MiniMax H3 Two-Stage HD Video Repaint

RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2089986546912030722?inviteCode=rh-v1111

MiniMax H3 Two-Stage HD Video Repaint

Purpose: a MiniMax H3 upscale/re-render graph that takes an existing clip, injects its latent as the base, applies a learned 3D latent 2x upscale, rebuilds high-resolution conditioning, runs a short low-noise refinement pass, then re-attaches the original audio. Intended for users who can resolve the named custom nodes and will check the source video and audio routing before queueing.

Graph

28 nodes / 45 links / 24 node types / 5 groups, all saved active. Group A: source video input. B: existing model settings. C: source video injection (MiniMaxH3ImageToVideo + H3InjectVideoLatent). D: learned 3D latent 2x upgrade with rebuilt high-res conditioning. E: original audio lock + original sampling parameters + output.

Saved configuration (as in the JSON)

• UNETLoader: minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors (default weight dtype).
• LoraLoaderModelOnly: minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors at strength 1.
• CLIPLoader: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, type minimax.
• VAELoader: video minimax_h3_video_vae_int8_convrot.safetensors, audio minimax_h3_audio_vae_fp32.safetensors.
• MinimaxH3LatentUpscalerNode3D: minimax_h3_latent_upscaler_3d_fp16.safetensors, scale 2, cuda/fp16.
• MiniMaxH3SigmaShift: 12 / 3. KSamplerSelect: euler. BasicScheduler: simple, 4 steps, denoise 0.65. H3SigmaRefiner: extra_steps 1, start_at_sigma 0.7, end 0, cosine.
• Source clip: LoadVideoUI saved at start 0 / end 5 s / 24 fps / maintain-aspect-ratio; base latent canvas 1344x768, length 124 frames.
• High-res conditioning stage: 1344x768 primitives with width/height supplied through math expressions. Output: VHS_VideoCombine, prefix MH3_Universal_Upscale_FL2VA_, h264 mp4, crf 19, 24 fps, original audio routed through MiniMaxH3NativeAudioLock.

How to use

Replace the source video and set the crop window. The high-res re-render prompt is explicit about preserving motion, identity, composition and camera movement; that is a prompt instruction, not a guarantee. Keep the distillation-matched sampler settings unless you deliberately retune them. Check the saved 5-second window and frame cap against your own clip before queueing.

Limits and not-verified items

No inference was run for this listing: no speed, VRAM, upscale-fidelity or audio-sync result is claimed. The saved filename prefix and group names still say FL2VA, which is a naming leftover, not evidence that the graph is a first/last-frame generator. The embedded VHS preview metadata points at a historical cloud run of the author's; it has no runtime effect and is not a credential. 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.