@Rey_Spartan RefMods (Reference Latent Adapters) are pre-encoded conditioning-level adapters (.safetensors) for MiniMax-H3. They allow zero-training, instant-load facial likeness, anatomical consistency, and multi-character scaling without the live VAE encoding overhead or prompt syntax constraints of legacy reference image workflows.

In short it is more quality, better character precision (likeness) for Minimax.

@Rey_Spartan basically precomputed reference. GitHub - Luisacaotica/ComfyUI-MiniMaxH3Mod: 🚧 Under construction — no-training RefMod reference adapters for MiniMax H3 (compress reference images/videos into tiny latents, no training needed). · GitHub

@HugMeIntoFace interested

why ? ref mods suck. Ref audio suck. Each image + to generation time.... . Lora works better. You generation time is the same if its not a 1-2gb lora

@gambikules858 You should learn more about it.

@gambikules858 ref mods are super sensitive
the audio needs to be clear
the images need same res (600 by 600 for example)
lora destroys the weights of the model anyway
thats why models like seedance 2.5 do not use loras
they use somthing like refmod for new characters or style
this models are more sensitive to own wights theyre system is not ltx or wan they are way more advance
so loras either look like shit or just destoys the model capability
lora is good for the models that does not have this capability or just you want ai to learn
something new like a new movment that the ai dont know