AI & Creative Tools

Flux Noto Emoji LoRA

A Flux.1-dev LoRA trained on Google's Noto emoji style, with preprocessing, BLIP captioning, and Replicate-based training scripts.

Live demoRepository
PythonFlux.1LoRABLIPReplicate

This project fine-tunes Flux.1-dev toward Google's open-source Noto emoji visual language. The model uses the trigger word NOTO, making it possible to prompt for objects and characters that do not exist in the original emoji set while keeping the flat, friendly Noto style.

Generated bear wearing a space suit emoji
A generated NOTO-style bear in a space suit.
Generated flying horse emoji
A generated flying horse emoji, outside the original Noto set.

Training pipeline

The repo contains the supporting scripts around the model, not only the final examples:

  • preprocess.py composites every transparent PNG onto a white background so the training images are consistent.
  • generate_img_caption.py runs BLIP over each emoji and writes paired text captions beside the images.
  • train.py starts a Replicate training job using the Flux LoRA trainer.
  • config.yaml captures a local trainer configuration with a NOTO trigger word, LoRA rank 16, 512px training data, flowmatch scheduling, and EMA enabled.

Result

The fun part was watching the LoRA generalise past the dataset. Prompts such as NOTO bear wearing space suit emoji and NOTO flying-horse emoji produced clean images that still felt like they belonged in the Noto family. The exercise gave me a complete pass through the small-model adaptation loop: dataset cleanup, caption generation, trainer configuration, hosted inference, and publishing.

Weights are available on Hugging Face.