New AI edits images based on text instructions

imagine --tile-x -w 1024 -h 512 "360 degree equirectangular panorama photograph of the desert"  --upscale

  3d-term, adj-architecture, adj-beauty, adj-detailed, adj-emotion, adj-general, adj-horror, animal, art-movement, 
  art-site, artist, artist-botanical, artist-surreal, aspect-ratio, bird, body-of-water, body-pose, camera-brand,
  camera-model, color, cosmic-galaxy, cosmic-nebula, cosmic-star, cosmic-term, dinosaur, eyecolor, f-stop, 
  fantasy-creature, fantasy-setting, fish, flower, focal-length, food, fruit, games, gen-modifier, hair, hd,
  iso-stop, landscape-type, national-park, nationality, neg-weight, noun-beauty, noun-fantasy, noun-general, 
  noun-horror, occupation, photo-term, pop-culture, pop-location, punk-style, quantity, rpg-item, scenario-desc, 
  skin-color, spaceship, style, tree-species, trippy, world-heritage-site

• It makes images from text descriptions! ????
• Generate images either in code or from command line.
• It just works. Proper requirements are installed. model weights are automatically downloaded. No huggingface account needed.
  (if you have the right hardware… and aren’t on windows)
• No more distorted faces!
• Noisy logs are gone (which was surprisingly hard to accomplish)
• WeightedPrompts let you smash together separate prompts (cat-dog)
• Tile Mode creates tileable images
• Prompt metadata saved into image file metadata
• Edit images by describing the part you want edited (see example above)
• Have AI generate captions for images aimg describe
• Interactive prompt: just run aimg
• ???? finetune your own image model. kind of like dreambooth. Read instructions on “Concept Training” page

How To

For full command line instructions run aimg --help

from imaginairy import imagine, imagine_image_files, ImaginePrompt, WeightedPrompt, LazyLoadingImage

url = "https://upload.wikimedia.org/wikipedia/commons/thumb/6/6c/Thomas_Cole_-_Architect%E2%80%99s_Dream_-_Google_Art_Project.jpg/540px-Thomas_Cole_-_Architect%E2%80%99s_Dream_-_Google_Art_Project.jpg"
prompts = [
    ImaginePrompt("a scenic landscape", seed=1, upscale=True),
    ImaginePrompt("a bowl of fruit"),
    ImaginePrompt([
        WeightedPrompt("cat", weight=1),
        WeightedPrompt("dog", weight=1),
    ]),
    ImaginePrompt(
        "a spacious building", 
        init_image=LazyLoadingImage(url=url)
    ),
    ImaginePrompt(
        "a bowl of strawberries", 
        init_image=LazyLoadingImage(filepath="mypath/to/bowl_of_fruit.jpg"),
        mask_prompt="fruit OR stem{*2}",  # amplify the stem mask x2
        mask_mode="replace",
        mask_modify_original=True,
    ),
    ImaginePrompt("strawberries", tile_mode=True),
]
for result in imagine(prompts):
    # do something
    result.save("my_image.jpg")

# or

imagine_image_files(prompts, outdir="./my-art")

Requirements

• ~10 gb space for models to download
• A CUDA supported graphics card with >= 11gb VRAM (and CUDA installed) or an M1 processor.
• Python installed. Preferably Python 3.10. (not conda)
• For macOS rust and setuptools-rust must be installed to compile the tokenizer library.
  They can be installed via: curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh and pip install setuptools-rust

Running in Docker

See example Dockerfile (works on machine where you can pass the gpu into the container)

docker build . -t imaginairy
# you really want to map the cache or you end up wasting a lot of time and space redownloading the model weights
docker run -it --gpus all -v $HOME/.cache/huggingface:/root/.cache/huggingface -v $HOME/.cache/torch:/root/.cache/torch -v `pwd`/outputs:/outputs imaginairy /bin/bash

Running on Google Colab

Example Colab

ChangeLog

8.0.0

• feature: ???? edit images with instructions alone!
• feature: when editing an image add --gif to create a comparision gif
• feature: aimg edit --suprise-me --gif my-image.jpg for some fun pre-programmed edits
• feature: prune-ckpt command also removes the non-ema weights

7.6.0

• fix: default model config was broken
• feature: print version with --version
• feature: ability to load safetensors
• feature: ???? outpainting. Examples: --outpaint up10,down300,left50,right50 or --outpaint all100 or --outpaint u100,d200,l300,r400

7.4.3

• fix: handle old pytorch lightning imports with a graceful failure (fixes #161)
• fix: handle failed image generations better (fixes #83)

7.4.2

• fix: run face enhancement on GPU for 10x speedup

7.4.1

• fix: incorrect config files being used for non-1.0 models

7.4.0

• feature: ???? finetune your own image model. kind of like dreambooth. Read instructions on “Concept Training” page
• feature: image prep command. crops to face or other interesting parts of photo
• fix: back-compat for hf_hub_download
• feature: add prune-ckpt command
• feature: allow specification of model config file

7.3.0

• feature: ???? depth-based image-to-image generations (and inpainting)
• fix: k_euler_a produces more consistent images per seed (randomization respects the seed again)

7.2.0

• feature: ???? tile in a single dimension (“x” or “y”). This enables, with a bit of luck, generation of 360 VR images.
  Try this for example: imagine --tile-x -w 1024 -h 512 "360 degree equirectangular panorama photograph of the mountains" --upscale

7.1.1

• fix: memory/speed regression introduced in 6.1.0
• fix: model switching now clears memory better, thus avoiding out of memory errors

7.1.0

• feature: ???? Stable Diffusion 2.1. Generated people are no longer (completely) distorted.
  Use with --model SD-2.1 or --model SD-2.0-v

7.0.0

• feature: negative prompting. --negative-prompt or ImaginePrompt(..., negative_prompt="ugly, deformed, extra arms, etc")
• feature: a default negative prompt is added to all generations. Images in SD-2.0 don’t look bad anymore. Images in 1.5 look improved as well.

6.1.2

• fix: add back in memory-efficient algorithms

6.1.1

• feature: xformers will be used if available (for faster generation)
• fix: version metadata was broken

6.1.0

• feature: use different default steps and image sizes depending on sampler and model selceted
• fix: #110 use proper version in image metadata
• refactor: samplers all have their own class that inherits from ImageSampler
• feature: ???????????? Stable Diffusion 2.0
  • --model SD-2.0 to use (it makes worse images than 1.5 though…)
  • Tested on macOS and Linux
  • All samplers working for new 512×512 model
  • New inpainting model working
  • 768×768 model working for all samplers except PLMS (--model SD-2.0-v )

5.1.0

• feature: add progress image callback

5.0.1

• fix: support larger images on M1. Fixes #8
• fix: support CPU generation by disabling autocast on CPU. Fixes #81

5.0.0

• feature: ???? inpainting support using new inpainting model from RunwayML. It works really well! By default, the
  inpainting model will automatically be used for any image-masking task
• feature: ???? new default sampler makes image generation more than twice as fast
• feature: added DPM++ 2S a and DPM++ 2M samplers.
• feature: improve progress image logging
• fix: fix bug with --show-work. fixes #84
• fix: add workaround for pytorch bug affecting macOS users using the new DPM++ 2S a and DPM++ 2M samplers.
• fix: add workaround for pytorch mps bug affecting k_dpm_fast sampler. fixes #75
• fix: larger image sizes now work on macOS. fixes #8

4.1.0

• feature: allow dynamic switching between models/weights --model SD-1.5 or --model SD-1.4 or --model path/my-custom-weights.ckpt)
• feature: log total progress when generating images (image X out of Y)

4.0.0

• feature: stable diffusion 1.5 (slightly improved image quality)
• feature: dilation and erosion of masks
  Previously the + and - characters in a mask (example: face{+0.1}) added to the grayscale value of any masked areas. This wasn’t very useful. The new behavior is that the mask will expand or contract by the number of pixel specified. The technical terms for this are dilation and erosion. This allows much greater control over the masked area.
• feature: update k-diffusion samplers. add k_dpm_adaptive and k_dpm_fast
• feature: img2img/inpainting supported on all samplers
• refactor: consolidates img2img/txt2img code. consolidates schedules. consolidates masking
• ci: minor logging improvements

3.0.1

• fix: k-samplers were broken

3.0.0

• feature: improved safety filter

2.4.0

• ???? feature: prompt expansion
• feature: make (blip) photo captions more descriptive

2.3.1

• fix: face fidelity default was broken

2.3.0

• feature: model weights file can be specified via --model-weights-path argument at the command line
• fix: set face fidelity default back to old value
• fix: handle small images without throwing exception. credit to @NiclasEriksen
• docs: add setuptools-rust as dependency for macos

2.2.1

• fix: init image is fully ignored if init-image-strength = 0

2.2.0

• feature: face enhancement fidelity is now configurable

2.1.0

• improved masking accuracy from clipseg

2.0.3

• fix memory leak in face enhancer
• fix blurry inpainting
• fix for pillow compatibility

2.0.0

• ???? fix: inpainted areas correlate with surrounding image, even at 100% generation strength. Previously if the generation strength was high enough the generated image
  would be uncorrelated to the rest of the surrounding image. It created terrible looking images.
• ???? feature: interactive prompt added. access by running aimg
• ???? feature: Specify advanced text based masks using boolean logic and strength modifiers. Mask descriptions must be lowercase. Keywords uppercase.
  Valid symbols: AND, OR, NOT, (), and mask strength modifier {+0.1} where + can be any of + - * /. Single character boolean operators also work (|, &, !)
• ???? feature: apply mask edits to original files with mask_modify_original (on by default)
• feature: auto-rotate images if exif data specifies to do so
• fix: mask boundaries are more accurate
• fix: accept mask images in command line
• fix: img2img algorithm was wrong and wouldn’t at values close to 0 or 1

1.6.2

• fix: another bfloat16 fix

1.6.1

• fix: make sure image tensors come to the CPU as float32 so there aren’t compatibility issues with non-bfloat16 cpus

1.6.0

• fix: maybe address #13 with expected scalar type BFloat16 but found Float
  • at minimum one can specify --precision full now and that will probably fix the issue
• feature: tile mode can now be specified per-prompt

1.5.3

• fix: missing config file for describe feature

1.5.1

• img2img now supported with PLMS (instead of just DDIM)
• added image captioning feature aimg describe dog.jpg => a brown dog sitting on grass
• added new commandline tool aimg for additional image manipulation functionality

1.4.0

• support multiple additive targets for masking with | symbol. Example: “fruit|stem|fruit stem”

1.3.0

• added prompt based image editing. Example: “fruit => gold coins”
• test coverage improved

1.2.0

• allow urls as init-images

previous

• img2img actually does # of steps you specify
• performance optimizations
• numerous other changes

Not Supported

• a GUI. this is a python library
• exploratory features that don’t work well

Todo

• Performance Optimizations
  • ✅ https://github.com/huggingface/diffusers/blob/main/docs/source/optimization/fp16.mdx
  • ✅ https://github.com/CompVis/stable-diffusion/compare/main…Doggettx:stable-diffusion:autocast-improvements#
  • ✅ https://www.reddit.com/r/StableDiffusion/comments/xalaws/test_update_for_less_memory_usage_and_higher/
  • https://github.com/neonsecret/stable-diffusion
  • CompVis/stable-diffusion#177
  • https://github.com/huggingface/diffusers/pull/532/files
  • https://github.com/HazyResearch/flash-attention
  • ✅ xformers improvements https://www.photoroom.com/tech/stable-diffusion-100-percent-faster-with-memory-efficient-attention/
• Development Environment
  • ✅ add tests
  • ✅ set up ci (test/lint/format)
  • ✅ unified pipeline (txt2img & img2img combined)
  • ✅ setup parallel testing
  • add docs
  • remove yaml config
  • delete more unused code
  • faster latent logging https://discuss.huggingface.co/t/decoding-latents-to-rgb-without-upscaling/23204/9
• Interface improvements
  • ✅ init-image at command line
  • ✅ prompt expansion
  • ✅ interactive cli
• Image Generation Features
  • ✅ add k-diffusion sampling methods
  • ✅ tiling
  • generation videos/gifs
  • Compositional Visual Generation
    • https://github.com/energy-based-model/Compositional-Visual-Generation-with-Composable-Diffusion-Models-PyTorch
    • https://colab.research.google.com/github/energy-based-model/Compositional-Visual-Generation-with-Composable-Diffusion-Models-PyTorch/blob/main/notebooks/demo.ipynb#scrollTo=wt_j3uXZGFAS
  • ✅ negative prompting
    • some syntax to allow it in a text string
  • ???? images as actual prompts instead of just init images.
    • not directly possible due to model architecture.
    • requires model fine-tuning since SD1.4 expects 77×768 text encoding input
    • https://twitter.com/Buntworthy/status/1566744186153484288
    • https://github.com/justinpinkney/stable-diffusion
    • https://github.com/LambdaLabsML/lambda-diffusers
    • https://www.reddit.com/r/MachineLearning/comments/x6k5bm/n_stable_diffusion_image_variations_released/
• Image Editing
  • outpainting
    • https://github.com/parlance-zz/g-diffuser-bot/search?q=noise&type=issues
    • lama cleaner
  • ✅ inpainting
    • https://github.com/Jack000/glid-3-xl-stable
    • https://github.com/andreas128/RePaint
    • ✅ img2img but keeps img stable
    • https://www.reddit.com/r/StableDiffusion/comments/xboy90/a_better_way_of_doing_img2img_by_finding_the/
    • https://gist.github.com/trygvebw/c71334dd127d537a15e9d59790f7f5e1
    • https://github.com/pesser/stable-diffusion/commit/bbb52981460707963e2a62160890d7ecbce00e79
    • https://github.com/SHI-Labs/FcF-Inpainting https://praeclarumjj3.github.io/fcf-inpainting/
  • ✅ text based image masking
    • ✅ ClipSeg – https://github.com/timojl/clipseg
    • https://github.com/facebookresearch/detectron2
  • Attention Control Methods
    • https://github.com/bloc97/CrossAttentionControl
    • https://github.com/ChenWu98/cycle-diffusion
• Image Enhancement
  • Photo Restoration – https://github.com/microsoft/Bringing-Old-Photos-Back-to-Life
  • Upscaling
    • ✅ realesrgan
    • ldm
    • https://github.com/lowfuel/progrock-stable
    • gobig
    • stable super-res?
      • todo: try with 1-0-0-0 mask at full image resolution (rencoding entire image+predicted image at every step)
      • todo: use a gaussian pyramid and only include the “high-detail” level of the pyramid into the next step
      • https://www.reddit.com/r/StableDiffusion/comments/xkjjf9/upscale_to_huge_sizes_and_add_detail_with_sd/
  • ✅ face enhancers
    • ✅ gfpgan – https://github.com/TencentARC/GFPGAN
    • ✅ codeformer – https://github.com/sczhou/CodeFormer
  • ✅ image describe feature –
    • ✅ https://github.com/salesforce/BLIP
    • ???? CLIP brute-force prompt reconstruction
      • The accuracy of this approach is too low for me to include it in imaginAIry
      • https://github.com/rmokady/CLIP_prefix_caption
      • https://github.com/pharmapsychotic/clip-interrogator (blip + clip)
    • https://github.com/KaiyangZhou/CoOp
  • ???? CPU support. While the code does actually work on some CPUs, the generation takes so long that I don’t think it’s
    worth the effort to support this feature
  • ✅ img2img for plms
  • ✅ img2img for kdiff functions
• Other
  • Enhancement pipelines
  • text-to-3d https://dreamfusionpaper.github.io/
    • https://shihmengli.github.io/3D-Photo-Inpainting/
    • thygate/stable-diffusion-webui-depthmap-script#50
    • Depth estimation
      • what is SOTA for monocular depth estimation?
      • https://github.com/compphoto/BoostingMonocularDepth
  • make a video https://github.com/lucidrains/make-a-video-pytorch
  • animations
    • https://github.com/francislabountyjr/stable-diffusion/blob/main/inferencing_notebook.ipynb
    • https://www.youtube.com/watch?v=E7aAFEhdngI
    • https://github.com/pytti-tools/frame-interpolation
  • guided generation
    • https://colab.research.google.com/drive/1dlgggNa5Mz8sEAGU0wFCHhGLFooW_pf1#scrollTo=UDeXQKbPTdZI
    • https://colab.research.google.com/github/aicrumb/doohickey/blob/main/Doohickey_Diffusion.ipynb#scrollTo=PytCwKXCmPid
    • https://github.com/mlfoundations/open_clip
    • https://github.com/openai/guided-diffusion
  • image variations https://github.com/lstein/stable-diffusion/blob/main/VARIATIONS.md
  • textual inversion
    • https://www.reddit.com/r/StableDiffusion/comments/xbwb5y/how_to_run_textual_inversion_locally_train_your/
    • https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb#scrollTo=50JuJUM8EG1h
    • https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_diffusion_textual_inversion_library_navigator.ipynb
    • https://github.com/Jack000/glid-3-xl-stable
  • fix saturation at high CFG https://www.reddit.com/r/StableDiffusion/comments/xalo78/fixing_excessive_contrastsaturation_resulting/
  • https://www.reddit.com/r/StableDiffusion/comments/xbrrgt/a_rundown_of_twenty_new_methodsoptions_added_to/
  • ✅ deploy to pypi
  • find similar images https://knn5.laion.ai/?back=https%3A%2F%2Fknn5.laion.ai%2F&index=laion5B&useMclip=false
  • https://github.com/vicgalle/stable-diffusion-aesthetic-gradients
• Training
  • Finetuning “dreambooth” style
  • Textual Inversion
  • Performance Improvements
  • ColoassalAI – almost got it working but it’s not easy enough to install to merit inclusion in imaginairy. We should check back in on this.
  • Xformers
  • Deepspeed

Notable Stable Diffusion Implementations

• https://github.com/ahrm/UnstableFusion
• https://github.com/AUTOMATIC1111/stable-diffusion-webui
• https://github.com/blueturtleai/gimp-stable-diffusion
• https://github.com/hafriedlander/stable-diffusion-grpcserver
• https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines/stable_diffusion
• https://github.com/lkwq007/stablediffusion-infinity
• https://github.com/lstein/stable-diffusion
• https://github.com/parlance-zz/g-diffuser-lib
• https://github.com/hafriedlander/idea2art

Online Stable Diffusion Services

• https://stablecog.com/

Further Reading

• Prompt Engineering Handbook
• Differences between samplers
  • https://www.reddit.com/r/StableDiffusion/comments/xbeyw3/can_anyone_offer_a_little_guidance_on_the/
• https://www.reddit.com/r/bigsleep/comments/xb5cat/wiskkeys_lists_of_texttoimage_systems_and_related/
• https://huggingface.co/blog/annotated-diffusion
• https://github.com/jessevig/bertviz
• https://www.youtube.com/watch?v=5pIQFQZsNe8
• https://jalammar.github.io/illustrated-transformer/
• https://huggingface.co/blog/assets/78_annotated-diffusion/unet_architecture.jpg