Questions on Building a Flux LoRA for Synthetic Illustration Style
Decide if a single‑resolution dataset suffices or if you should add varied aspect ratios for your style LoRA.
Plan your dataset resolution strategy and decide on a training workflow for style and character LoRAs.
Summary
The author aims to create a style LoRA for a corporate‑memphis‑like illustration style using Flux 2 Dev Image Edit workflow. They generate a dataset from reference images and ask whether a single resolution (512×512) suffices or if multiple aspect ratios should be included to maintain consistency across varied outputs.
They also inquire about recurring characters: whether to train a style LoRA first, generate a character dataset, then train separate character LoRAs, and how combining LoRAs might cause conflicts. Guidance is sought for scenes with multiple mascots interacting and for a comprehensive guide on synthetic style LoRA training.
No recent bible exists for this specific use case; the author requests resources on clean, flat vector style illustrations and advice on training strategies.
Key changes
- Goal: create a style LoRA for a corporate‑memphis‑like illustration style.
- Using Flux 2 Dev Image Edit workflow to generate dataset from reference images.
- Question 1: single‑resolution (512×512) vs multi‑resolution; concern about consistency across aspect ratios.
- Question 2: recurring characters; typical approach: train style LoRA first, generate character dataset, train separate character LoRA per character.
- Combining LoRAs can cause conflicts; lower weights may mitigate.
- Need guidance on generating scenes with multiple mascots interacting.
- No recent comprehensive guide for synthetic style LoRA training.
- Seeking tips for clean, flat vector style illustrations.