A LoRA-based Stable Diffusion model trained to generate images of a man named “Steve” in a wide variety of scenarios. This model is fine-tuned from black-forest-labs/FLUX.1-dev using a Flow Matching–based noise scheduler.
Trigger Words
Use steve in your prompt to activate the specific style and character details for this LoRA.
Model Information
LoRA Rank / Alpha: 16 / 16
Number of Steps: 4000
Batch Size: 1
Learning Rate: 0.0001
Noise Scheduler: flowmatch
Optimizer: adamw8bit
Precision: bf16
Gradient Checkpointing: true
EMA: true (decay = 0.99)
Quantization: enabled
How to Use
This LoRA can be merged or applied to the FLUX.1-dev base model through Diffusers or a compatible UI/tool.
Example pseudocode:
python
1from diffusers import StableDiffusionPipeline
2import torch
34base_model ="black-forest-labs/FLUX.1-dev"5lora_model ="YOUR_USERNAME/steve_lora_flux_1_dev_v1.2"67pipe = StableDiffusionPipeline.from_pretrained(base_model, torch_dtype=torch.float16).to("cuda")8# Load your LoRA weights (implementation depends on the UI or method)9# pipe.load_lora_weights(lora_model) # Example call1011prompt ="steve, man lounging in fitted athletic wear on crisp white linens, strong and confident"12image = pipe(prompt).images[0]13image.save("steve_example.jpg")
Download Model
Weights for this LoRA are available in Safetensors format.
Download them from the Files & versions tab.
License
This model is provided under a flux-1-dev-non-commercial-license. Please review the license file for details on acceptable use.
Acknowledgements
Trained with AI Toolkit by Ostris
Based on the FLUX.1-dev base model
Disclaimer:
Use responsibly. This model is intended for artistic, non-commercial purposes. The creators are not responsible for any misuse, generation of disallowed content, or potential harm caused by outputs. Always review and curate model outputs before sharing.
steveant/steve-lora-v1.2
This is a LoRA-based Stable Diffusion model fine-tuned on a custom image dataset to generate images featuring a man named “Steve” in various settings and scenarios. It has been trained using the FLUX.1-dev base model, leveraging a Flow Matching–based noise scheduler and LoRA network adapters.
Note: This model is in version v1.2 and is currently considered experimental.
Model Details
Model type: LoRA adapter for Stable Diffusion (sd_trainer)
Datasets: Private dataset containing images and associated textual captions.
Model Architecture and Training
This LoRA was trained using the following key parameters:
Training steps: 4000
Batch size: 1
Gradient accumulation steps: 1
Learning rate: 0.0001
Noise Scheduler: flowmatch
Optimizer: adamw8bit
Precision: bf16
LoRA settings:
Linear rank: 16
Linear alpha: 16
Sampling configuration (for sample images):
Sampler: flowmatch
Resolution: 1200 x 1600
Guidance scale: 4.1
Sample steps: 29
During training, image captions were drawn from .txt files. Some techniques applied include:
Caption dropout: 0.00
Token shuffling: true
Gradient checkpointing: true
Exponential moving average: use_ema = true with ema_decay = 0.99
Intended Use
This model is intended to generate images of a “Steve” character in various poses, outfits, and scenarios. Possible use cases include:
Creative media and content generation
Character concepting for artistic projects
Test and experimentation with Flow Matching–based schedulers in Stable Diffusion
Important: This model is not intended to generate explicit or harmful content. Users are advised to comply with local regulations and handle outputs responsibly.
How to Use
Installation
Make sure you have the Diffusers library or another Stable Diffusion–compatible framework installed.
Loading the Model
python
1from diffusers import StableDiffusionPipeline
2import torch
34# Example: Pseudocode for loading the base model + LoRA5base_model_id ="black-forest-labs/FLUX.1-dev"6lora_model_id ="steveant/steve-lora-v1.2"# hypothetical path on HF hub78pipeline = StableDiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16).to("cuda")9# Load LoRA weights10# Typically, you would merge or apply the LoRA as per your chosen library's method.
Prompting
Use the trigger wordsteve in your prompt to invoke the specific style or character details. For instance:
python
1prompt =(2"steve, man lounging in fitted athletic wear on crisp white linens, "3"strong and confident expression, warm ambient lighting, full-body shot, "4"textured fabric details"5)6result = pipeline(prompt).images[0]7result.save("steve_lounging.png")
Negative Prompting (Optional)
Provide a neg (negative) prompt parameter to omit or reduce undesired elements.
python
1neg_prompt ="low resolution, bad quality"2result = pipeline(prompt=prompt, negative_prompt=neg_prompt).images[0]
Sample Prompts
Below are some sample prompts used during training:
A photorealistic business headshot of steve, a Caucasian man in his 40s exuding confidence in a modern office setting.
A professional portrait of steve, a Caucasian man in his 40s and full head of hair, dressed sharply in a navy suit, captured with soft natural light.
A high-quality headshot of steve, a Caucasian man in his 40s and full head of hair, taken with an 85mm lens, emphasizing realism and authority.
A corporate headshot of steve, a Caucasian man in his 40s, standing before a sleek office backdrop with a confident expression.
Limitations and Biases
The model’s outputs depend on the style and content of the dataset.
Since the training data is limited to “Steve” images in specific scenarios, the model may not generalize well to drastically different contexts.
Bias: Any biases in the original dataset might be reflected in the generated images.
Training Data
Private dataset of images featuring “Steve,” labeled with text captions.
Resolution used for latent caching: 720, 960, and 1440.
Data augmentation: Slight caption dropout, token shuffling, etc.
Citation
If you use this model or find it helpful for your research/projects, please cite:
@misc{steve_lora_flux_1_dev_v1.2,
author = {steveant},
title = {steve_lora_flux_1_dev_v1.2 (LoRA model)},
year = {2024},
howpublished = {\url{https://huggingface.co/steveant/steve-lora-v1.2}},
}
License
This model and code are available under CreativeML Open RAIL-M or your chosen license. Please refer to the repository’s license or contact the author for more details.
Contributing
Contributions are welcome! If you wish to improve this model card or have new use cases and improvements to propose:
Disclaimer:
This model is for research and educational purposes. Always validate and review images generated to ensure they align with your intended use and do not violate any regulations or ethical standards.
