damian0815/compel

A prompting enhancement library for transformers-type text embedding systems

What it solves

Compel provides a way to precisely control the influence of specific words or phrases in a text prompt for image generation models. Standard text encoders often treat all words with equal weight, making it difficult to emphasize or de-emphasize certain elements of an image without rewriting the entire prompt.

How it works

The library implements a flexible syntax for weighting and blending embeddings. By using markers like ++ to upweight a term, it modifies the embedding tensor produced by the text encoder. It supports advanced operations such as:

  • Weighting: Increasing or decreasing the importance of specific tokens.
  • Blending: Concatenating different prompt segments using a .and() syntax to improve image quality for complex prompts.
  • Style Prompts: Separate handling of style and content prompts for models like SDXL and Flux.
  • Long Prompt Handling: Chunking and padding prompts that exceed the model's maximum token length.

Who it’s for

It is designed for developers and researchers using diffusers-based systems (such as Stable Diffusion, SDXL, and Flux) who need fine-grained control over how their text prompts are interpreted by the model.

Highlights

  • Broad Model Support: Compatible with Stable Diffusion v1.5, SDXL, and Flux.
  • Intuitive Syntax: Uses a simple string-based weighting system (e.g., ball++).
  • Advanced Embedding Control: Supports negative prompts, style prompts, and textual inversion managers.
  • Flexible Tokenization: Offers multiple modes for splitting long text to avoid truncation.
  • Integration: Seamlessly integrates with Hugging Face's StableDiffusionPipeline and other diffusers pipelines.

Related

  • Dispatch
    Sentence Transformers v5.4: Multimodal Embedding and Reranker Models

    Sentence Transformers v5.4 introduces multimodal embedding and reranker models for encoding and comparing text, images, audio, and video.

  • Dispatch
    Hugging Face TextImage Augmentation pipeline release

    Hugging Face and Albumentations AI released a TextImage Augmentation pipeline that jointly modifies document images and their text, enabling realistic synthetic data generation and robust fine‑tuning of vision‑language models on limited document datasets.

  • Project
    huggingface/sentence-transformers

    A framework for computing embeddings, retrieval, and reranking using state-of-the-art dense, sparse, and multi-vector encoder models.

  • Project
    MinishLab/model2vec

    Model2Vec is an open‑source Python library that distills any sentence‑transformer into a tiny, static‑embedding model (8‑30 MB) that runs up to 500× faster on CPU. It offers zero‑data distillation, optional fine‑tuning for classification, and integrates with Hugging Face, Sentence‑Transformers, and LangChain. Install with `pip install model2vec`, load pre‑built models via `StaticModel.from_pretrained`, or distill your own in ~30 seconds on a CPU.

  • Project
    IBM/prompt-declaration-language

    A declarative YAML-based language for creating reliable, composable LLM prompts and integrating them into software systems with support for control flow and code execution.