Skittles v2

Model Summary

Skittles v2 is a cutting-edge text-to-image generation model, created by merging components from the FLUX.1 Schnell architecture. By combining the precision of FLUX.1 Schnell with advanced tweaks, Skittles v2 is designed to offer high-quality image outputs.

Type: Text-to-Image Generation
Architecture: Merged FLUX.1 Schnell with CFG capabilities
Output Quality: Seems to be on par with FLUX.1 Dev
Performance: Optimized for both image fidelity

Features

CFG Integration: Skittles v2 unlocks CFG (Classifier-Free Guidance) capabilities, offering fine-grained control over image generation.
High Fidelity: Produces ultra-realistic and detailed images.
Customizable Output: Supports a wide range of prompts, styles, and configurations.

Model Details

Base Model: FLUX.1 Schnell
Merge Approach: The model was combined using a custom merging strategy, blending FLUX.1 Schnell’s architecture with optimized CFG decoding.
Training Paradigm: Not retrained, but restructured for improved inference performance.
Output Size: Supports resolutions up to 1024x1024 pixels.

Intended Use

Applications

Generating ultra-realistic images for creative projects
Creating concept art, visual prototypes, and artistic renderings
Exploration of text-to-image synthesis for research or artistic purposes

Examples

Prompt	Image Description
"Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"	Hyper-detailed astronaut surrounded by lush, muted jungle tones
"A futuristic cityscape at sunset, ultra-realistic, cinematic, 4K"	Vibrant, glowing cityscape with dynamic lighting effects
"A delicious ceviche cheesecake slice"	Highly detailed and realistic rendition of a culinary masterpiece

How to Use

from diffusers import DiffusionPipeline

# Load the model
pipe = DiffusionPipeline.from_pretrained("miike-ai/skittles-v2")
pipe.to("cuda")  # Ensure CUDA is available

# Generate an image
prompt = "An ultra-realistic image of a futuristic cityscape."
image = pipe(prompt, guidance_scale=3.5, num_inference_steps=28).images[0]

# Save the result
image.save("generated_image.png")

Limitations and Biases

The model may produce biased or stereotypical outputs based on the provided text prompts.
Outputs are deterministic but rely heavily on the prompt quality. Results may vary with ambiguous descriptions.
The model is not trained to handle NSFW content or harmful prompts.

Acknowledgments

Built on top of FLUX.1 Schnell by Black Forest Labs
Contributions from miike-ai
Integrated with Hugging Face Diffusers for seamless inference

Citation

If you use Skittles v2 in your work, please cite:

@misc{miike2024skittlesv2,
  title={Skittles v2: A Merged Text-to-Image Generation Model},
  author={miike-ai},
  year={2024},
  url={https://huggingface.co/miike-ai/skittles-v2},
}

miike-ai
/

skittles-v2