Project Overview
- Primary Objective
- To create a LoRA endpoint that can be plugged into our game.
- Experiment Date
- October 14, 2025
Model & Endpoint
- LoRA Endpoint (on Replicate)
-
patrickjohncyh/aibg:7db1be8883adcfdf7481b9d6842883ea3099ca41ac447ce05afdd5c1011ee1b5 - Base Model Architecture
- Fine-tuning was performed on FLUX dev. The resulting LoRA can also be applied to the FLUX schnelll base model.
Dataset Details
The dataset was prepared by Jose and consists of 26 image-text pairs in our target art style.
Critical Pre-processing Step:
In addition to the image description, the text for each image was appended with the trigger word: "In the style of AIBG". This helps the model activate the LoRA during generation.
Training Process
Training was conducted using the Replicate web interface:
- Used the training form at replicate.com/replicate/fast-flux-trainer/train.
- Uploaded a ZIP file containing the image-text pairs.
- Adjusted parameters (see below) and started the training.
Key Hyperparameters:
lora_type: set to"style"training_steps: increased to3000
Outcomes & Observations
Evaluation Method
Informal qualitative analysis ("I eyeballed it!"). No quantitative metrics (e.g., FID, CLIPS) were used for this experiment.
Qualitative Findings
- Works well for "known" concepts like animals.
- Struggles with more abstract creatures or concepts.
Recommendations & Next Steps
- The training dataset's text descriptions can be improved to be more visually descriptive, which would help the model map text to image more effectively.
- During inference (actual use), input prompts should also be more visually specific. This may require prompt rewriting.