MiniMax H3 Ref2VA Two-Stage Multi-Reference Redraw

RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2088888730617729025?inviteCode=rh-v1111

MiniMax H3 Ref2VA Two-Stage Multi-Reference Redraw

Purpose: the two-pass multi-reference redraw. Multi-reference conditioning runs PASS 1; one pixel-space RTX upscale carries the video into PASS 2, which is re-conditioned at the actual upscaled size with the 4-step turbo LoRA at low denoise. The saved note calls this the V2 fix: PASS 2 no longer inherits PASS 1's old conditioning size.

Graph (as saved)

63 nodes / 80 links / 5 groups.

• Model path: MiniMaxH3HybridLoader "H25 + LightX2V 8" active, UNETLoader minimax_h3_fl2va_bf16.safetensors also active in this saved state.
• 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, ref LoRA rank 256 at 0.75.
• PASS 1: SamplerCustomAdvanced + res_multistep + BasicScheduler beta 8 steps denoise 1; sigma shift 6 / 3. PASS 2: 4 steps denoise 0.65 (simple); shift 12 / 3; H3SigmaRefiner 1 / 0.7 / 0 / cosine.
• Upscale: one Yuan_RTXVideoUpscaleH3 (title ×1.5, saved multiplier widget 2.0 — check before queueing).
• References: six LoadImage slots; two MiniMaxH3ReferenceToVideo conditioners (one per pass), each saved 1344x768 / 124 / max.
• Duration 15 s; ResolutionSelector 16:9 at 0.4 MP / 64. Output prefixes MH3x2Smpl_ and MH3_Preview.

How to use

This is the cheaper multi-reference option — one upscale instead of two. Check the RTX multiplier and the PASS 2 conditioning recompute, then keep PASS 1 audio for the export. If you need more detail than this produces, move to the three-stage file rather than tweaking denoise blindly.

Limits and not-verified items

No inference was run for this listing: no speed, VRAM or reference-consistency claim is made; values are quoted as saved. 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.