Turning ReActor face embeddings into LoRAs: a hands-on Krea2 training workflow

PropagandaOfTheDude · reddit · 2026-09-03

A Redditor shares a working pipeline for converting ReActor face embeddings into LoRAs trained on Krea2. Key points:

• Embeddings are math-friendly: ReActor encodes faces as vectors; averaging two people's embeddings element-wise yields an in-between face. More flexible than LoRAs/edit models, but re-rendering breaks on extreme head angles and only works on photographic images.
• Synthetic training data: training images (varied distances, outfits, hairstyles, head poses) are generated with ReActor, and the generation prompts double as training captions — accurate by definition. The tilted-head pose still requires manual loops of 10 generations to pick one with sane eyes.
• Hidden prompts to fix the face substrate: since re-rendering never fully overwrites the base face (eye mismatches, seams, skin-tone drift), the author appends a fixed block describing a consistent, photogenic subject (age, ethnicity, facial structure, makeup) so the embedding has clean features to fit onto.
• Keep three debug images: original prompt, prompt+traits block, and post-swap.

The post is truncated mid-way, but the shared portion already gives a reproducible method skeleton.

Original post →

More from Multimodal

• One-Shot Fable 5.1 Remakes a Call of Duty Demo, Fueling 'Software Priced at Electricity Cost' Take — nptacek · 2026-09-03
• Krea Launches Beta of Its 'World's Best Agents for Creative Work' — angrypenguinPNG · 2026-09-03
• MiniMax H3 video generation shows blur artifacts on fast motion, asks Reddit — MixZealousideal9359 · 2026-09-03
• DreamX-Creator 1.0: new video+sound model built on Wan 2.2 5B — cc_aa_tt_zz · 2026-09-03
• Letting an AI agent direct: one concept prompt yields a full retro pop-art short film — SimplyAnnisa · 2026-09-03
• Fine-tuned SDXL on 60 childhood photos to simulate memory recall — Chuka444 · 2026-09-03

Multimodal channel →