likewendy/pandora-edit

This is a lora from Qwen-Image-Edit-2509.

Like Pandora, it's part of a series.

However, it uses significantly fewer images than Pandora, which used over 6,000 images, including 5,000 low-quality and 1,000 high-quality images.

It still employs a two-stage training method: coarse training and fine training. Coarse training used 233 images of slightly lower quality in terms of clarity and artistic appeal, with a total file size of 592.4 MB including the final image and reference images.

Fine training used 141 images, with a total file size of 1.8 GB including the final image and reference images. These images are of high quality in terms of artistic appeal, resolution, and clarity, possessing studio-level image quality.

When building the dataset, some reference images were generated using AI, and some prompts were manually edited.

Checkpoints: Coarse training had three checkpoints: epochs 1, 2, and 3.

Fine training had four checkpoints: epochs 0, 1, 2, and 3.

The coarse training repeated each sample 6 times, and the fine-tuning repeated each sample 10 times, meaning approximately 1,000 steps per epoch.

Each checkpoint is 900MB, and I don't think uploading them all to HF is a good idea.

It's difficult to determine which checkpoint is best. The coarse training checkpoints generate lower quality images, but the images in the dataset are more diverse, preventing overfitting to any particular image.

The fine-tuning checkpoints are good in terms of image quality and artistic appeal, but they overfit the dataset, making each image in the dataset similar.

Therefore, I uploaded the last checkpoint.

During inference, using Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors to reduce steps and speed up inference, and using eigen-banana-qwen-image-edit-2509-fp16-lora.safetensors to adjust aesthetics, is a good idea.