---
language:
- en
license: apache-2.0
library_name: diffusers
pipeline_tag: text-to-image
tags:
- text-to-image
- image-generation
- shuttle
widget:
- text: >-
    Venus floating market at dawn, fantasy digital art, highly detailed, atmospheric lighting with film-like light leaks, impressive background, studio photo style, cinematic, intricate details.
  output:
    url: gallery/1.webp
- text: >-
    Silent forest, sun barely piercing treetops, mysterious lake turns dark red at dawn, reflecting colorful sky. Lone tree on shore with diamond-like dewdrops, photorealistic.
  output:
    url: gallery/2.webp
- text: >-
    A beautiful photo showcases a night waterfall in the jungle, illuminated with a subtle blue tint that adds an ethereal touch. Fireflies float delicately around, their gentle glow enhancing the magical ambiance of the scene.
  output:
    url: gallery/3.webp
    
instance_prompt: null
---

# Shuttle 3 Diffusion

<Gallery />

## Model Variants
These model variants provide different precision levels and formats optimized for diverse hardware capabilities and use cases
- [bfloat16](https://huggingface.co/shuttleai/shuttle-3-diffusion)
- [GGUF](https://huggingface.co/shuttleai/shuttle-3-diffusion-GGUF)
- [fp8](https://huggingface.co/shuttleai/shuttle-3-diffusion-fp8)

Shuttle 3 Diffusion is a text-to-image AI model designed to create detailed and diverse images from textual prompts in just 4 steps. It offers enhanced performance in image quality, typography, understanding complex prompts, and resource efficiency.

![image/png](https://huggingface.co/shuttleai/shuttle-3-diffusion/resolve/main/demo.png)

You can try out the model through a website at https://chat.shuttleai.com/images

## Using the model via API
You can use Shuttle 3 Diffusion via API through ShuttleAI
- [ShuttleAI](https://shuttleai.com/)
- [ShuttleAI Docs](https://docs.shuttleai.com/)

## Using the model with 🧨 Diffusers
Install or upgrade diffusers
```shell
pip install -U diffusers
```
Then you can use `DiffusionPipeline` to run the model
```python
import torch
from diffusers import DiffusionPipeline

# Load the diffusion pipeline from a pretrained model, using bfloat16 for tensor types.
pipe = DiffusionPipeline.from_pretrained(
    "shuttleai/shuttle-3-diffusion", torch_dtype=torch.bfloat16
).to("cuda")

# Uncomment the following line to save VRAM by offloading the model to CPU if needed.
# pipe.enable_model_cpu_offload()

# Uncomment the lines below to enable torch.compile for potential performance boosts on compatible GPUs.
# Note that this can increase loading times considerably.
# pipe.transformer.to(memory_format=torch.channels_last)
# pipe.transformer = torch.compile(
#     pipe.transformer, mode="max-autotune", fullgraph=True
# )

# Set your prompt for image generation.
prompt = "A cat holding a sign that says hello world"

# Generate the image using the diffusion pipeline.
image = pipe(
    prompt,
    height=1024,
    width=1024,
    guidance_scale=3.5,
    num_inference_steps=4,
    max_sequence_length=256,
    # Uncomment the line below to use a manual seed for reproducible results.
    # generator=torch.Generator("cpu").manual_seed(0)
).images[0]

# Save the generated image.
image.save("shuttle.png")
```
To learn more check out the [diffusers](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux) documentation

## Using the model with ComfyUI

To run local inference with Shuttle 3 Diffusion using [ComfyUI](https://github.com/comfyanonymous/ComfyUI), you can use this [safetensors file](https://huggingface.co/shuttleai/shuttle-3-diffusion/blob/main/shuttle-3-diffusion.safetensors).

## Comparison to other models
Shuttle 3 Diffusion can produce images better images than Flux Dev in just four steps, while being licensed under Apache 2.
![image/png](https://huggingface.co/shuttleai/shuttle-3-diffusion/resolve/main/comparison.png)
[More examples](https://docs.shuttleai.com/getting-started/shuttle-diffusion)

## Training Details
Shuttle 3 Diffusion uses Flux.1 Schnell as its base. It can produce images similar to Flux Dev or Pro in just 4 steps, and it is licensed under Apache 2. The model was partially de-distilled during training. When used beyond 10 steps, it enters "refiner mode," enhancing image details without altering the composition. We overcame the limitations of the Schnell-series models by employing a special training method, resulting in improved details and colors.