This is a LoRA (Low-Rank Adaptation) model trained on the Qwen-Image base model to generate images of "saurian_oracle" - a fantastical character concept combining ancient wisdom with futuristic elements.

Trigger Word: saurian_oracle

Training Details

Base Model

Model: Qwen/Qwen-Image
Architecture: Qwen Image (Flow Matching based diffusion model)
Quantization: uint3 quantization with accuracy recovery adapters

Training Configuration

Training Steps: 3,000
Batch Size: 1
Learning Rate: 5e-05
Optimizer: AdamW 8-bit
LoRA Rank (Linear): 32
LoRA Alpha (Linear): 64
LoRA Rank (Conv): 16
LoRA Alpha (Conv): 16
Resolution: 768x512
Precision: bfloat16
Noise Scheduler: Flow Matching
Dataset Size: 23 images
Gradient Accumulation: 1
Weight Decay: 0.0001

Training Framework

Toolkit: AI-Toolkit
Save Format: Diffusers-compatible safetensors
Checkpoints Saved: Every 300 steps (keeping last 4)

Usage

With Diffusers

python

from diffusers import DiffusionPipeline
import torch

# Load base model
pipe = DiffusionPipeline.from_pretrained(
    "Qwen/Qwen-Image",
    torch_dtype=torch.bfloat16
)

# Load LoRA weights
pipe.load_lora_weights("Limbicnation/saurian-oracle-qwen-lora")

# Move to GPU
pipe = pipe.to("cuda")

# Generate image
prompt = "saurian_oracle, remarkable being with ancient and advanced metallic armor in a high-tech futuristic setting, captivating atmosphere, highly detailed, professional photography, masterpiece, best quality"
image = pipe(
    prompt,
    num_inference_steps=20,
    guidance_scale=4.0,
    width=768,
    height=768
).images[0]

image.save("saurian_oracle.png")

Recommended Parameters

Guidance Scale: 3.5 - 4.5
Steps: 20-25
Resolution: 768x768 or 512x768
Sampler: Flow Matching (default)

Prompt Examples

Here are some prompt examples that work well with this LoRA:

Futuristic Oracle:

saurian_oracle, remarkable being, with ancient and the advanced, metallic armor, in a high-tech futuristic setting, captivating atmosphere, highly detailed, professional photography, masterpiece, best quality
Cybernetic Oracle:

saurian_oracle, mysterious and enigmatic figure, with ancient and the modern collide, cybernetic suit that glints with a deep blue hue, dark, blue, organic, in a mysterious area environment, highly detailed, professional photography, masterpiece, best quality
Fantastical Oracle:

saurian_oracle, fantastical, with tail, armor, red, in a professional studio setting, highly detailed, professional photography, masterpiece, best quality
Ancient Tech Oracle:

saurian_oracle, power of ancient machine, with organic and the mechanical, ancient machines, dark, blue, organic, in a dark void environment, highly detailed, professional photography, masterpiece, best quality

Model Versions

This repository contains multiple checkpoints from the training run:

saurian_oracle_qwen.safetensors - Final model (3,000 steps)
saurian_oracle_qwen_000002700.safetensors - Checkpoint at 2,700 steps
saurian_oracle_qwen_000002400.safetensors - Checkpoint at 2,400 steps
saurian_oracle_qwen_000002100.safetensors - Checkpoint at 2,100 steps
saurian_oracle_qwen_000001800.safetensors - Checkpoint at 1,800 steps

The final model is recommended for general use.

Limitations and Bias

The model is trained on a small dataset (23 images) and may have limited variation
Best results are achieved when using the trigger word "saurian_oracle"
The model is optimized for the specific character design concept and may not generalize well to other concepts
Generated images may inherit biases present in the base Qwen-Image model

Technical Specifications

Model Size: ~282 MB per checkpoint
Format: SafeTensors
LoRA Modules: 840 U-Net modules
Text Encoder: Not trained (base model text encoder used)

License

This model is released under the Apache 2.0 license, following the base Qwen-Image model's licensing.

Citation

If you use this model, please cite:

bibtex

@misc{saurian_oracle_qwen_lora,
  title={Saurian Oracle - Qwen Image LoRA},
  author={Gero Doll},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/Limbicnation/saurian-oracle-qwen-lora}}
}

Acknowledgments

Base model: Qwen/Qwen-Image
Training toolkit: AI-Toolkit by Ostris
Quantization adapters: ostris/accuracy_recovery_adapters

Target Examples

These are the target oracle images used for training this LoRA: