J-Rouge: photoreal feature performer for MiniMax-H3 Ref2V. Face locked, motion alive.

Trained as a Ref2V identity adapter on MiniMax-H3 (not SDXL, not Hunyuan). Built for slow, continuous motion: the neon stage strut, bedroom window light, an intimate camera. No frozen poses and no stiff AI smile.

Settings that work: trigger j-rogue, LoRA strength 0.70 (0.80 and above overbakes the face), H3 Turbo LoRA loaded first at 1.0, a clean face still as Image 1, the video VAE on the face encode (never the audio VAE), and sage attention bypassed on an RTX 5090.

Pairs with the official MiniMax H3 4-step Turbo LoRA. Part of Project PXE Boot, a Dallas Design District club built on AI models.