import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "neonforestmist/Clover-Image-Tiny",
    torch_dtype=torch.float16,
).to("cuda")
pipe.load_lora_weights(
    "neonforestmist/clover-image-tiny-watercolor-anime-lora"
)

image = pipe(
    "watercolor anime, a small blue cat beside a quiet garden stream",
    num_inference_steps=20,
    guidance_scale=7.5,
).images[0]

Training

Base revision: 63b0e9f6be9c00888ff464f342a9ef052bf76681
Dataset revision: a08ad49bc28e9cebd704fd1010846357220c5c5d
Resolution: 512 × 512
Optimizer steps: 1,200
Rank: 16
Batch size: 1
Learning rate: 1e-4 with cosine decay and 120 warmup steps
Min-SNR gamma: 5
Precision: fp16
Seed: 20260730
Trainer: Diffusers 0.39.0 train_text_to_image_lora.py

The dataset revision embeds each image in its Parquet row so the exact training input remains portable in remote jobs. The reproducible pinned-data job configuration is included in the Clover source repository under training/.

Provenance

Base model: neonforestmist/Clover-Image-Tiny
Training dataset: neonforestmist/GPT_Watercolor_Anime_Style_Images
Core ML base: neonforestmist/Clover-Image-Tiny-CoreML
Training and conversion source: neonforestmist/clover-image-tiny-lora-trainer

License and limitations

These adapter weights are a derivative of Clover Image Tiny and use the CreativeML Open RAIL-M license. The training dataset is Apache-2.0. Generated content can inherit limitations and biases from the base checkpoint and training data; review outputs before use.