Lore

Flux Personal LoRA Training via Replicate

Overview

Training a personal Flux LoRA (a small fine-tuned add-on that teaches the base Flux model a specific person's likeness) can be done directly on Replicate.com using the 'AI toolkit for flux Lora training' model (by Replicate contributor Luca Taco), replacing older Dream Booth/Stable Diffusion 1.4 workflows run in Google Colab.

Method

• Minimum 12 training images (20 used in practice).
• Requires a Hugging Face account with a fine-grain access token; set HF_REPO_ID to auto-upload the trained LoRA to a public Hugging Face repo — Replicate can't reach a private one.
• Training rents an Nvidia A100 at roughly $5/hr (~$0.001/sec); one reported run took ~26 min and cost ~$2.18. Treat as illustrative, not a guaranteed price — pricing and vendor terms drift.
• Inference afterward runs a few cents per image; default settings are 28 inference steps, 1.0 LoRA scale, 1-4 outputs, configurable aspect ratio.

Prompting tip

Placing the trained trigger word as the very first word of a prompt appears to produce more reliable inclusion of the likeness than embedding it mid-prompt — an informally observed, unconfirmed effect, not documented model behavior.

Related

A sibling technique to Flux Turbo Lora (8-step workflow) (speed-oriented LoRA workflow) and Flux Realism LoRA & Guidance-Scale Tuning for Photorealism (photorealism tuning) — this one covers training your own likeness LoRA rather than tuning an existing one. A finished image/prompt pair can be animated by feeding it into Runway Gen 3 as the first frame.

Из тем: Flux & ComfyUI: Local Model Tuning