Ray0821/ivgeng-2026-lab2-style-lora

Lab 2 Style LoRA

A geometric illustration style adapter trained on the four artistic-custom example images supplied by the course starter repository. The course README identifies these images as GPT-4 generated. Captions: a bear, a car, a cat, a dog. This adapter is a course experiment, not a claim to reproduce any named artist.

Training

Base: Stable Diffusion 1.4. UNet attention LoRA rank 4; 797,184 trainable adapter parameters. The text encoder and base parameters remain frozen. Resolution 512, batch size 1, learning rate 1e-4, constant schedule, training seed 42, fp16. Completed 400 optimizer steps. The supplied course algorithm implementation was used without modification.

Dataset and training code source: course starter repository.

Inference

import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16
).to("cuda")
pipe.load_lora_weights("Ray0821/ivgeng-2026-lab2-style-lora")
image = pipe(
    "a bear", num_inference_steps=30, guidance_scale=7.5,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
image.save("sample.png")

Final comparisons use the base model's PNDM scheduler, 30 inference steps and seeds 42 and 123, with matched base/adapter settings. Only adapter weights are provided, not the base model.

Limitations

Style transfer is strongest for the training object classes. Some unseen subjects, especially a cottage beside a river, retain a landscape-painting or photographic appearance. Colors and transfer strength vary with the seed.

Integrity

File: pytorch_lora_weights.safetensors (3,226,184 bytes).

SHA-256: 0616ec65bba635e6f0b44067c560b8a1f9114f13ed0868e2eb2a9dd378e19394.

The adapter is distributed under the base model's CreativeML Open RAIL-M license and use restrictions.