glowforge-dev/stable-diffusion-2-1-img2img-jonathan2
046
1---2license: openrail++3tags:4- stable-diffusion5- image-to-image6widget:7 - src: >-8 https://huggingface.co/datasets/mishig/sample_images/resolve/main/canny-edge.jpg9 prompt: Girl with Pearl Earring10inference:11 parameters:12 height: 51213 width: 51214 negative_prompt: ''15duplicated_from: stabilityai/stable-diffusion-2-1-base16---17 18# Stable Diffusion v2-1-base Model Card19This model card focuses on the model associated with the Stable Diffusion v2-1-base model.20 21This `stable-diffusion-2-1-base` model fine-tunes [stable-diffusion-2-base](https://huggingface.co/stabilityai/stable-diffusion-2-base) (`512-base-ema.ckpt`) with 220k extra steps taken, with `punsafe=0.98` on the same dataset.22 23- Use it with the [`stablediffusion`](https://github.com/Stability-AI/stablediffusion) repository: download the `v2-1_512-ema-pruned.ckpt` [here](https://huggingface.co/stabilityai/stable-diffusion-2-1-base/resolve/main/v2-1_512-ema-pruned.ckpt).24- Use it with 🧨 [`diffusers`](#examples)25 26## Model Details27- **Developed by:** Robin Rombach, Patrick Esser28- **Model type:** Diffusion-based text-to-image generation model29- **Language(s):** English30- **License:** [CreativeML Open RAIL++-M License](https://huggingface.co/stabilityai/stable-diffusion-2/blob/main/LICENSE-MODEL)31- **Model Description:** This is a model that can be used to generate and modify images based on text prompts. It is a [Latent Diffusion Model](https://arxiv.org/abs/2112.10752) that uses a fixed, pretrained text encoder ([OpenCLIP-ViT/H](https://github.com/mlfoundations/open_clip)).32- **Resources for more information:** [GitHub Repository](https://github.com/Stability-AI/).33- **Cite as:**34 35 @InProceedings{Rombach_2022_CVPR,36 author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},37 title = {High-Resolution Image Synthesis With Latent Diffusion Models},38 booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},39 month = {June},40 year = {2022},41 pages = {10684-10695}42 }43 44 45## Examples46 47Using the [🤗's Diffusers library](https://github.com/huggingface/diffusers) to run Stable Diffusion 2 in a simple and efficient manner.48 49```bash50pip install diffusers transformers accelerate scipy safetensors51```52Running the pipeline (if you don't swap the scheduler it will run with the default PNDM/PLMS scheduler, in this example we are swapping it to EulerDiscreteScheduler):53 54```python55from diffusers import StableDiffusionPipeline, EulerDiscreteScheduler56import torch57 58model_id = "stabilityai/stable-diffusion-2-1-base"59 60scheduler = EulerDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler")61pipe = StableDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler, torch_dtype=torch.float16)62pipe = pipe.to("cuda")63 64prompt = "a photo of an astronaut riding a horse on mars"65image = pipe(prompt).images[0]66 67image.save("astronaut_rides_horse.png")68```69 70**Notes**:71- Despite not being a dependency, we highly recommend you to install [xformers](https://github.com/facebookresearch/xformers) for memory efficient attention (better performance)72- If you have low GPU RAM available, make sure to add a `pipe.enable_attention_slicing()` after sending it to `cuda` for less VRAM usage (to the cost of speed)73 74# Uses75 76## Direct Use77The model is intended for research purposes only. Possible research areas and tasks include78 79- Safe deployment of models which have the potential to generate harmful content.80- Probing and understanding the limitations and biases of generative models.81- Generation of artworks and use in design and other artistic processes.82- Applications in educational or creative tools.83- Research on generative models.84 85Excluded uses are described below.86 87 ### Misuse, Malicious Use, and Out-of-Scope Use88_Note: This section is originally taken from the [DALLE-MINI model card](https://huggingface.co/dalle-mini/dalle-mini), was used for Stable Diffusion v1, but applies in the same way to Stable Diffusion v2_.89 90The model should not be used to intentionally create or disseminate images that create hostile or alienating environments for people. This includes generating images that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.91 92#### Out-of-Scope Use93The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.94 95#### Misuse and Malicious Use96Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:97 98- Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc.99- Intentionally promoting or propagating discriminatory content or harmful stereotypes.100- Impersonating individuals without their consent.101- Sexual content without consent of the people who might see it.102- Mis- and disinformation103- Representations of egregious violence and gore104- Sharing of copyrighted or licensed material in violation of its terms of use.105- Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use.106 107## Limitations and Bias108 109### Limitations110 111- The model does not achieve perfect photorealism112- The model cannot render legible text113- The model does not perform well on more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”114- Faces and people in general may not be generated properly.115- The model was trained mainly with English captions and will not work as well in other languages.116- The autoencoding part of the model is lossy117- The model was trained on a subset of the large-scale dataset118 [LAION-5B](https://laion.ai/blog/laion-5b/), which contains adult, violent and sexual content. To partially mitigate this, we have filtered the dataset using LAION's NFSW detector (see Training section).119 120### Bias121While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.122Stable Diffusion vw was primarily trained on subsets of [LAION-2B(en)](https://laion.ai/blog/laion-5b/),123which consists of images that are limited to English descriptions.124Texts and images from communities and cultures that use other languages are likely to be insufficiently accounted for.125This affects the overall output of the model, as white and western cultures are often set as the default. Further, the126ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts.127Stable Diffusion v2 mirrors and exacerbates biases to such a degree that viewer discretion must be advised irrespective of the input or its intent.128 129 130## Training131 132**Training Data**133The model developers used the following dataset for training the model:134 135- LAION-5B and subsets (details below). The training data is further filtered using LAION's NSFW detector, with a "p_unsafe" score of 0.1 (conservative). For more details, please refer to LAION-5B's [NeurIPS 2022](https://openreview.net/forum?id=M3Y74vmsMcY) paper and reviewer discussions on the topic.136 137**Training Procedure**138Stable Diffusion v2 is a latent diffusion model which combines an autoencoder with a diffusion model that is trained in the latent space of the autoencoder. During training,139 140- Images are encoded through an encoder, which turns images 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 4141- Text prompts are encoded through the OpenCLIP-ViT/H text-encoder.142- The output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross-attention.143- The loss is a reconstruction objective between the noise that was added to the latent and the prediction made by the UNet. We also use the so-called _v-objective_, see https://arxiv.org/abs/2202.00512.144 145We currently provide the following checkpoints, for various versions:146 147### Version 2.1148- `512-base-ema.ckpt`: Fine-tuned on `512-base-ema.ckpt` 2.0 with 220k extra steps taken, with `punsafe=0.98` on the same dataset.149- `768-v-ema.ckpt`: Resumed from `768-v-ema.ckpt` 2.0 with an additional 55k steps on the same dataset (`punsafe=0.1`), and then fine-tuned for another 155k extra steps with `punsafe=0.98`.150 151### Version 2.0152- `512-base-ema.ckpt`: 550k steps at resolution `256x256` on a subset of [LAION-5B](https://laion.ai/blog/laion-5b/) filtered for explicit pornographic material, using the [LAION-NSFW classifier](https://github.com/LAION-AI/CLIP-based-NSFW-Detector) with `punsafe=0.1` and an [aesthetic score](https://github.com/christophschuhmann/improved-aesthetic-predictor) >= `4.5`.153 850k steps at resolution `512x512` on the same dataset with resolution `>= 512x512`.154- `768-v-ema.ckpt`: Resumed from `512-base-ema.ckpt` and trained for 150k steps using a [v-objective](https://arxiv.org/abs/2202.00512) on the same dataset. Resumed for another 140k steps on a `768x768` subset of our dataset.155- `512-depth-ema.ckpt`: Resumed from `512-base-ema.ckpt` and finetuned for 200k steps. Added an extra input channel to process the (relative) depth prediction produced by [MiDaS](https://github.com/isl-org/MiDaS) (`dpt_hybrid`) which is used as an additional conditioning.156The additional input channels of the U-Net which process this extra information were zero-initialized.157- `512-inpainting-ema.ckpt`: Resumed from `512-base-ema.ckpt` and trained for another 200k steps. Follows the mask-generation strategy presented in [LAMA](https://github.com/saic-mdal/lama) which, in combination with the latent VAE representations of the masked image, are used as an additional conditioning.158The additional input channels of the U-Net which process this extra information were zero-initialized. The same strategy was used to train the [1.5-inpainting checkpoint](https://github.com/saic-mdal/lama).159- `x4-upscaling-ema.ckpt`: Trained for 1.25M steps on a 10M subset of LAION containing images `>2048x2048`. The model was trained on crops of size `512x512` and is a text-guided [latent upscaling diffusion model](https://arxiv.org/abs/2112.10752).160In addition to the textual input, it receives a `noise_level` as an input parameter, which can be used to add noise to the low-resolution input according to a [predefined diffusion schedule](configs/stable-diffusion/x4-upscaling.yaml).161 162- **Hardware:** 32 x 8 x A100 GPUs163- **Optimizer:** AdamW164- **Gradient Accumulations**: 1165- **Batch:** 32 x 8 x 2 x 4 = 2048166- **Learning rate:** warmup to 0.0001 for 10,000 steps and then kept constant167 168## Evaluation Results169Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,1705.0, 6.0, 7.0, 8.0) and 50 steps DDIM sampling steps show the relative improvements of the checkpoints:171 172173 174Evaluated using 50 DDIM steps and 10000 random prompts from the COCO2017 validation set, evaluated at 512x512 resolution. Not optimized for FID scores.175 176## Environmental Impact177 178**Stable Diffusion v1** **Estimated Emissions**179Based on that information, we estimate the following CO2 emissions using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). The hardware, runtime, cloud provider, and compute region were utilized to estimate the carbon impact.180 181- **Hardware Type:** A100 PCIe 40GB182- **Hours used:** 200000183- **Cloud Provider:** AWS184- **Compute Region:** US-east185- **Carbon Emitted (Power consumption x Time x Carbon produced based on location of power grid):** 15000 kg CO2 eq.186 187## Citation188 @InProceedings{Rombach_2022_CVPR,189 author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},190 title = {High-Resolution Image Synthesis With Latent Diffusion Models},191 booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},192 month = {June},193 year = {2022},194 pages = {10684-10695}195 }196 197*This model card was written by: Robin Rombach, Patrick Esser and David Ha and is based on the [Stable Diffusion v1](https://github.com/CompVis/stable-diffusion/blob/main/Stable_Diffusion_v1_Model_Card.md) and [DALL-E Mini model card](https://huggingface.co/dalle-mini/dalle-mini).*198