MiniMax H3 Ref2VA Three-Stage 2K Multi-Reference Redraw

RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2089240711672713217?inviteCode=rh-v1111

MiniMax H3 Ref2VA Three-Stage 2K Multi-Reference Redraw

Purpose: a three-pass Ref2VA redraw. Multi-reference conditioning drives PASS 1; two pixel-space RTX upscales carry the video through PASS 2 and PASS 3, each re-conditioned at the actual new size with the 4-step turbo LoRA at low denoise. This is the multi-reference route rather than the first/last-frame route.

Graph (as saved)

75 nodes / 109 links / 6 groups.

• Model path: MiniMaxH3HybridLoader titled "H25 + LightX2V 8" (block range adaln 25-49) is the active loader; the plain UNETLoader on minimax_h3_fl2va_bf16.safetensors is saved bypassed.
• CLIPLoader qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (minimax); VAELoader video minimax_h3_video_vae_int8_convrot.safetensors + audio minimax_h3_audio_vae_fp32.safetensors.
• LoRAs: fl2v turbo 8-step at 1, ref2v turbo 4-step at 1, and minimax_h3_ref_lora_rank_256_bf16.safetensors at 0.75 saved bypassed.
• PASS 1: SamplerCustomAdvanced + res_multistep + BasicScheduler beta 8 steps denoise 1; sigma shift 6 / 3. PASS 2/3: 4 steps denoise 0.65 (simple); shift 12 / 3; H3SigmaRefiner 1 / 0.7 / 0 / cosine.
• Upscales: two Yuan_RTXVideoUpscaleH3 nodes ×1.5 (PASS 1→2 and PASS 2→3).
• References: five LoadImage slots (one main identity/composition anchor, the rest optional attribute references); three MiniMaxH3ReferenceToVideo conditioners — one per pass — each saved 1344x768 / 124 / max.
• Duration 15 s; ResolutionSelector 16:9 at 0.3 MP / 64. Output prefix MH3_Ref2VA_3Pass_ (plus MH3x2Smpl_ / MH3_Preview taps and a bypassed MH3-2Pass_ saver).

How to use

Picture 1 is the main identity/composition anchor; pictures 2–5 are optional attribute references. Per the saved note, do not connect a video companion audio unless the matching reference video exists. Keep PASS 1 audio for the final export. Three passes at ×1.5 each is a heavy 15-second job; the notes themselves warn about VRAM at 1.5–2 MP for long clips.

Limits and not-verified items

No inference was run for this listing: no speed, VRAM, 2K-quality or reference-consistency claim is made. H3 weights, LoRAs, encoder and VAEs are not bundled. Source notes carry third-party attribution (MarkDKBerry.com); preserved as-is, no originality claim. The cover image is a promotional still, not a frame produced by this graph.

Tutorial: YouTube · Bilibili. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA and cover rights separately. No blanket commercial or redistribution license is asserted.