ImageGen — SDXL LoRA (general-purpose)

LoRA fine-tune of stabilityai/stable-diffusion-xl-base-1.0 on a broad, filtered aesthetic dataset for improved prompt adherence and image quality over stock SDXL.

Training status

• Step: 6,200
• Epoch: 2.00
• Validation loss: 0.1364
• LoRA: rank 32, alpha 64, targets to_q, to_k, to_v, to_out.0
• Precision: bf16, gradient checkpointing on

Checkpoints

• latest/ — most recent adapter weights
• best/ — lowest validation-loss adapter weights

Dataset sources

• Spawning/PD12M (target ~30,000)
• common-canvas/commoncatalog-cc-by (target ~20,000)
• laion/laion2B-en-aesthetic (target ~23,000)
• laion/laion-art (target ~7,500)
• poloclub/diffusiondb (target ~12,500)
• kakaobrain/coyo-700m (target ~5,500)

License note: trained on research datasets of scraped image-text pairs. Released under cc-by-nc-4.0 (non-commercial). Verify each source dataset's terms before any downstream commercial use.

Usage

from diffusers import StableDiffusionXLPipeline
import torch

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("Dikshan1234/ImageGen", subfolder="best")
image = pipe("a cozy cabin in a snowy forest, golden hour").images[0]