A text prompt weighting and blending library for transformers-type text embedding systems, by @damian0815.
With a flexible and intuitive syntax, you can re-weight different parts of a prompt string and thus re-weight the different parts of the embedding tensor produced from the string.
Tested and developed against Hugging Face's StableDiffusionPipeline but it should work with any diffusers-based system that uses an Tokenizer and a Text Encoder of some kind.
Adapted from the InvokeAI prompting code (also by @damian0815).
Note that cross-attention control .swap() is currently ignored by Compel, but you can use it by calling build_conditioning_tensor_for_prompt_object() yourself, and implementing cross-attention control in your diffusion loop.
pip install compel
Documentation is here.
with Hugging Face diffusers >=0.12:
from diffusers import StableDiffusionPipeline
from compel import Compel
pipeline = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
compel = Compel(tokenizer=pipeline.tokenizer, text_encoder=pipeline.text_encoder)
# upweight "ball"
prompt = "a cat playing with a ball++ in the forest"
conditioning = compel.build_conditioning_tensor(prompt)
# or: conditioning = compel([prompt])
# generate image
images = pipeline(prompt_embeds=conditioning, num_inference_steps=20).images
images[0].save("image.jpg")
For batched input, use the call interface to compel:
from diffusers import StableDiffusionPipeline
from compel import Compel
pipeline = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
compel = Compel(tokenizer=pipeline.tokenizer, text_encoder=pipeline.text_encoder)
prompts = ["a cat playing with a ball++ in the forest", "a dog playing with a ball in the forest"]
prompt_embeds = compel(prompts)
images = pipeline(prompt_embeds=prompt_embeds).images
images[0].save("image0.jpg")
images[1].save("image1.jpg")
If you want to have access to 🤗diffusers textual inversions, instantiate a DiffusersTextualInversionManager and pass it on Compel init:
pipeline = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
textual_inversion_manager = DiffusersTextualInversionManager(pipeline)
compel = Compel(tokenizer=pipeline.tokenizer, text_encoder=pipeline.text_encoder,
textual_inversion_manager=textual_inversion_manager)
If you run into memory issues, please make sure you're running compel inside with torch.no_grad(): blocks.
If this doesn't help, you could try this advice offered by @kshieh1:
After image generation, you should explictly de-reference the tensor object (i.e., prompt_embeds = None) and call gc.collect()
See https://github.com/damian0815/compel/issues/24 for more details. Thanks @kshieh1 !
pipeline.enable_sequential_cpu_offloading() with SDXL models (you need to pass device='cuda' on compel init).and()With big thanks to Patrick von Platen from Hugging Face for the pull request, Compel now supports SDXL. Use it like this:
from compel import Compel, ReturnedEmbeddingsType
from diffusers import DiffusionPipeline
import torch
pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", variant="fp16", use_safetensors=True, torch_dtype=torch.float16).to("cuda")
compel = Compel(tokenizer=[pipeline.tokenizer, pipeline.tokenizer_2] , text_encoder=[pipeline.text_encoder, pipeline.text_encoder_2], returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED, requires_pooled=[False, True])
# upweight "ball"
prompt = "a cat playing with a ball++ in the forest"
conditioning, pooled = compel(prompt)
# generate image
image = pipeline(prompt_embeds=conditioning, pooled_prompt_embeds=pooled, num_inference_steps=30).images[0]
Please note that this is a breaking change if you've been using clip skip: the old boolean arg use_penultimate_clip_layer has been replaced with an enum ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NORMALIZED.
.and() weights.and()For Stable Diffusion 2.1 I've been experimenting with a new feature: concatenated embeddings. What I noticed, for example, is that for more complex prompts image generation quality becomes wildly better when the prompt is broken into multiple parts and fed to OpenCLIP separately.
TL;DR: you can now experiment with breaking up your prompts into segments, which for SD2.1 appears to improve the generated image. The syntax is ("prompt part 1", "prompt part 2").and(). You can have more than one part, and you can also weight them, eg ("a man eating an apple", "sitting on the roof of a car", "high quality, trending on artstation, 8K UHD").and(1, 0.5, 0.5) which will assign weight 1 to man eating an apple and 0.5 to sitting on the roof of a car and high quality, trending on artstation, 8K UHD.
Here's a nonsense example from the InvokeAI discord #garbage-bin channel, created by gogurt enjoyer's incredible nightmare prompt generator:
a moist sloppy pindlesackboy sloppy hamblin' bogomadong, Clem Fandango is pissed-off, Wario's Woods in background, making a noise like ga-woink-a
Plugging this straight into SD2.1 we get this, which is really not a good image:
However, if the prompt is broken up into chunks and fed into OpenCLIP separately as four separate prompts, and then concatenated:
a moist sloppy pindlesackboy sloppy hamblin' bogomadong
Clem Fandango is pissed-off
Wario's Woods in background
making a noise like ga-woink-a
then output image with the same seed is so much better:
In the new .and() syntax you would prompt this as follows:
("a moist sloppy pindlesackboy sloppy hamblin' bogomadong", "Clem Fandango is pissed-off", "Wario's Woods in background", "making a noise like ga-woink-a").and()
The effect can be more or less subtle. Here for example is
A dream of a distant galaxy, by Caspar David Friedrich, matte painting, trending on artstation, HQ
And the same split into two parts:
A dream of a distant galaxy, by Caspar David Friedrich, matte painting
trending on artstation, HQ
The Compel prompt for this is:
("A dream of a distant galaxy, by Caspar David Friedrich, matte painting", "trending on artstation, HQ").and()
DiffusersTextualInversionManager (thanks @pdoane)To use, pass use_penultimate_clip_layer=True when initializing your Compel instance. Note that there's no need to pass this flag for SD2.0/SD2.1 because diffusers already throws away the last hidden layer when loading the SD2.0+ text encoder.
. characters inside parentheses)withLora/useLora on parse_prompt_string().Compel.parse_prompt_string() now returns a ConjunctionwithLora(name[, weight]) or useLora(name[, weight]) anywhere in the prompt string will be parsed to LoraWeight instances, and returned on the outermost Conjunction returned by parse_prompt_string().also fix test case for default swap parameters
cat.swap("")) when truncation is disabled--free_gpu_mem optionDownweighting now works by applying an attention mask to remove the downweighted tokens, rather than literally removing them from the sequence. This behaviour is the default, but the old behaviour can be re-enabled by passing downweight_mode=DownweightMode.REMOVE on init of the Compel instance.
Formerly, downweighting a token worked by both multiplying the weighting of the token's embedding, and doing an inverse-weighted blend with a copy of the token sequence that had the downweighted tokens removed. The intuition is that as weight approaches zero, the tokens being downweighted should be actually removed from the sequence. However, removing the tokens resulted in the positioning of all downstream tokens becoming messed up. The blend ended up blending a lot more than just the tokens in question.
As of v1.0.0, taking advice from @keturn and @bonlime (https://github.com/damian0815/compel/issues/7) the procedure is by default different. Downweighting still involves a blend but what is blended is a version of the token sequence with the downweighted tokens masked out, rather than removed. This correctly preserves positioning embeddings of the other tokens.
Also a bugfix: fix black images on weight 0 (https://github.com/invoke-ai/InvokeAI/issues/2832)
To enable, initialize Compel with truncate_long_prompts=False (default is True). Prompts that are longer than the model's max_token_length will be chunked and padded out to an integer multiple of max_token_length.
Note that even if you don't use a negative prompt, you'll need to build a conditioning tensor for a negative prompt of at least "", and use compel.pad_conditioning_tensors_to_same_length(), otherwise the you'll get an error about mismatched conditioning tensor lengths:
compel = Compel(..., truncate_long_prompts=False)
prompt = "a cat playing with a ball++ in the forest, amazing, exquisite, stunning, masterpiece, skilled, powerful, incredible, amazing, trending on gregstation, greg, greggy, greggs greggson, greggy mcgregface, ..." # very long prompt
conditioning = compel.build_conditioning_tensor(prompt)
negative_prompt = "" # it's necessary to create an empty prompt - it can also be very long, if you want
negative_conditioning = compel.build_conditioning_tensor(negative_prompt)
[conditioning, negative_conditioning] = compel.pad_conditioning_tensors_to_same_length([conditioning, negative_conditioning])
FAQs
A prompting enhancement library for transformers-type text embedding systems.
We found that compel demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Did you know?
Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.
Security News
Research
An advanced npm supply chain attack is leveraging Ethereum smart contracts for decentralized, persistent malware control, evading traditional defenses.
Security News
Research
Attackers are impersonating Sindre Sorhus on npm with a fake 'chalk-node' package containing a malicious backdoor to compromise developers' projects.
Security News
The npm package for the LottieFiles Player web component was hit with a supply chain attack after a software engineer's npmjs credentials were compromised.