Nodes/ComfyUI_StarNodes/⭐ Star Minimax All In One
ComfyUI Node

⭐ Star Minimax All In One

Complete MiniMax H3 reference-to-video pipeline in a single node: loads the diffusion model (or uses the optional MODEL override input), the minimax text encoder and both VAEs, builds <Picture i> / <Video k> / <Audio j> reference conditioning, samples with the chosen sampler/scheduler and decodes video + audio. Reference image/video/audio slots grow automatically, exactly like the core MiniMax H3 Reference to Video node.

By Starnodes2024·Created 2 years ago·Updated 7 days ago· 103
⭐ Star Minimax All In One
  • model_override
  • ref_images
  • ref_videos
  • ref_video_audios
  • ref_audios
  • IMAGE
  • AUDIO
  • FPS
  • LATENT
  • MODEL
  • CLIP
  • VAE
  • AUDIO_VAE
modevideo
prompt
aspect_ratio16:9 (Widescreen)
megapixels0.5
match_ratio_from_imagefalse
duration5.0
ref_image_sizematch
seed0
steps20
sampler_nameres_multistep
schedulersimple
denoise1.00
diffusion_modelminimax_h3_ref2va_pruned_int8_convrot.safetensors
weight_dtypedefault
clip_nameqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
clip_typeminimax
clip_devicedefault
vae_nameminimax_h3_video_vae_fp16.safetensors
audio_vae_nameminimax_h3_audio_vae_fp32.safetensors
audio_vae_precisionfp32
audio_vae_devicemain_device
Category⭐StarNodes/Video

Inputs (26)

NameTypeDefaultDescription
modeCOMBOvideo'video' renders the full clip with audio. 'image' renders 9 frames at the selected ratio and size and outputs only frame index 8 as a still image (best-quality frame); duration and audio decoding are skipped. Reference inputs work exactly like in video mode.
promptSTRINGReference inputs by tag in connection order, e.g. <Picture 1>, <Video 1>, <Audio 1>, then describe scene, motion and audio.
aspect_ratioCOMBO16:9 (Widescreen)Aspect ratio for the output dimensions.
megapixelsCOMBO0.5Target total megapixels (output pixel size). 0.5 MP ~ 960x544 at 16:9; 2.0 MP ~ 1920x1088. Select "audio only" for a fixed 32x32 canvas when you only need audio output.
match_ratio_from_imageBOOLEANfalseIf enabled and a reference image is connected, the closest aspect ratio of the first reference image is used at the selected pixel size.
durationFLOAT5.00.2–150Video duration in seconds at 24 fps. Internally snapped to the model's 17k+5 frame grid (5s -> 124 frames).
ref_image_sizeCOMBOmatchReference image sizing. 'match' scales each ref (down only, keeping aspect) to the generation's pixel area; 'max' uses a 2048px short edge for best identity fidelity but is much slower.
seedINT00–18446744073709550000
stepsINT201–10000
sampler_nameCOMBOres_multistep44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38
schedulerCOMBOsimple'beta' or 'normal' tends to outperform 'simple' for reference-heavy prompts.
denoiseFLOAT1.000–1
diffusion_modelCOMBOminimax_h3_ref2va_pruned_int8_convrot.safetensorsIgnored when a MODEL is connected to the model_override input.
weight_dtypeCOMBOdefault4 options: default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2
clip_nameCOMBOqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors0 options:
clip_typeCOMBOminimax28 options: stable_diffusion, stable_cascade, sd3, stable_audio, mochi, ltxv, +22
clip_deviceCOMBOdefault2 options: default, cpu
vae_nameCOMBOminimax_h3_video_vae_fp16.safetensorsVideo VAE (e.g. minimax_h3_video_vae_fp16).
audio_vae_nameCOMBOminimax_h3_audio_vae_fp32.safetensorsAudio VAE (e.g. minimax_h3_audio_vae_fp32).
audio_vae_precisionCOMBOfp32Precision the audio VAE runs at. fp32 is recommended (same as the KJ loader preset).
audio_vae_deviceCOMBOmain_device2 options: main_device, cpu
model_overrideoptMODELOptional external MODEL (e.g. a sage-attention patched MiniMax H3). When connected, the internal diffusion_model dropdown is ignored.
ref_imagesoptCOMFY_AUTOGROW_V3
ref_videosoptCOMFY_AUTOGROW_V3
ref_video_audiosoptCOMFY_AUTOGROW_V3
ref_audiosoptCOMFY_AUTOGROW_V3

Outputs (8)

NameTypeDescription
IMAGEIMAGE
AUDIOAUDIO
FPSFLOATFixed frame rate of the generated video (24.0). Connect straight into your video combine/save node.
LATENTLATENTThe combined processed latent from the sampler (NestedTensor with video+audio), before VAE decoding.
MODELMODELThe diffusion model used for sampling.
CLIPCLIPThe loaded text encoder.
VAEVAEThe video VAE.
AUDIO_VAEVAEThe audio VAE (None in image mode without audio references).