RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2088992117300482049?inviteCode=rh-v1111
MiniMax H3 Universal Video HD Redraw (Single-Pass Upscale)
Purpose: the simplest route in this set — take any video, scale it up, inject its latent into MiniMax H3, and run one 4-step low-denoise redraw that preserves the original motion, identity, composition and camera movement while restoring detail. No reference images, no multi-pass chain.
Graph (as saved)
26 nodes / 34 links / 6 groups.
- Model loader:
DiffusionModelLoaderKJin this file (there is no UNETLoader node), plus CLIPLoaderqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors(minimax) and VAELoader videominimax_h3_video_vae_int8_convrot.safetensors+ audiominimax_h3_audio_vae_fp32.safetensors. - LoRA:
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensorsat 1. Sigma shift 12 / 3. - Input:
LoadVideoUIsaved at 0–5 s, 24 fps, maintain-aspect-ratio; RTX upscale node (title ×1.5, saved multiplier widget 2.0);MiniMaxH3ImageToVideoconditioning 1344x768 / 124 +H3InjectVideoLatentinjects the source clip latent. - Sampling:
SamplerCustomAdvanced+KSamplerSelecteuler +BasicSchedulersimple 4 steps denoise 0.65;H3SigmaRefiner1 / 0.7 / 0 / cosine;MiniMaxH3NativeAudioLockkeeps the original audio from being regenerated. - Output: VHS_VideoCombine h264 mp4 crf 19, prefix
MH3_Universal_Upscale_FL2VA_.
How to use
Trim the source clip before queueing — the graph's own notes cap H3 length at 362 frames (~15.1 s at 24 fps) and recommend ≤15 s inputs. The default ×1.5 is only appropriate for low-to-mid-resolution sources; for high-resolution input, lower the scale or pre-scale. The low 0.65 denoise is what preserves the composition; raise it and the model starts redrawing instead of restoring.
Limits and not-verified items
No inference was run for this listing: no speed, VRAM, detail-gain or audio-integrity claim is made. H3 weights, LoRA, encoder and VAEs are not bundled. The graph 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.
Description
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2088992117300482049?inviteCode=rh-v1111
MiniMax H3 Universal Video HD Redraw (Single-Pass Upscale)Purpose: the simplest route in this set — take any video, scale it up, inject its latent into MiniMax H3, and run one 4-step low-denoise redraw that preserves the original motion, identity, composition and camera movement while restoring detail. No reference images, no multi-pass chain.
Graph (as saved)26 nodes / 34 links / 6 groups.
- Model loader:
DiffusionModelLoaderKJin this file (there is no UNETLoader node), plus CLIPLoaderqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors(minimax) and VAELoader videominimax_h3_video_vae_int8_convrot.safetensors+ audiominimax_h3_audio_vae_fp32.safetensors. - LoRA:
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensorsat 1. Sigma shift 12 / 3. - Input:
LoadVideoUIsaved at 0–5 s, 24 fps, maintain-aspect-ratio; RTX upscale node (title ×1.5, saved multiplier widget 2.0);MiniMaxH3ImageToVideoconditioning 1344x768 / 124 +H3InjectVideoLatentinjects the source clip latent. - Sampling:
SamplerCustomAdvanced+KSamplerSelecteuler +BasicSchedulersimple 4 steps denoise 0.65;H3SigmaRefiner1 / 0.7 / 0 / cosine;MiniMaxH3NativeAudioLockkeeps the original audio from being regenerated. - Output: VHS_VideoCombine h264 mp4 crf 19, prefix
MH3_Universal_Upscale_FL2VA_.
Trim the source clip before queueing — the graph's own notes cap H3 length at 362 frames (~15.1 s at 24 fps) and recommend ≤15 s inputs. The default ×1.5 is only appropriate for low-to-mid-resolution sources; for high-resolution input, lower the scale or pre-scale. The low 0.65 denoise is what preserves the composition; raise it and the model starts redrawing instead of restoring.
Limits and not-verified itemsNo inference was run for this listing: no speed, VRAM, detail-gain or audio-integrity claim is made. H3 weights, LoRA, encoder and VAEs are not bundled. The graph 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.