Fork: 192 Star: 4642 (更新于 2023-01-25 09:42:22)

license: MIT , Language: Python.

AI imagined images. Pythonic generation of stable diffusion images.


ImaginAIry 🤖🧠

Downloads image image Code style: black Python Checks

AI imagined images. Pythonic generation of stable diffusion images.

"just works" on Linux and macOS(M1) (and maybe windows?).


# on macOS, make sure rust is installed first
>> pip install imaginairy
>> imagine "a scenic landscape" "a photo of a dog" "photo of a fruit bowl" "portrait photo of a freckled woman"
# Stable Diffusion 2.1
>> imagine --model SD-2.1 "a forest"
Console Output
🤖🧠 received 4 prompt(s) and will repeat them 1 times to create 4 images.
Loading model onto mps backend...
Generating 🖼  : "a scenic landscape" 512x512px seed:557988237 prompt-strength:7.5 steps:40 sampler-type:PLMS
    PLMS Sampler: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40/40 [00:29<00:00,  1.36it/s]
    🖼  saved to: ./outputs/000001_557988237_PLMS40_PS7.5_a_scenic_landscape.jpg
Generating 🖼  : "a photo of a dog" 512x512px seed:277230171 prompt-strength:7.5 steps:40 sampler-type:PLMS
    PLMS Sampler: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40/40 [00:28<00:00,  1.41it/s]
    🖼  saved to: ./outputs/000002_277230171_PLMS40_PS7.5_a_photo_of_a_dog.jpg
Generating 🖼  : "photo of a fruit bowl" 512x512px seed:639753980 prompt-strength:7.5 steps:40 sampler-type:PLMS
    PLMS Sampler: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40/40 [00:28<00:00,  1.40it/s]
    🖼  saved to: ./outputs/000003_639753980_PLMS40_PS7.5_photo_of_a_fruit_bowl.jpg
Generating 🖼  : "portrait photo of a freckled woman" 512x512px seed:500686645 prompt-strength:7.5 steps:40 sampler-type:PLMS
    PLMS Sampler: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40/40 [00:29<00:00,  1.37it/s]
    🖼  saved to: ./outputs/000004_500686645_PLMS40_PS7.5_portrait_photo_of_a_freckled_woman.jpg

🎉 Edit Images with Instructions alone! by InstructPix2Pix

Just tell imaginairy how to edit the image and it will do it for you!
Use prompt strength to control how strong the edit is. For extra control you can combine with prompt-based masking.

>> aimg edit scenic_landscape.jpg "make it winter" --prompt-strength 20
>> aimg edit dog.jpg "make the dog red" --prompt-strength 5
>> aimg edit bowl_of_fruit.jpg "replace the fruit with strawberries"
>> aimg edit freckled_woman.jpg "make her a cyborg" --prompt-strength 13
>> aimg edit pearl_girl.jpg "make her wear clown makup"
>> aimg edit mona-lisa.jpg "make it a color professional photo headshot" --negative-prompt "old, ugly"

Want just quickly have some fun? Try --surprise-me to apply some pre-defined edits.

>> aimg edit --gif --surprise-me pearl_girl.jpg
>> aimg edit --gif --surprise-me mona-lisa.jpg
>> aimg edit --gif --surprise-me luke.jpg
>> aimg edit --gif --surprise-me spock.jpg

Prompt Based Masking by clipseg

Specify advanced text based masks using boolean logic and strength modifiers. Mask syntax:

  • mask descriptions must be lowercase
  • keywords (AND, OR, NOT) must be uppercase
  • parentheses are supported
  • mask modifiers may be appended to any mask or group of masks. Example: (dog OR cat){+5} means that we'll select any dog or cat and then expand the size of the mask area by 5 pixels. Valid mask modifiers:
    • {+n} - expand mask by n pixels
    • {-n} - shrink mask by n pixels
    • {*n} - multiply mask strength. will expand mask to areas that weakly matched the mask description
    • {/n} - divide mask strength. will reduce mask to areas that most strongly matched the mask description. probably not useful

When writing strength modifiers keep in mind that pixel values are between 0 and 1.

>> imagine \
    --init-image pearl_earring.jpg \ 
    --mask-prompt "face AND NOT (bandana OR hair OR blue fabric){*6}" \
    --mask-mode keep \
    --init-image-strength .2 \
    --fix-faces \
    "a modern female president" "a female robot" "a female doctor" "a female firefighter"


>> imagine \
    --init-image fruit-bowl.jpg \
    --mask-prompt "fruit OR fruit stem{*6}" \
    --mask-mode replace \
    --mask-modify-original \
    --init-image-strength .1 \
    "a bowl of kittens" "a bowl of gold coins" "a bowl of popcorn" "a bowl of spaghetti"


Face Enhancement by CodeFormer

>> imagine "a couple smiling" --steps 40 --seed 1 --fix-faces


Upscaling by RealESRGAN

>> imagine "colorful smoke" --steps 40 --upscale


Tiled Images

>> imagine  "gold coins" "a lush forest" "piles of old books" leaves --tile

360 degree images

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


Use depth maps for amazing "translations" of existing images.

>> imagine --model SD-2.0-depth --init-image girl_with_a_pearl_earring_large.jpg --init-image-strength 0.05  "professional headshot photo of a woman with a pearl earring" -r 4 -w 1024 -h 1024 --steps 50



Given a starting image, one can generate it's "surroundings".

Example: imagine --init-image pearl-earring.jpg --init-image-strength 0 --outpaint all250,up0,down600 "woman standing" ➡️

Prompt Expansion

You can use {} to randomly pull values from lists. A list of values separated by | and enclosed in { } will be randomly drawn from in a non-repeating fashion. Values that are surrounded by _ _ will pull from a phrase list of the same name. Folders containing .txt phraselist files may be specified via --prompt_library_path. The option may be specified multiple times. Built-in categories:

  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


imagine "a {lime|blue|silver|aqua} colored dog" -r 4 --seed 0 (note that it generates a dog of each color without repetition)

imagine "a {_color_} dog" -r 4 --seed 0 will generate four, different colored dogs. The colors will be pulled from an included phraselist of colors.

imagine "a {_spaceship_|_fruit_|hot air balloon}. low-poly" -r 4 --seed 0 will generate images of spaceships or fruits or a hot air balloon

Credit to noodle-soup-prompts where most, but not all, of the wordlists originate.

Generate image captions (via BLIP)

>> aimg describe assets/mask_examples/bowl001.jpg
a bowl full of gold bars sitting on a table


  • 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 <filename-or-url>
  • 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 = ""
prompts = [
    ImaginePrompt("a scenic landscape", seed=1, upscale=True),
    ImaginePrompt("a bowl of fruit"),
        WeightedPrompt("cat", weight=1),
        WeightedPrompt("dog", weight=1),
        "a spacious building", 
        "a bowl of strawberries", 
        mask_prompt="fruit OR stem{*2}",  # amplify the stem mask x2
    ImaginePrompt("strawberries", tile_mode=True),
for result in imagine(prompts):
    # do something"my_image.jpg")

# or

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


  • ~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 | 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



  • feature: image sizes can now be multiples of 8 instead of 64. Inputs will be silently rounded down.


  • fix: bypass huggingface cache retrieval bug


  • fix: limit attention slice size on MacOS machines with 64gb (#175)


  • fix: use python 3.7 compatible lru_cache
  • fix: use windows compatible filenames


  • fix: hf_hub_download() got an unexpected keyword argument 'token'


  • fix: spelling mistake of "surprise"


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


  • 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


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


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


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


  • 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


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


  • 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


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


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


  • 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.


  • fix: add back in memory-efficient algorithms


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


  • 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 512x512 model
    • New inpainting model working
    • 768x768 model working for all samplers except PLMS (--model SD-2.0-v )


  • feature: add progress image callback


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


  • 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


  • 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)


  • 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


  • fix: k-samplers were broken


  • feature: improved safety filter


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


  • fix: face fidelity default was broken


  • 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


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


  • feature: face enhancement fidelity is now configurable



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


  • 🎉 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


  • fix: another bfloat16 fix


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


  • 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


  • fix: missing config file for describe feature


  • 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


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


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


  • allow urls as init-images


  • 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


Notable Stable Diffusion Implementations

Online Stable Diffusion Services

Further Reading

最近版本更新:(数据更新于2023-01-25 09:42:27)


2023-02-05 11:08:25 acheong08/ChatGPT

2023-02-05 08:17:20 LAION-AI/Open-Assistant

2023-02-04 12:26:13 Zero6992/chatGPT-discord-bot

2023-02-04 06:50:19 hwchase17/langchain

2023-02-03 11:26:07 kovidgoyal/calibre

2023-02-02 01:27:59 sqlalchemy/sqlalchemy