Shuttle 3 Diffusion - AI Image Models Tool

Overview

Shuttle 3 Diffusion is a text-to-image diffusion model from ShuttleAI that emphasizes fast, resource-efficient image generation and improved handling of typography and complex prompts. The model is designed to produce usable 1024×1024 images in as few as four inference steps (the default fast mode), while also supporting longer runs for detail refinement. The project provides multiple distribution formats (bfloat16, GGUF, fp8) and integration options so developers can run the model locally or through ShuttleAI's API and UI. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3-diffusion)) Shuttle 3 Diffusion targets creators and developers who need quick iterations (prototyping, previews, or real-time workflows) while retaining an option to refine and upscale outputs when higher fidelity is required. It ships under an Apache-2.0 license and has been repackaged into quantized variants (GGUF and fp8) for lower-memory inference on consumer hardware. The model is available on Hugging Face with example pipelines for Diffusers and guidance on running in ComfyUI or via ShuttleAI's hosted endpoints. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3-diffusion))

Model Statistics

• Downloads: 870
• Likes: 212
• Pipeline: text-to-image

License: apache-2.0

Model Details

Architecture and base: Shuttle 3 Diffusion is built on the FLUX family (base: FLUX.1 Schnell) and uses a diffusion pipeline tuned for aggressive step reduction (4-step sampling) with a "refiner mode" that improves detail if you run more than ~10 steps. The developers describe a partial de‑distillation/finetuning approach to overcome Schnell-series limitations, improving color and detail fidelity. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3-diffusion)) Model variants and formats: Official releases include bfloat16 (bf16), GGUF and fp8 formats. The GGUF variant is published with a reported 12B-parameter footprint and provides quantized runtime sizes such as Q4_K_S (~6.88 GB), Q6_K (~9.94 GB), Q8_0 (~12.8 GB), and BF16 (~23.9 GB), allowing selection of tradeoffs between VRAM usage and quality. These variants let you run Shuttle 3 Diffusion on a range of GPU configurations or use lower-precision formats for CPU/GGUF workflows. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3-diffusion-GGUF)) Performance and usage: The recommended fast configuration commonly used in examples is num_inference_steps=4, guidance_scale≈3.5, and typical canvas sizes up to 1024×1024. For integration, the model exposes a Diffusers-compatible pipeline (example code on the Hugging Face model card) and is packaged for ComfyUI and ShuttleAI API access. When used beyond the 10-step threshold it operates as a refiner, enhancing fine detail without changing overall composition. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3-diffusion))

Key Features

• 4-step fast sampling for rapid image previews and interactive workflows.
• Refiner mode: running >10 steps increases detail without changing composition.
• Multiple formats: bfloat16, GGUF, and fp8 for varied hardware targets.
• Optimized typography and improved prompt understanding for legible text.
• Apache-2.0 license allowing broad reuse and downstream integration.

Example Usage

Example (python):

import torch
from diffusers import DiffusionPipeline

# Load the Shuttle 3 Diffusion pipeline (example uses bfloat16). Adjust device and precision to your hardware.
pipe = DiffusionPipeline.from_pretrained(
    "shuttleai/shuttle-3-diffusion",
    torch_dtype=torch.bfloat16
).to("cuda")

prompt = "A photorealistic portrait of an astronaut holding a sunflower, studio lighting"

image = pipe(
    prompt,
    height=1024,
    width=1024,
    guidance_scale=3.5,
    num_inference_steps=4,
    max_sequence_length=256,
).images[0]

image.save("shuttle3_output.png")

# Notes: you can enable CPU offload or experiment with torch.compile per Hugging Face examples.
# See the model card for ComfyUI or API integration examples. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3-diffusion))

Pricing

Shuttle 3 Diffusion model files are freely available on Hugging Face under an Apache-2.0 license; ShuttleAI provides hosted API access under tiered subscription plans (Free, Basic $10/month, Premium $25/month, Scale $75/month) — check ShuttleAI pricing and docs for exact API usage quotas and whether the model is available on your chosen plan. ([huggingface.co](https://huggingface.co/shuttleai/shuttle-3-diffusion))

Benchmarks

Hugging Face likes: 212 likes (model page) (Source: https://huggingface.co/shuttleai/shuttle-3-diffusion)

Downloads (last month, model page): 849 downloads (last month, reported on model card) (Source: https://huggingface.co/shuttleai/shuttle-3-diffusion)

GGUF model size: 12B parameters (GGUF variant) (Source: https://huggingface.co/shuttleai/shuttle-3-diffusion-GGUF)

Quantized memory footprints (GGUF): Q4_K_S 6.88 GB | Q6_K 9.94 GB | Q8_0 12.8 GB | BF16 23.9 GB (Source: https://huggingface.co/shuttleai/shuttle-3-diffusion-GGUF)

Pipeline type: Text-to-image (Diffusers / FluxPipeline) (Source: https://huggingface.co/shuttleai/shuttle-3-diffusion)

Key Information

• Category: Image Models
• Type: AI Image Models Tool