# adobe-research/custom-diffusion

Custom Diffusion: Multi-Concept Customization of Text-to-Image Diffusion (CVPR 2023)

Repository: https://github.com/adobe-research/custom-diffusion
Canonical: https://ross.abutalabs.com/products/custom-diffusion
Homepage: https://www.cs.cmu.edu/~custom-diffusion
Language: Python
License: NOASSERTION
License Family: other
Topics: customization, fine-tuning, text-to-image-generation, computer-vision, diffusion-models, few-shot, pytorch
Last push: 2026-05-24T19:31:46+00:00

## Health v2 (maintenance only)
Score: 69/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 84, release rhythm 35, longevity 97
- inputs: {"age_days": 1364, "days_push": 101, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1978, forks 140 (observed 2026-08-28T04:06:01.566516+00:00)

## What it is
Custom Diffusion is a research codebase for efficiently fine-tuning text-to-image diffusion models like Stable Diffusion on a few example images (4-20) of a new concept. It updates only key and value projection matrices in cross-attention layers, enabling fast training (~6 minutes on 2 A100 GPUs), small per-concept storage (75MB), and composition of multiple customized concepts.

## Use cases
- teach stable diffusion to generate images of my pet from a few photos
- fine-tune a text-to-image model on a custom artistic style
- combine multiple learned concepts like an object plus a style in one generation
- create images of a personal toy or product in new settings
- evaluate concept customization methods with the CustomConcept101 dataset
- train a lightweight adapter for stable diffusion with low storage cost

## When to choose
- you want to personalize Stable Diffusion with only a handful of example images
- you need fast fine-tuning with minimal GPU time and small per-concept checkpoints
- you want to compose multiple custom concepts (objects, styles, categories) in one model
- you need a research baseline or dataset for comparing personalization methods

## When to avoid
- you need a production-grade, well-supported training pipeline rather than research code
- you have no GPU resources available
- you want to fine-tune models other than Stable Diffusion-family diffusion models
- you prefer a maintained integration - in that case use the Custom Diffusion implementation in Hugging Face diffusers instead

## Facets
- artifact type: library
- maturity: stable
- function: machine-learning, deep-learning, llm-training, image-processing
- domain: machine-learning, computer-vision, image-processing, artificial-intelligence
- platform: python
- tags: diffusion-models, stable-diffusion, fine-tuning, text-to-image, few-shot-learning, customization, pytorch, cvpr-2023, research-code, gpu, linux

## Member repositories
- adobe-research/custom-diffusion (main) score 69

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:01.566516+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:04:44.000175+00:00, confidence not recorded.
  - readme: https://github.com/adobe-research/custom-diffusion (fetched 2026-08-28T04:06:01.566516+00:00, sha 2cbca50bb7e8)
  - homepage: https://www.cs.cmu.edu/~custom-diffusion (fetched 2026-08-29T10:43:43.105186+00:00, sha 03139f345f1a)
- Data as of 2026-08-30T08:39:29.467469+00:00.