gramatchi/llava-1.5-7b-restoration-lora

LLaVA-1.5-7B restoration planner (LoRA, checkpoint-28800)

LoRA adapter for liuhaotian/llava-v1.5-7b. Given a photo, it answers with the distortions it finds (JPEG compression, Gaussian noise, exposure shift; zero to three), their severity, the order in which to undo them, and the restoration method for each step.

Files

File	Size	What
adapter_model.bin	640 MB	LoRA weights (r=64, alpha=128, all attention and MLP projections)
non_lora_trainables.bin	84 MB	fine-tuned multimodal projector
adapter_config.json, config.json	small	configs

Loading

The base model is downloaded from Hugging Face separately. With the LLaVA repo (see the training repo for the exact commit and our changes):

from llava.model.builder import load_pretrained_model

tokenizer, model, image_processor, _ = load_pretrained_model(
    model_path="path/to/this/folder",
    model_base="liuhaotian/llava-v1.5-7b",
    model_name="llava-lora-restoration",
)

The model name has to contain lora, otherwise LLaVA does not merge the adapter and you silently get the base model. The same goes for the folder name if the name is derived from the path.

Training

Trained on 38400 examples (2400 photos, 16 examples per photo, quarter each with 0/1/2/3 distortions) with the distortions applied on the fly. 3 epochs, lr 2e-4, batch size 4, one L40S. Training code and data: see the GitHub repo. Evaluation on 1000 held-out photos: https://github.com/gramatchi/final-test-llava

Limitations

Trained and tested only on synthetic distortions of three types applied to ordinary photos. Mild distortions are often missed, and exposure shifts are hard to tell from a naturally dark or bright scene. Not tested on real forensic or surveillance material.

License

The adapter is a derivative of LLaVA-1.5-7B, so the terms of the base model (and of the Llama 2 model it builds on) apply. Check them before reuse.