Load H3 RefLoRA (LoRA + RefMod)

Load a RefLoRA container: LoRA weights onto MODEL/CLIP and the embedded reference(s) out as H3_REF_MODS.

By malcolmamal·Created about 14 hours ago·Updated about 14 hours ago· 0

◄reflora_name▾►

◄lora_strength1.00►

◄refmod_retention1.00►

◄membersall►

◄attach_to_modelfalse►

◄lora_strength_clip1.00►

◄componentsAll►

◄max_total_tokens0►

◄show_infotrue►

CategoryMiniMax-H3/reflora

Inputs (11)

Name	Type	Default	Description
model	MODEL		The model the LoRA weights patch.
reflora_name	COMBO		A RefLoRA container in models/loras/. An ordinary LoRA also works -- it simply has no references to load.
lora_strength	FLOAT	1.00-10–10	LoRA weight strength on the model. 0 skips patching entirely.
refmod_retention	FLOAT	1.000–1	Reference strength. 1.0 = fully preserved, 0.7 = partially preserved, 0.4 = attribute transfer, 0.15 = weak. 0 loads no references at all.
members	STRING	all	Which references to load, numbered as the Inspect node shows them: 'all', 'none', or a list like '1' or '1,3-4'. Unselected members are never read off disk.
attach_to_model	BOOLEAN	false	ON: the references ride the MODEL output, so nothing has to be wired into the conditioning -- one node, one wire. OFF (default): use the H3_REF_MODS output with an Apply node. Do not do both, or every reference is injected twice.
clipopt	CLIP		Optional. MiniMax H3 LoRAs are lora_unet_* only, so this is usually left unconnected.
lora_strength_clipopt	FLOAT	1.00-10–10	—
componentsopt	COMBO	All	Filter references by modality before loading them.
max_total_tokensopt	INT	00–1048576	Reference token budget; 0 disables it. References are not free -- a full-resolution identity reference can cost several thousand tokens in every DiT block.
show_infoopt	BOOLEAN	true	Print the container's contents to the console on load.