keras/stable_diffusion_3.5_large

Model Overview

Stable Diffusion 3.5 is a Multimodal Diffusion Transformer (MMDiT) text-to-image model that features greatly improved performance in image quality, typography, complex prompt understanding, and resource-efficiency.

For more technical details, please refer to the Research paper.

Please note: this model is released under the Stability Community License. For Enterprise License visit Stability.ai or contact us for commercial licensing details.

Links

• SD3.5 Quickstart Notebook
• SD3.5 API Documentation
• SD3.5 Model Card
• KerasHub Beginner Guide
• KerasHub Model Publishing Guide

Presets

The following model checkpoints are provided by the Keras team. Full code examples for each are available below.

Preset name	Parameters	Description
stable_diffusion_3.5_large	9.05B	9 billion parameter, including CLIP L and CLIP G text encoders, MMDiT generative model, and VAE autoencoder. Developed by Stability AI.
stable_diffusion_3.5_large_turbo	9.05B	9 billion parameter, including CLIP L and CLIP G text encoders, MMDiT generative model, and VAE autoencoder. A timestep-distilled version that eliminates classifier-free guidance and uses fewer steps for generation. Developed by Stability AI.

Model Description

• Developed by: Stability AI
• Model type: MMDiT text-to-image generative model
• Model Description: This is a model that can be used to generate images based on text prompts. It is a Multimodal Diffusion Transformer that uses three fixed, pretrained text encoders (OpenCLIP-ViT/G, CLIP-ViT/L and T5-xxl), and QK-normalization to improve training stability.

Example Usage

!pip install -U keras-hub
!pip install -U keras

# Pretrained Stable Diffusion 3 model.
model = keras_hub.models.StableDiffusion3Backbone.from_preset(
    "stable_diffusion_3.5_large"
)

# Randomly initialized Stable Diffusion 3 model with custom config.
vae = keras_hub.models.VAEBackbone(...)
clip_l = keras_hub.models.CLIPTextEncoder(...)
clip_g = keras_hub.models.CLIPTextEncoder(...)
model = keras_hub.models.StableDiffusion3Backbone(
    mmdit_patch_size=2,
    mmdit_num_heads=4,
    mmdit_hidden_dim=256,
    mmdit_depth=4,
    mmdit_position_size=192,
    vae=vae,
    clip_l=clip_l,
    clip_g=clip_g,
)

# Image to image example
image_to_image = keras_hub.models.StableDiffusion3ImageToImage.from_preset(
        "stable_diffusion_3.5_large", height=512, width=512
)
image_to_image.generate(
    {
        "images": np.ones((512, 512, 3), dtype="float32"),
        "prompts": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    }
)

# Generate with batched prompts.
image_to_image.generate(
    {
        "images": np.ones((2, 512, 512, 3), dtype="float32"),
        "prompts": ["cute wallpaper art of a cat", "cute wallpaper art of a dog"],
    }
)

# Generate with different `num_steps`, `guidance_scale` and `strength`.
image_to_image.generate(
    {
        "images": np.ones((512, 512, 3), dtype="float32"),
        "prompts": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    }
    num_steps=50,
    guidance_scale=5.0,
    strength=0.6,
)

# Generate with `negative_prompts`.
text_to_image.generate(
    {
        "images": np.ones((512, 512, 3), dtype="float32"),
        "prompts": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
        "negative_prompts": "green color",
    }
)

# inpainting example
reference_image = np.ones((1024, 1024, 3), dtype="float32")
reference_mask = np.ones((1024, 1024), dtype="float32")
inpaint = keras_hub.models.StableDiffusion3Inpaint.from_preset(
    "stable_diffusion_3.5_large", height=512, width=512
)
inpaint.generate(
    reference_image,
    reference_mask,
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
)

# Generate with batched prompts.
reference_images = np.ones((2, 512, 512, 3), dtype="float32")
reference_mask = np.ones((2, 1024, 1024), dtype="float32")
inpaint.generate(
    reference_images,
    reference_mask,
    ["cute wallpaper art of a cat", "cute wallpaper art of a dog"]
)

# Generate with different `num_steps`, `guidance_scale` and `strength`.
inpaint.generate(
    reference_image,
    reference_mask,
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    num_steps=50,
    guidance_scale=5.0,
    strength=0.6,
)

# text to image example
text_to_image = keras_hub.models.StableDiffusion3TextToImage.from_preset(
   "stable_diffusion_3.5_large", height=512, width=512
)
text_to_image.generate(
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
)

# Generate with batched prompts.
text_to_image.generate(
    ["cute wallpaper art of a cat", "cute wallpaper art of a dog"]
)

# Generate with different `num_steps` and `guidance_scale`.
text_to_image.generate(
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    num_steps=50,
    guidance_scale=5.0,
)

# Generate with `negative_prompts`.
text_to_image.generate(
    {
        "prompts": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
        "negative_prompts": "green color",
    }
)

Example Usage with Hugging Face URI

!pip install -U keras-hub
!pip install -U keras

# Pretrained Stable Diffusion 3 model.
model = keras_hub.models.StableDiffusion3Backbone.from_preset(
    "hf://keras/stable_diffusion_3.5_large"
)

# Randomly initialized Stable Diffusion 3 model with custom config.
vae = keras_hub.models.VAEBackbone(...)
clip_l = keras_hub.models.CLIPTextEncoder(...)
clip_g = keras_hub.models.CLIPTextEncoder(...)
model = keras_hub.models.StableDiffusion3Backbone(
    mmdit_patch_size=2,
    mmdit_num_heads=4,
    mmdit_hidden_dim=256,
    mmdit_depth=4,
    mmdit_position_size=192,
    vae=vae,
    clip_l=clip_l,
    clip_g=clip_g,
)

# Image to image example
image_to_image = keras_hub.models.StableDiffusion3ImageToImage.from_preset(
        "hf://keras/stable_diffusion_3.5_large", height=512, width=512
)
image_to_image.generate(
    {
        "images": np.ones((512, 512, 3), dtype="float32"),
        "prompts": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    }
)

# Generate with batched prompts.
image_to_image.generate(
    {
        "images": np.ones((2, 512, 512, 3), dtype="float32"),
        "prompts": ["cute wallpaper art of a cat", "cute wallpaper art of a dog"],
    }
)

# Generate with different `num_steps`, `guidance_scale` and `strength`.
image_to_image.generate(
    {
        "images": np.ones((512, 512, 3), dtype="float32"),
        "prompts": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    }
    num_steps=50,
    guidance_scale=5.0,
    strength=0.6,
)

# Generate with `negative_prompts`.
text_to_image.generate(
    {
        "images": np.ones((512, 512, 3), dtype="float32"),
        "prompts": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
        "negative_prompts": "green color",
    }
)

# inpainting example
reference_image = np.ones((1024, 1024, 3), dtype="float32")
reference_mask = np.ones((1024, 1024), dtype="float32")
inpaint = keras_hub.models.StableDiffusion3Inpaint.from_preset(
    "hf://keras/stable_diffusion_3.5_large", height=512, width=512
)
inpaint.generate(
    reference_image,
    reference_mask,
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
)

# Generate with batched prompts.
reference_images = np.ones((2, 512, 512, 3), dtype="float32")
reference_mask = np.ones((2, 1024, 1024), dtype="float32")
inpaint.generate(
    reference_images,
    reference_mask,
    ["cute wallpaper art of a cat", "cute wallpaper art of a dog"]
)

# Generate with different `num_steps`, `guidance_scale` and `strength`.
inpaint.generate(
    reference_image,
    reference_mask,
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    num_steps=50,
    guidance_scale=5.0,
    strength=0.6,
)

# text to image example
text_to_image = keras_hub.models.StableDiffusion3TextToImage.from_preset(
   "hf://keras/stable_diffusion_3.5_large", height=512, width=512
)
text_to_image.generate(
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
)

# Generate with batched prompts.
text_to_image.generate(
    ["cute wallpaper art of a cat", "cute wallpaper art of a dog"]
)

# Generate with different `num_steps` and `guidance_scale`.
text_to_image.generate(
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    num_steps=50,
    guidance_scale=5.0,
)

# Generate with `negative_prompts`.
text_to_image.generate(
    {
        "prompts": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
        "negative_prompts": "green color",
    }
)