MiniMax H3 Ref2VA - Multi-Reference Video Generation

RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2090060185078419458?inviteCode=rh-v1111

MiniMax H3 Ref2VA - Multi-Reference Video Generation

Purpose: a MiniMax H3 reference-to-video graph with the full multi-reference input layout (image, video, paired-video-audio and standalone-audio bays) plus the same two-stage progressive latent-upscale route. Intended for users who can resolve the named custom nodes and will inspect which bays are actually connected or bypassed in this saved state.

Graph

59 nodes / 101 links / 33 node types / 10 groups. Main chain: base UNET -> turbo LoRA -> H3 sigma shift -> euler/simple custom sampler (stage 1) -> SplitSigmas -> 3D latent upscaler 2x -> rebuilt stage-2 reference conditioning -> low-noise refinement -> video + audio decode -> SaveVideo.

What is actually connected in the saved file

• Five reference images are wired to ref_image_0..4; the sixth image bay exists but is unconnected, and the seventh image socket is not wired. All five wired image loaders are active.
• The three reference-video loaders, their three paired audio crops, and the three standalone LoadAudio/AudioCrop pairs are saved with mode=4 (bypass) and empty selections. The layout is present but inactive in this export.
• Global reference duration control: 5 seconds. Global audio start: 0. Ref image size mode: match.

Saved configuration (as in the JSON)

• UNETLoader: minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors (default weight dtype). Note: the graph's own Markdown note names a different base and a 4-step LoRA; the loader nodes are the saved state and are what this listing quotes.
• LoraLoaderModelOnly: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors at strength 1.
• CLIPLoader: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, type minimax. VAEs: minimax_h3_video_vae_fp16.safetensors, minimax_h3_audio_vae_fp32.safetensors.
• MiniMaxH3SigmaShift: 12 / 3. KSamplerSelect: euler. BasicScheduler: simple, 8 steps, denoise 1. SplitSigmas at step 4. H3SigmaRefiner kept but extra_steps 0 (disabled).
• ResolutionSelector: 9:16 portrait, 0.5 MP, multiple of 32. Duration control: 10 seconds (frame budget derived by a 17N+5 expression). Output: SaveVideo prefix video/AIKSK_H3_Ref2VA_RH, CreateVideo at 24 fps.

How to use

Start from the five wired image bays and the prompt, then decide deliberately whether to activate the reference-video and audio bays: they are saved bypassed, so treating them as "already configured" would be wrong. The global reference-duration control drives the audio crops only when the corresponding bays are connected. Keep the distillation-matched sampler settings unless you retune them.

Limits and not-verified items

No inference was run for this listing: no speed, identity-consistency, lip-sync or visual-quality claim is made. The note/loader mismatch documented above is preserved, not silently repaired. Saved run labels and example prompt text are the author's demo content, not evidence of an output. The cover image is a promotional still shipped with the source video, not a frame produced by this graph. Weights and input media are not bundled.

Tutorial: YouTube · Bilibili. The workflows are supplied through AI-KSK; verify workflow, node, model, source-video, image, audio and voice rights separately. No blanket commercial or redistribution license is asserted.