MiniMax H3 FL2VA Two-Stage Upscale (First/Last Frame)

RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2089240886684250114?inviteCode=rh-v1111

MiniMax H3 FL2VA Two-Stage Upscale (First/Last Frame)

Purpose: the two-pass variant of the FL2VA upscale. First and last frame images condition PASS 1; a pixel-space RTX upscale then carries the result into PASS 2, which re-conditions at the actual upscaled size and runs the 4-step turbo LoRA at low denoise.

Graph (as saved)

59 nodes / 68 links / 5 groups.

• UNETLoader minimax_h3_fl2va_bf16.safetensors; 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: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors at 1 (PASS 1) and minimax_h3_fl2v_turbo_4step_v1_768p_comfyui_bf16.safetensors at 1 (PASS 2).
• Inputs: first-frame and last-frame LoadImage nodes, MiniMaxH3ImageToVideo (1344x768 / 124 saved), duration 10 s; ResolutionSelector is saved bypassed (21:9 / 0.7 MP).
• PASS 1: SamplerCustomAdvanced + KSamplerSelect res_multistep + BasicScheduler beta 8 steps denoise 1; shift 6 / 3.
• Upscale: Yuan_RTXVideoUpscaleH3 (node title ×1.5, saved multiplier widget 2.0 — check it before queueing), then PASS 2 re-conditioning at the actual 1.5× size.
• PASS 2: KSamplerSelect er_sde + BasicScheduler simple 4 steps denoise 0.65; shift 12 / 3; H3SigmaRefiner 1 / 0.7 / 0 / cosine. Stale title: one PASS 2 scheduler node is titled "8 steps / denoise 0.30" while its saved widgets are 4 steps / 0.65 — quote the widgets.
• Output prefix MH3x2Smpl_ (plus MH3_Preview and a bypassed MH3-2Pass_ saver).

How to use

Replace the first and last frames and the prompt, keep the audio from PASS 1 for the final export, and confirm the RTX multiplier matches the resolution you planned. The two-stage route is the cheaper option when the three-stage graph's last pass is not buying you visible detail.

Limits and not-verified items

No inference was run for this listing: no speed, VRAM or quality 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.