# Hyper-Realistic AI Twin Generation Workflow

> Workflow for training and generating realistic personal portraits using FLUX LoRAs

This workflow outlines how to train a personalized FLUX.1 \[dev\] LoRA using Fal.ai's Portrait Trainer and generate authentic-looking self-portraits. It pairs a custom dataset and specific training hyperparameters with phone-camera prompt cues and skin texture LoRAs to avoid typical synthetic AI artifacts.

## How it works

1. Dataset Collection: Gather ~200 clear selfies in diverse lighting and environments with face fully visible. Make ~20% of the dataset full-body photos.
2. Model Training: On Fal.ai using the FLUX LoRA Portrait Trainer (fal-ai/flux-lora-portrait-trainer), configure:
3. Learning Rate: 0.00009
4. Steps: 2500 to 4000 (roughly 1000 steps for 4 images, 3000-4000 for up to 100 images; author used 2500 steps)
5. Multiresolution Training: ON
6. Subject Crop: ON
7. Trigger Word: Choose a unique phrase or token (e.g., 'Sirio' or 'AlexPortrait').
8. Generation: Run FLUX.1 \[dev\] with LoRAs. Add your trained model path (scale 1.0).
9. Prompting: Emulate phone camera capture, e.g.: 'a photo of \[TriggerWord\] man wearing a black hoodie, iPhone photo, IMG\_1080.HEIC'
10. Skin Texture: Stack a skin texture LoRA weighted at 0.7-1.0 to bring out pores and realistic texture.

## What it is for

- Train a custom FLUX LoRA to create consistent personal AI avatars
- Generate authentic-looking casual selfies without standard AI over-smoothing
- Produce realistic stock or personal branding photos using custom prompts

## Key points

- Gather a dataset of ~200 diverse selfies, including ~20% full-body shots with visible faces
- Train on Fal.ai using fal-ai/flux-lora-portrait-trainer with a learning rate of 0.00009 and 2500–4000 steps
- Enable multiresolution training and subject crop during training
- Assign a unique trigger word to reliably call the subject in prompts
- Prompt with camera-like artifacts by referencing phone filenames such as IMG\_1080.HEIC
- Stack a secondary skin texture LoRA weighted between 0.7 and 1.0 for realistic pores and details

## Pricing

Paid. Requires Fal.ai compute credits to run training and inference

## Details

- Type: Tip
- Category: [AI](https://xiry.app/topics/ai.md)
- Topic: [Image Generation](https://xiry.app/topics/ai/image-generation.md)
- Open source: no
- Platforms: Web, API
- For: AI creators, photographers, and digital artists looking for photorealistic personal AI models
- Alternative to: PhotoMaker, InstantID
- Tags: flux, lora, fal ai, ai portraits, prompting, fine tuning

## Why it was saved

- [@heysirio, 26 November 2024](https://www.instagram.com/p/DC175tRv6le/): Produces photorealistic AI selfies that capture authentic facial features and skin texture rather than the typical synthetic AI look.

## Related

- [Krea](https://xiry.app/resources/krea): AI creative suite for real-time image generation, enhancement, and video creation (Web app, Freemium)
- [Recraft](https://xiry.app/resources/recraft): AI design platform generating editable vectors, 3D graphics, and photorealistic images (Web app, Freemium)
- [Adobe Firefly](https://xiry.app/resources/adobe-firefly): Adobe's creative generative AI suite for image, video, and audio creation (Web app, Freemium)
- [Bing Image Creator](https://xiry.app/resources/bing-image-creator): Free AI text-to-image generator powered by DALL-E and Microsoft models (Web app, Free)
- [Ideogram](https://xiry.app/resources/ideogram): AI image generator known for accurate text rendering and graphic design. (Web app, Freemium)
- [Midjourney](https://xiry.app/resources/midjourney): AI model that generates photorealistic and artistic images from text prompts (Web app, Paid)

Updated 3 October 2026. Page: https://xiry.app/resources/hyper-realistic-ai-twin-generation-workflow

Make your own library: Xiry takes your saved reels, tweets, Reddit bookmarks and Watch Later videos, makes them searchable and pulls the useful stuff out, like this library. It's invite-only for now; join the waitlist at https://xiry.app/#waitlist.