Notes:

• Use 1337 5p34k trigger word. E.g., “woodcut” becomes “w00dcu7”
• Small datasets (15 to 30 images) are effective
  • I used only 7 images of me for a likeness lora trained on FAL as shown here
  • Trigger word was “ohwx man”

Guides:

• Replicate: https://replicate.com/blog/fine-tune-flux-with-faces (pasting some higlights below)
  • Step 1: Gather your training images
    • You can fine-tune Flux with as few as two training images, but for best results you’ll want to use at least 10 images or more. In theory you’ll get continually better results as you include more images in the training data, but the training process can take longer the more images you add.
      • Filenames don’t matter. Name your files whatever you like.
      • Aspect ratio doesn’t matter. Images can be square, landscape, portrait, etc. 10 images is a good minimum.
  • Step 2: Choose a unique trigger word
    • Whenever you fine-tune an image model, you also choose a unique “trigger word” that you’ll use later in your text prompts when generating images: photo of YOUR_TRIGGER_WORD_HERE looking super-cool, riding on a segway scooter
    • Here are some things to consider when choosing a trigger word:
      • It should be something unique like MY_UNIQ_TRGGR. Think “vanity license plates”, but without any length limits.
      • It should not be an existing word in any language, like dog or cyberpunk.
      • It should not be TOK, because it will clash with other fine-tunes if you ever want to combine them.
      • Case doesn’t matter, but capital letters can help visually distinguish the trigger word from the rest of the text prompt.
      • For my zeke/ziki-flux fine-tune, I chose ZIKI as a trigger word. Short, unique, and memorable.

Code:

Simple Tuner: https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md AI Toolkit: https://github.com/ostris/ai-toolkit?tab=readme-ov-file#flux1-training

Online Services for LoRA training:

Fal.ai: https://fal.ai/models/fal-ai/flux-lora-general-training

Also Fal has a fast trainer, $2, 5 minutes: https://fal.ai/models/fal-ai/flux-lora-fast-training

Civitai: https://civitai.com/models/train

Astrai.ai https://www.astria.ai/

Haven’t had a chance to try this one yet. Roope has some examples

https://replicate.com/ostris/flux-dev-lora-trainer/train

Uses Ostris at-toolkit, which is the same as Fal.