Had a great question come up from about creating consistent characters. Response is probably helpful for all!
IMO Midjourney is still far and away the best tool for this.
You can do both consistency in style and character.
—sref is the style consistency command
—cref is the character consistency command.
Here’s a quick example of a cref I did.
More info here.
Your other option is to use something like flux and fine tune a LoRA on your character. This is what I did to train flux on my face.
Instructions
To train a Flux LoRA model on yourself using Replicate, follow these steps:
1. Prepare your training images:
- Gather 10-20 high-quality images of yourself[3].
- Ensure the images show your face clearly with some variety in expressions and angles[3].
- Name your image files using the format: "photo_of_[trigger_word]_[number].jpg" (e.g., "photo_of_john_1.jpg")[4].
- Compress all images into a single ZIP file[4].
2. Set up the training process on Replicate:
- For the "destination" input, choose an existing model or create a new one[3].
- For "input_images", upload your ZIP file containing the training images[3].
- Set your "trigger_word" (e.g., "john")[3].
- Leave "steps" at the default 1000[3].
- Keep other inputs at their default values[3].
3. Start the training:
- Click "Create training" to begin the process[3].
- The training will take approximately 20 minutes for 10 images and 1000 steps[3].
4. Use your trained model:
- Once training is complete, your model will be ready to use[3].
- You can generate images directly on Replicate's web interface or use the API[3].
- When creating prompts, include your trigger word to activate your trained concept[3].
Some additional tips:
- The training process typically costs around $2-3 on Replicate[1][6].
- You can fine-tune with as few as 2 images, but 10 or more is recommended for better results[3].
- Ensure your trigger word is unique to avoid conflicts with existing concepts[3].
- For best results, use AI-generated images rather than real photos for training[1].
By following these steps, you should be able to create a Flux LoRA model trained on your own images using Replicate's platform.