Merge pull request #5 from Stability-AI/main

update

.gitignore

@@ -0,0 +1,165 @@
+# Generated by project
+outputs/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# General MacOS
+.DS_Store
+.AppleDouble
+.LSOverride
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# IDEs
+.idea/
+.vscode/

README.md

@@ -1,12 +1,24 @@
-# Stable Diffusion 2.0
+# Stable Diffusion Version 2
 ![t2i](assets/stable-samples/txt2img/768/merged-0006.png)
 ![t2i](assets/stable-samples/txt2img/768/merged-0002.png)
 ![t2i](assets/stable-samples/txt2img/768/merged-0005.png)
 
 This repository contains [Stable Diffusion](https://github.com/CompVis/stable-diffusion) models trained from scratch and will be continuously updated with
 new checkpoints. The following list provides an overview of all currently available models. More coming soon.
+
 ## News
-**November 2022**
+
+**December 7, 2022**
+
+*Version 2.1*
+
+- New stable diffusion model (_Stable Diffusion 2.1-v_, [HuggingFace](https://huggingface.co/stabilityai/stable-diffusion-2-1)) at 768x768 resolution and (_Stable Diffusion 2.1-base_, [HuggingFace](https://huggingface.co/stabilityai/stable-diffusion-2-1-base)) at 512x512 resolution, both based on the same number of parameters and architecture as 2.0 and fine-tuned on 2.0, on a less restrictive NSFW filtering of the [LAION-5B](https://laion.ai/blog/laion-5b/) dataset.
+Per default, the attention operation of the model is evaluated at full precision when `xformers` is not installed. To enable fp16 (which can cause numerical instabilities with the vanilla attention module on the v2.1 model) , run your script with `ATTN_PRECISION=fp16 python <thescript.py>`
+
+**November 24, 2022**
+
+*Version 2.0*
+
 - New stable diffusion model (_Stable Diffusion 2.0-v_) at 768x768 resolution. Same number of parameters in the U-Net as 1.5, but uses [OpenCLIP-ViT/H](https://github.com/mlfoundations/open_clip) as the text encoder and is trained from scratch. _SD 2.0-v_ is a so-called [v-prediction](https://arxiv.org/abs/2202.00512) model. 
 - The above model is finetuned from _SD 2.0-base_, which was trained as a standard noise-prediction model on 512x512 images and is also made available.
 - Added a [x4 upscaling latent text-guided diffusion model](#image-upscaling-with-stable-diffusion).
@@ -80,11 +92,11 @@ The weights are available via [the StabilityAI organization at Hugging Face](htt
 
 
 
-## Stable Diffusion v2.0
+## Stable Diffusion v2
 
-Stable Diffusion v2.0 refers to a specific configuration of the model
+Stable Diffusion v2 refers to a specific configuration of the model
 architecture that uses a downsampling-factor 8 autoencoder with an 865M UNet
-and OpenCLIP ViT-H/14 text encoder for the diffusion model. The _SD 2.0-v_ model produces 768x768 px outputs. 
+and OpenCLIP ViT-H/14 text encoder for the diffusion model. The _SD 2-v_ model produces 768x768 px outputs. 
 
 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 DDIM sampling steps show the relative improvements of the checkpoints:
@@ -97,16 +109,16 @@ Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,
 ![txt2img-stable2](assets/stable-samples/txt2img/merged-0003.png)
 ![txt2img-stable2](assets/stable-samples/txt2img/merged-0001.png)
 
-Stable Diffusion 2.0 is a latent diffusion model conditioned on the penultimate text embeddings of a CLIP ViT-H/14 text encoder.
+Stable Diffusion 2 is a latent diffusion model conditioned on the penultimate text embeddings of a CLIP ViT-H/14 text encoder.
 We provide a [reference script for sampling](#reference-sampling-script).
 #### Reference Sampling Script
 
 This script incorporates an [invisible watermarking](https://github.com/ShieldMnt/invisible-watermark) of the outputs, to help viewers [identify the images as machine-generated](scripts/tests/test_watermark.py).
-We provide the configs for the _SD2.0-v_ (768px) and _SD2.0-base_ (512px) model.
+We provide the configs for the _SD2-v_ (768px) and _SD2-base_ (512px) model.
 
-First, download the weights for [_SD2.0-v_](https://huggingface.co/stabilityai/stable-diffusion-2) and [_SD2.0-base_](https://huggingface.co/stabilityai/stable-diffusion-2-base). 
+First, download the weights for [_SD2.1-v_](https://huggingface.co/stabilityai/stable-diffusion-2-1) and [_SD2.1-base_](https://huggingface.co/stabilityai/stable-diffusion-2-1-base). 
 
-To sample from the _SD2.0-v_ model, run the following:
+To sample from the _SD2.1-v_ model, run the following:
 
 ```
 python scripts/txt2img.py --prompt "a professional photograph of an astronaut riding a horse" --ckpt <path/to/768model.ckpt/> --config configs/stable-diffusion/v2-inference-v.yaml --H 768 --W 768  

ldm/models/diffusion/ddpm.py

@@ -390,7 +390,7 @@ class DDPM(pl.LightningModule):
         elif self.parameterization == "v":
             target = self.get_v(x_start, noise, t)
         else:
-            raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
+            raise NotImplementedError(f"Parameterization {self.parameterization} not yet supported")
 
         loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
 

ldm/modules/attention.py

@@ -16,6 +16,9 @@ try:
 except:
     XFORMERS_IS_AVAILBLE = False
 
+# CrossAttn precision handling
+import os
+_ATTN_PRECISION = os.environ.get("ATTN_PRECISION", "fp32")
 
 def exists(val):
     return val is not None
@@ -167,9 +170,16 @@ class CrossAttention(nn.Module):
 
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
 
-        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+        # force cast to fp32 to avoid overflowing
+        if _ATTN_PRECISION =="fp32":
+            with torch.autocast(enabled=False, device_type = 'cuda'):
+                q, k = q.float(), k.float()
+                sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+        else:
+            sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+        
         del q, k
-
+    
         if exists(mask):
             mask = rearrange(mask, 'b ... -> b (...)')
             max_neg_value = -torch.finfo(sim.dtype).max

modelcard.md

@@ -80,7 +80,7 @@ Stable Diffusion v2 mirrors and exacerbates biases to such a degree that viewer
 **Training Data**
 The model developers used the following dataset for training the model:
 
-- 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.
+- LAION-5B and subsets (details below). The training data is further filtered using LAION's NSFW detector. For more details, please refer to LAION-5B's [NeurIPS 2022](https://openreview.net/forum?id=M3Y74vmsMcY) paper and reviewer discussions on the topic.
 
 **Training Procedure**
 Stable 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, 
@@ -90,7 +90,13 @@ Stable Diffusion v2 is a latent diffusion model which combines an autoencoder wi
 - The 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 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.
 
-We currently provide the following checkpoints:
+We currently provide the following checkpoints, for various versions:
+
+### Version 2.1
+
+- `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.
+- `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`.
+### Version 2.0
 
 - `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`.
   850k steps at resolution `512x512` on the same dataset with resolution `>= 512x512`.

requirements.txt

@@ -13,4 +13,7 @@ transformers==4.19.2
 webdataset==0.2.5
 open-clip-torch==2.7.0
 gradio==3.11
+kornia==0.6
+invisible-watermark>=0.1.5
+streamlit-drawable-canvas==0.8.0
 -e .