docs/inferene/inference_doc.md

Inference Arguments Guide

This guide describes the command-line arguments for the infer.py script, which performs Stable Diffusion inference using either full fine-tuning or LoRA methods.

Arguments

Argument	Type	Default Value	Description
--method	String	Required (choices: full_finetuning, lora)	Specifies the inference method: full_finetuning or lora.
--prompt	String	"donald trump in ghibli style"	Text prompt for image generation.
--height	Integer	512	Height of the output image (must be divisible by 8).
--width	Integer	512	Width of the output image (must be divisible by 8).
--num_inference_steps	Integer	50	Number of inference steps for the generation process.
--guidance_scale	Float	3.5	Guidance scale for classifier-free guidance.
--batch_size	Integer	1	Batch size for inference.
--seed	Integer	42	Random seed for reproducibility.
--lora_scale	Float	1.2	Scaling factor for LoRA weights (applicable when using LoRA method).
--config_path	String	"configs/model_ckpts.yaml"	Path to the model configuration YAML file.
--output_path	String	"test_data/ghibli_style_{method}_output.png"	Path to save the output image.

Usage Example

• To run inference with full fine-tuning:

python infer.py --method full_finetuning --prompt "donald trump in ghibli style" --height 512 --width 512 --num_inference_steps 50 --guidance_scale 3.5 --batch_size 1 --seed 42 --config_path configs/model_ckpts.yaml

The output_path is tests/test_data/ghibli_style_output_full_finetuning.png.

• To run inference with LoRA:

python infer.py --method lora --prompt "donald trump in ghibli style" --height 512 --width 512 --num_inference_steps 50 --guidance_scale 3.5 --batch_size 1 --seed 42 --lora_scale 1.2 --config_path configs/model_ckpts.yaml

The output_path is tests/test_data/ghibli_style_output_lora.png.

Notes

• The output image is saved as test_data/ghibli_style_{method}_output.png, where {method} is either full_finetuning or lora.
• Ensure the --height and --width values are divisible by 8 to avoid errors in the Stable Diffusion pipeline.
• The --lora_scale argument is only relevant when using the LoRA method.
• The script uses different model IDs based on the method:
  • Full fine-tuning: danhtran2mind/Ghibli-Stable-Diffusion-2.1-Base-finetuning
  • LoRA: danhtran2mind/Ghibli-Stable-Diffusion-2.1-LoRA