RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2088625693134512130?inviteCode=rh-v1111
MiniMax H3 FL2VA Three-Stage 2K Upscale
Purpose: a three-pass FL2VA upscale for MiniMax H3 video. PASS 1 generates at the planned stage-1 resolution with the turbo 8-step LoRA; a pixel-space RTX upscale then doubles the working size twice, with PASS 2 and PASS 3 re-conditioning at each new size and running the 4-step LoRA at low denoise.
Graph (as saved)
72 nodes / 98 links / 6 groups.
- UNETLoader
minimax_h3_fl2va_bf16.safetensors; CLIPLoaderqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors(type minimax); VAELoader videominimax_h3_video_vae_int8_convrot.safetensors+ audiominimax_h3_audio_vae_fp32.safetensors. - LoRAs:
minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensorsat 1 for PASS 1;minimax_h3_fl2v_turbo_4step_v1_768p_comfyui_bf16.safetensorsat 1 for PASS 2/3. - PASS 1:
SamplerCustomAdvanced+KSamplerSelectres_multistep +BasicSchedulerbeta 8 steps denoise 1; sigma shift 6 / 3 (dynamic-first). - Upscale:
Yuan_RTXVideoUpscaleH3×1.5 after PASS 1 and again after PASS 2. - PASS 2 / PASS 3:
SamplerCustomAdvanced+KSamplerSelecter_sde and res_multistep (saved per pass) +BasicSchedulersimple 4 steps denoise 0.65; sigma shift 12 / 3;MiniMaxH3ImageToVideore-conditioned at the actual size of each stage;H3SigmaRefiner1 / 0.7 / 0 / cosine. - Duration control 10 s; ResolutionSelector 9:16 portrait at 0.3 MP / 64 multiple. Output prefix
MH3_FL2VA_3Pass_(plus MH3x2Smpl_ and MH3_Preview taps).
How to use
The graph's own notes are explicit that the first pass matters most for motion and general look, and that the audio from PASS 1 should be kept for the final export. PASS 2/3 re-condition at the real upscaled size; the graph includes patchify/shape fixes so the stage sizes stay divisible.
Limits and not-verified items
No inference was run for this listing: no speed, VRAM, 2K-result or temporal-quality claim is made. The H3 weights, LoRAs, CLIP and VAEs are not bundled; check their licenses separately. 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.
Description
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2088625693134512130?inviteCode=rh-v1111
MiniMax H3 FL2VA Three-Stage 2K UpscalePurpose: a three-pass FL2VA upscale for MiniMax H3 video. PASS 1 generates at the planned stage-1 resolution with the turbo 8-step LoRA; a pixel-space RTX upscale then doubles the working size twice, with PASS 2 and PASS 3 re-conditioning at each new size and running the 4-step LoRA at low denoise.
Graph (as saved)72 nodes / 98 links / 6 groups.
- UNETLoader
minimax_h3_fl2va_bf16.safetensors; CLIPLoaderqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors(type minimax); VAELoader videominimax_h3_video_vae_int8_convrot.safetensors+ audiominimax_h3_audio_vae_fp32.safetensors. - LoRAs:
minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensorsat 1 for PASS 1;minimax_h3_fl2v_turbo_4step_v1_768p_comfyui_bf16.safetensorsat 1 for PASS 2/3. - PASS 1:
SamplerCustomAdvanced+KSamplerSelectres_multistep +BasicSchedulerbeta 8 steps denoise 1; sigma shift 6 / 3 (dynamic-first). - Upscale:
Yuan_RTXVideoUpscaleH3×1.5 after PASS 1 and again after PASS 2. - PASS 2 / PASS 3:
SamplerCustomAdvanced+KSamplerSelecter_sde and res_multistep (saved per pass) +BasicSchedulersimple 4 steps denoise 0.65; sigma shift 12 / 3;MiniMaxH3ImageToVideore-conditioned at the actual size of each stage;H3SigmaRefiner1 / 0.7 / 0 / cosine. - Duration control 10 s; ResolutionSelector 9:16 portrait at 0.3 MP / 64 multiple. Output prefix
MH3_FL2VA_3Pass_(plus MH3x2Smpl_ and MH3_Preview taps).
The graph's own notes are explicit that the first pass matters most for motion and general look, and that the audio from PASS 1 should be kept for the final export. PASS 2/3 re-condition at the real upscaled size; the graph includes patchify/shape fixes so the stage sizes stay divisible.
Limits and not-verified itemsNo inference was run for this listing: no speed, VRAM, 2K-result or temporal-quality claim is made. The H3 weights, LoRAs, CLIP and VAEs are not bundled; check their licenses separately. 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.