, not from a new measured benchmark.

Bilibili tutorial: https://www.bilibili.com/video/BV1F2nZ6SEA7/