PK MiniMax H3 Image

Still-image variant of PK MiniMax H3 Combined, built on the native p-h3-image internals: reference images sized to the generation's pixel area (width/height 0 = match the first reference image, or 2912), a RefMod bundle, and a still-image latent (one video temporal cell + its two audio latents). The model needs ModelSamplingMiniMaxH3 at 1.0/1.0.

By PlagueKind·Created 6 months ago·Updated a day ago· 217

PK MiniMax H3 Image

clip
vae
reference_images
mods

conditioning
latent
reference_map

◄prompt—►

◄width0►

◄height0►

◄reference_fps24.00►

◄max_total_tokens0►

◄vlm_resolution384►

◄vlm_video_resolution384►

CategoryPlagueKind/minimax

Inputs (11)

Name	Type	Default	Description
clip	CLIP		—
prompt	STRING		—
width	INT	00–16384	0 uses the width of the first reference image after scaling, or 2912 without images.
height	INT	00–16384	0 uses the height of the first reference image after scaling, or 2912 without images.
reference_fps	FLOAT	24.001–120	Playback rate assumed when reconstructing a RefMod video reference's timestamps.
max_total_tokens	INT	00–2147483647	Refuse RefMod bundles over this many reference tokens. 0 = no limit.
vlm_resolution	INT	3840–4096	Qwen3-VL presentation resolution for still reference images, independent of the resolution actually VAE-encoded as the DiT reference. 0 (or out of 256-4096) keeps the same resolution as the DiT reference.
vlm_video_resolution	INT	3840–4096	Qwen3-VL presentation resolution for RefMod video reference frames. Higher values use more visual tokens. 0 (or out of 256-4096) keeps original size.
vaeopt	VAE		H3 video VAE. Required for reference images to be VAE-encoded as DiT references; without it they only condition the text encoder.
reference_imagesopt	COMFY_AUTOGROW_V3		—
modsopt	H3_REF_MODS		RefMod bundle from a RefMod Stack, or a Prompt Builder's 'mods' output. Labelled last.

Outputs (3)

Name	Type	Description
conditioning	CONDITIONING	—
latent	LATENT	—
reference_map	STRING	—