LoRA text2image fine-tuning - Miracle-2001/pokemon-lora

These are LoRA adaption weights for runwayml/stable-diffusion-v1-5. The weights were fine-tuned on the svjack/pokemon-blip-captions-en-zh dataset. You can find some example images in the following.

Intended uses & limitations

How to use

import torch
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler
from huggingface_hub import model_info

# LoRA weights ~3 MB
model_path = "Miracle-2001/pokemon-lora"

info = model_info(model_path)
model_base = info.cardData["base_model"]
pipe = StableDiffusionPipeline.from_pretrained(model_base, torch_dtype=torch.float16)
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

pipe.unet.load_attn_procs(model_path)
pipe.to("cuda")

image = pipe("Green pokemon with menacing face", num_inference_steps=25).images[0]
image.save("green_pokemon.png")

Limitations and bias

[TODO: provide examples of latent issues and potential remediations]

Training details

cd diffusers/examples/text_to_image

export MODEL_NAME="runwayml/stable-diffusion-v1-5"
export OUTPUT_DIR="/flash2/aml/kyzhang24/HW3/hw3-base/"
export HUB_MODEL_ID="pokemon-lora"
export DATASET_NAME="svjack/pokemon-blip-captions-en-zh"

HF_ENDPOINT=https://hf-mirror.com python train_text_to_image_lora.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --dataset_name=$DATASET_NAME \
  --dataloader_num_workers=8 \
  --resolution=512 --center_crop --random_flip \
  --train_batch_size=1 \
  --gradient_accumulation_steps=4 \
  --max_train_steps=15000 \
  --learning_rate=1e-04 \
  --max_grad_norm=1 \
  --lr_scheduler="cosine" --lr_warmup_steps=0 \
  --output_dir=${OUTPUT_DIR} \
  --push_to_hub \
  --hub_model_id=${HUB_MODEL_ID} \
  --report_to=wandb \
  --checkpointing_steps=500 \
  --validation_prompt="Totoro" \
  --seed=1337 \
  --caption_column="en_text"