gnicartweet/J-Rogue-H3-ref2va

J-Rogue - MiniMax H3 Ref2VA LoRA (v4)

TRIGGER WORD: j-rogue

Use exactly j-rogue (lowercase, with hyphen) at the start of every prompt. This exact string was used in every training caption. The model was trained under the internal dataset name jamie - same character, just an old nickname.

A reference-to-video LoRA for MiniMax H3 generating a consistent fictional character: J-Rogue, a 29-year-old feature dancer at PXE Boot (fictional; not based on any real person).

Character lock (quick reference)

• Soft honey-blonde chin-to-shoulder bob (not pixie, not past shoulders); warm brown eyes
• Fair skin with warm undertone; voluptuous hourglass build; always clearly an adult (29)
• Default mood: soft knowing smile, calm confidence
• Palette: white / black / gold
• Six fixed tattoos: anchor (sternum), camera (right shoulder), Celtic knot (left upper back), sun circle (right upper back), alien head (left forearm), mandala (right forearm)

Training details

Base model	Comfy-Org/MiniMax-H3, arch minimax_h3_ref2va
Trainer	Ostris AI Toolkit (RunComfy), step 500 of 500 - final checkpoint
LoRA rank / alpha	16 / 16, BF16, 208 modules (50 DiT blocks + 2 token-refiner blocks)
LR / optimizer	1e-4, AdamW8bit, flowmatch, guidance loss on
Dataset	12 clips + 10 reference images (ref2v smoke set), 39-frame grid @ 768p
Assistant LoRA	ostris/minimax_h3_training_adapter (ref2va v1)
Text encoder	qwen3vl_32b, nvfp4 quantized

Recommended usage (ComfyUI)

1. Load the MiniMax H3: Reference to Video workflow (official template)
2. Add this LoRA via a LoRA Loader between the model loader and the sampler - strength 1.0
3. Prompt example: j-rogue, walking through the club entrance, confident stride, purple club lighting
4. Resolution ~768p, 24fps, clip lengths on the 22/39/56/73/90/107/124-frame grid
5. Feed one or more reference images of J-Rogue through the reference branch

Related models by the same author

• gnicartweet/J-Rogue_v2 - SD-family stills LoRA (trigger: jrogue_pxe)
• gnicartweet/j-rogue-merged-v1 - Flux/SD merge (NOT for H3)

Version history

• 2026-10-12 (2): sample 4 added - first OpenH3-IR pipeline clip (j-rogue, lip-synced signature line "You will get used to it. Everybody does.").
• v4 (this file, step 500): ref2va smoke-test training run.