Stable Diffusion V1 5

🚀 Stable Diffusion v1-5 Model Card

Stable Diffusion is a latent text-to-image diffusion model that can generate photo-realistic images from any text input. For more details on how Stable Diffusion works, refer to 🤗's Stable Diffusion blog.

🚀 Quick Start

You can use this model with both the 🧨Diffusers library and the RunwayML GitHub repository.

✨ Features

• Capable of generating photo-realistic images from text input.
• Fine-tuned on specific datasets to improve performance.
• Can be used with different libraries and repositories.

📦 Installation

Diffusers

from diffusers import StableDiffusionPipeline
import torch

model_id = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

prompt = "a photo of an astronaut riding a horse on mars"
image = pipe(prompt).images[0]  
    
image.save("astronaut_rides_horse.png")

For more detailed instructions, use-cases, and examples in JAX, follow the instructions here.

Original GitHub Repository

1. Download the weights:
   • v1-5-pruned-emaonly.ckpt - 4.27GB, ema-only weight. Uses less VRAM, suitable for inference.
   • v1-5-pruned.ckpt - 7.7GB, ema+non-ema weights. Uses more VRAM, suitable for fine-tuning.
2. Follow the instructions here.

📚 Documentation

Model Details

💻 Usage Examples

Basic Usage

from diffusers import StableDiffusionPipeline
import torch

model_id = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

prompt = "a photo of an astronaut riding a horse on mars"
image = pipe(prompt).images[0]  
    
image.save("astronaut_rides_horse.png")

🔧 Technical Details

Training

Training Data The model was trained on the following dataset:

• LAION-2B (en) and subsets thereof.

Training Procedure Stable Diffusion v1 - 5 is a latent diffusion model that combines an autoencoder with a diffusion model trained in the autoencoder's latent space. During training:

• Images are encoded by an encoder to turn them into latent representations. The autoencoder uses a relative downsampling factor of 8 and maps images of shape H x W x 3 to latents of shape H/f x W/f x 4.
• Text prompts are encoded by a ViT - L/14 text - encoder.
• The non - pooled output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross - attention.
• The loss is a reconstruction objective between the noise added to the latent and the prediction made by the UNet.

Currently, six Stable Diffusion checkpoints are provided, trained as follows:

• stable-diffusion-v1-1: 237,000 steps at resolution 256x256 on laion2B-en. 194,000 steps at resolution 512x512 on laion-high-resolution (170M examples from LAION - 5B with resolution >= 1024x1024).
• stable-diffusion-v1-2: Resumed from stable-diffusion-v1-1. 515,000 steps at resolution 512x512 on "laion-improved-aesthetics" (a subset of laion2B - en, filtered to images with an original size >= 512x512, estimated aesthetics score > 5.0, and an estimated watermark probability < 0.5. The watermark estimate is from the LAION - 5B metadata, and the aesthetics score is estimated using an improved aesthetics estimator).
• stable-diffusion-v1-3: Resumed from stable-diffusion-v1-2 - 195,000 steps at resolution 512x512 on "laion-improved-aesthetics" and 10 % dropping of the text - conditioning to improve classifier-free guidance sampling.
• stable-diffusion-v1-4 Resumed from stable-diffusion-v1-2 - 225,000 steps at resolution 512x512 on "laion-aesthetics v2 5+" and 10 % dropping of the text - conditioning to improve classifier-free guidance sampling.
• stable-diffusion-v1-5 Resumed from stable-diffusion-v1-2 - 595,000 steps at resolution 512x512 on "laion-aesthetics v2 5+" and 10 % dropping of the text - conditioning to improve classifier-free guidance sampling.
• stable-diffusion-inpainting Resumed from stable-diffusion-v1-5 - then 440,000 steps of inpainting training at resolution 512x512 on “laion-aesthetics v2 5+” and 10% dropping of the text - conditioning. For inpainting, the UNet has 5 additional input channels (4 for the encoded masked - image and 1 for the mask itself) whose weights were zero - initialized after restoring the non - inpainting checkpoint. During training, synthetic masks are generated, and in 25% of cases, everything is masked.

Hardware: 32 x 8 x A100 GPUs Optimizer: AdamW Gradient Accumulations: 2 Batch: 32 x 8 x 2 x 4 = 2048 Learning rate: Warmed up to 0.0001 for 10,000 steps and then kept constant.

Limitations and Bias

Limitations

• The model doesn't achieve perfect photorealism.
• It can't render legible text.
• It performs poorly on more complex tasks involving compositionality, like rendering an image of “A red cube on top of a blue sphere”.
• Faces and people may not be generated accurately.
• Trained mainly with English captions, it doesn't work as well in other languages.
• The autoencoding part of the model is lossy.
• Trained on the large - scale dataset LAION - 5B, which contains adult material and isn't suitable for product use without additional safety measures.
• No additional measures were used to deduplicate the dataset, resulting in some memorization of duplicated images in the training data. The training data can be searched at https://rom1504.github.io/clip-retrieval/ to detect memorized images.

Bias

Image generation models can reinforce or exacerbate social biases. Stable Diffusion v1 was trained on subsets of LAION - 2B(en), mainly consisting of images with English descriptions. Texts and images from non - English communities and cultures may be under - represented, leading to a bias where white and western cultures are often the default. Also, the model performs significantly worse with non - English prompts.

Safety Module

The model is intended to be used with the Safety Checker in Diffusers. This checker compares model outputs against known hard - coded NSFW concepts. The concepts are hidden to prevent reverse - engineering. Specifically, it compares the class probability of harmful concepts in the embedding space of the CLIPTextModel after image generation. The concepts are passed into the model with the generated image and compared to a hand - engineered weight for each NSFW concept.

📄 License

This model is open access and available to all, under a CreativeML OpenRAIL - M license that further specifies rights and usage.

The CreativeML OpenRAIL License states:

1. You can't use the model to deliberately produce or share illegal or harmful outputs or content.
2. CompVis claims no rights on the outputs you generate. You are free to use them but accountable for their use, which must comply with the license provisions.
3. You may re - distribute the weights and use the model commercially or as a service. If you do, you must include the same use restrictions as in the license and share a copy of the CreativeML OpenRAIL - M with all your users. Read the full license here.

Uses

Direct Use

The model is for research purposes only. Possible research areas and tasks include:

• Safe deployment of models with the potential to generate harmful content.
• Probing and understanding the limitations and biases of generative models.
• Generation of artworks and use in design and other artistic processes.
• Applications in educational or creative tools.
• Research on generative models.

Misuse, Malicious Use, and Out - of - Scope Use

Note: This section is adapted from the [DALLE - MINI model card](https://huggingface.co/dalle - mini/dalle - mini) and applies to Stable Diffusion v1.

The model shouldn't be used to create or spread images that create hostile or alienating environments. This includes generating disturbing, distressing, or offensive content, or content that propagates stereotypes.

Out - of - Scope Use

The model wasn't trained to provide factual or true representations of people or events. Using it for such purposes is beyond its capabilities.

Misuse and Malicious Use

Using the model to generate cruel content towards individuals is a misuse. This includes, but is not limited to:

• Generating demeaning, dehumanizing, or harmful representations of people, their environments, cultures, religions, etc.
• Intentionally promoting or propagating discriminatory content or harmful stereotypes.
• Impersonating individuals without consent.
• Generating sexual content without the consent of those who might see it.
• Spreading mis - and disinformation.
• Representing egregious violence and gore.
• Sharing copyrighted or licensed material in violation of its terms of use.
• Sharing altered copyrighted or licensed material in violation of its terms of use.

Evaluation Results

Evaluations with different classifier - free guidance scales (1.5, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0) and 50 PNDM/PLMS sampling steps show the relative improvements of the checkpoints.