RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2090027189914918914?inviteCode=rh-v1111
MiniMax H3 FL2VA - Text and First/Last Frame Video
Purpose: a MiniMax H3 audio-video graph that keeps first-frame and last-frame inputs in one file and runs a two-stage (progressive latent upscale) route. Intended for users who can resolve the named custom nodes and assets and who will inspect the actual connections before queueing. This is a workflow resource, not a tested model release.
Graph
39 nodes / 51 links / 28 node types / 7 groups, all saved active. Main chain: base UNET -> turbo LoRA -> H3 sigma shift -> euler/simple custom sampler (stage 1) -> SplitSigmas -> 3D latent upscaler 2x -> rebuilt stage-2 conditioning -> low-noise refinement sampler -> video and audio decode -> SaveVideo.
Saved configuration (as in the JSON)
- UNETLoader:
minimax_h3_fl2va_pruned_int8_convrot.safetensors(default weight dtype). - LoraLoaderModelOnly:
minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensorsat strength 1. - CLIPLoader:
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, type minimax. - VAELoader:
minimax_h3_video_vae_fp16.safetensorsandminimax_h3_audio_vae_fp32.safetensors. - MinimaxH3LatentUpscalerNode3D:
minimax_h3_latent_upscaler_3d_fp16.safetensors, scale 2, cuda/fp16. - MiniMaxH3SigmaShift: video 12 / audio 3. KSamplerSelect: euler. BasicScheduler: simple, 8 steps, denoise 1. SplitSigmas at step 4. H3SigmaRefiner kept but extra_steps 0 (disabled).
- ResolutionSelector: 9:16 portrait, 0.5 MP, multiple of 32. With the saved size table that is a 544x960 stage-1 canvas and a 1088x1920 stage-2 canvas; the companion 16:9 table stores up to 1920x1088 at 2.0 MP.
- Duration control: 5 seconds. Output: SaveVideo prefix
video/AIKSK_H3_FL2VA_RH, CreateVideo at 24 fps.
How to use
Both frame inputs are optional in the underlying node: disconnect both for text-to-video, keep one for first-frame or last-frame image-to-video, keep both for FL2VA. Replace the prompt and the two images, and keep the distillation-matched sampler settings unless you deliberately retune them. Aspect and stage-1 resolution come from one selector; stage-2 size is derived from it.
Limits and not-verified items
No inference was run for this listing: no speed, step count, VRAM figure, or visual-quality claim is made here. Asset filenames are quoted exactly as saved and must be resolved in your own installation. The cover image is a promotional still shipped with the source video, not a frame produced by this graph. Weights, model files and input media are not bundled. The graph's own notes describe intended acceleration behaviour; treat them as author notes, not as measured results.
Tutorial: YouTube · Bilibili. The workflows are supplied through AI-KSK; verify workflow, node, model, source-video, image and voice rights separately. No blanket commercial or redistribution license is asserted.
Description
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2090027189914918914?inviteCode=rh-v1111
MiniMax H3 FL2VA - Text and First/Last Frame VideoPurpose: a MiniMax H3 audio-video graph that keeps first-frame and last-frame inputs in one file and runs a two-stage (progressive latent upscale) route. Intended for users who can resolve the named custom nodes and assets and who will inspect the actual connections before queueing. This is a workflow resource, not a tested model release.
Graph39 nodes / 51 links / 28 node types / 7 groups, all saved active. Main chain: base UNET -> turbo LoRA -> H3 sigma shift -> euler/simple custom sampler (stage 1) -> SplitSigmas -> 3D latent upscaler 2x -> rebuilt stage-2 conditioning -> low-noise refinement sampler -> video and audio decode -> SaveVideo.
Saved configuration (as in the JSON)- UNETLoader:
minimax_h3_fl2va_pruned_int8_convrot.safetensors(default weight dtype). - LoraLoaderModelOnly:
minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensorsat strength 1. - CLIPLoader:
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, type minimax. - VAELoader:
minimax_h3_video_vae_fp16.safetensorsandminimax_h3_audio_vae_fp32.safetensors. - MinimaxH3LatentUpscalerNode3D:
minimax_h3_latent_upscaler_3d_fp16.safetensors, scale 2, cuda/fp16. - MiniMaxH3SigmaShift: video 12 / audio 3. KSamplerSelect: euler. BasicScheduler: simple, 8 steps, denoise 1. SplitSigmas at step 4. H3SigmaRefiner kept but extra_steps 0 (disabled).
- ResolutionSelector: 9:16 portrait, 0.5 MP, multiple of 32. With the saved size table that is a 544x960 stage-1 canvas and a 1088x1920 stage-2 canvas; the companion 16:9 table stores up to 1920x1088 at 2.0 MP.
- Duration control: 5 seconds. Output: SaveVideo prefix
video/AIKSK_H3_FL2VA_RH, CreateVideo at 24 fps.
Both frame inputs are optional in the underlying node: disconnect both for text-to-video, keep one for first-frame or last-frame image-to-video, keep both for FL2VA. Replace the prompt and the two images, and keep the distillation-matched sampler settings unless you deliberately retune them. Aspect and stage-1 resolution come from one selector; stage-2 size is derived from it.
Limits and not-verified itemsNo inference was run for this listing: no speed, step count, VRAM figure, or visual-quality claim is made here. Asset filenames are quoted exactly as saved and must be resolved in your own installation. The cover image is a promotional still shipped with the source video, not a frame produced by this graph. Weights, model files and input media are not bundled. The graph's own notes describe intended acceleration behaviour; treat them as author notes, not as measured results.
Tutorial: YouTube · Bilibili. The workflows are supplied through AI-KSK; verify workflow, node, model, source-video, image and voice rights separately. No blanket commercial or redistribution license is asserted.