Hugging Face's logo

steveant
/
steve-lora-v1.1

Text-to-Image
Diffusers
flux
lora
template:sd-lora
Model card Files Community

steve_lora_flux_1_dev_v1.1

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.

Prompt
[trigger], man lounging in fitted athletic wear on crisp white linens, strong and confident expression, warm ambient lighting, full-body shot, textured fabric details
Prompt
[trigger], man standing in a steamy, industrial-style shower, tousled hair, water droplets glistening on muscular physique, frosted glass, subtle highlights on skin, intimate and rugged atmosphere, high resolution
Prompt
[trigger], man sunbathing by a private pool, board shorts, focused gaze over shoulder, chiseled skin, shimmering water, sharp focus, vibrant and intense sunlight
Prompt
[trigger], man reclining in a luxurious, stone-tiled bubble bath, dimmed Edison bulb candles, rising steam, strong and defined physique highlighted by soft intimate lighting, elegant and masculine decor
Prompt
[trigger], man posing in front of a full-length mirror, tailored blazer and dark jeans, confident smirk, dual reflections, soft lighting accentuating muscular build, full visibility
Prompt
[trigger], man riding a rugged vintage motorcycle down an open highway, leather jacket and sturdy boots, fierce and determined gaze, expansive desert landscape, dynamic motion blur, powerful stance
Prompt
[trigger], man dancing with strength and grace in a dimly lit ballroom, tailored suit with subtle flair, joyful and intense expression, sparkling chandeliers, elegant yet masculine architecture, full-body movement shot
Prompt
[trigger], man hiking through a dense, lush forest trail, durable outdoor gear, determined and focused look, dappled sunlight filtering through vibrant greenery, full-body action shot showcasing athletic build
Prompt
[trigger], man preparing a gourmet meal in a sleek, modern kitchen, stylish apron over casual attire, focused and purposeful expression, high-end appliances, fresh and hearty ingredients, bright and clean lighting
Prompt
[trigger], man performing on stage under a dramatic spotlight, sharp and stylish evening attire, intense and captivating gaze, theater backdrop with bold textures, dynamic and powerful poses, full-body and close-up shots
Prompt
[trigger], man meditating on a tranquil mountaintop at dusk, comfortable and functional yoga wear, serene and strong expression, panoramic sunset views, soft twilight lighting highlighting muscular physique
Prompt
[trigger], man exploring an ancient castle, adventurous yet sophisticated attire, curious and bold expression, grand stone hallways, historic and rugged decor, interplay of natural and candlelight emphasizing strength and elegance

Trigger Words

Use steve in your prompt to activate the specific style and character details for this LoRA.

Model Information

  • LoRA Rank / Alpha: 32 / 32
  • 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:

from diffusers import StableDiffusionPipeline
import torch

base_model = "black-forest-labs/FLUX.1-dev"
lora_model = "YOUR_USERNAME/steve_lora_flux_1_dev_v1.1"

pipe = StableDiffusionPipeline.from_pretrained(base_model, torch_dtype=torch.float16).to("cuda")
# Load your LoRA weights (implementation depends on the UI or method)
# pipe.load_lora_weights(lora_model)  # Example call

prompt = "steve, man lounging in fitted athletic wear on crisp white linens, strong and confident"
image = pipe(prompt).images[0]
image.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-v1.1

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 v1.1 and is currently considered experimental.

Model Details

  • Model type: LoRA adapter for Stable Diffusion (sd_trainer)
  • Trigger word: steve
  • Base model: black-forest-labs/FLUX.1-dev
  • Network: LoRA (rank: 32, alpha: 32)
  • Quantization: Enabled
  • 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: 32
    • Linear alpha: 32
  • Sampling configuration (for sample images):
    • Sampler: flowmatch
    • Resolution: 1024 x 1024
    • Guidance scale: 5
    • Sample steps: 23

During training, image captions were drawn from .txt files. Some techniques applied include:

  • Caption dropout: 0.05
  • 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

  1. Installation
    Make sure you have the Diffusers library or another Stable Diffusion–compatible framework installed.

  2. Loading the Model 

    from diffusers import StableDiffusionPipeline
import torch

# Example: Pseudocode for loading the base model + LoRA
base_model_id = "black-forest-labs/FLUX.1-dev"  
lora_model_id = "steveant/steve-lora-v1.1"  # hypothetical path on HF hub

pipeline = StableDiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16).to("cuda")
# Load LoRA weights
# Typically, you would merge or apply the LoRA as per your chosen library's method.

  3. Prompting
    Use the trigger word steve in your prompt to invoke the specific style or character details. For instance:

    prompt = (
    "steve, man lounging in fitted athletic wear on crisp white linens, "
    "strong and confident expression, warm ambient lighting, full-body shot, "
    "textured fabric details"
)
result = pipeline(prompt).images[0]
result.save("steve_lounging.png")

  4. Negative Prompting (Optional)
    Provide a neg (negative) prompt parameter to omit or reduce undesired elements. 

    neg_prompt = "low resolution, bad quality"
result = pipeline(prompt=prompt, negative_prompt=neg_prompt).images[0]

Sample Prompts

Below are some sample prompts used during training:

  • steve, man standing in a steamy, industrial-style shower, water droplets glistening on muscular physique
  • steve, man sunbathing by a private pool, shimmering water, sharp focus
  • steve, man riding a rugged vintage motorcycle down an open highway, leather jacket and sturdy boots
  • steve, man meditating on a tranquil mountaintop at dusk, comfortable and functional yoga wear

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_v1.1,
  author = {steveant},
  title = {steve_lora_flux_1_dev_v1.1 (LoRA model)},
  year = {2024},
  howpublished = {\url{https://huggingface.co/steveant/steve-lora-v1.1}},
}

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:

  1. Open an issue on the GitHub/Spaces project (if available).
  2. Submit pull requests or suggestions.
  3. Respect the usage and license guidelines.

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.

Downloads last month
30
Inference API
Text-to-Image
Examples

Model tree for steveant/steve-lora-v1.1

Base model

black-forest-labs/FLUX.1-dev
Adapter
(13054)
this model