BrushNet
Plug-and-play diffusion inpainting adapter that injects masked-image features into any pretrained diffusion model via a dual-branch design.
About
BrushNet (ECCV 2024, arXiv 2403.06976) is Tencent ARC Lab's approach to diffusion-based inpainting. Instead of fine-tuning the whole base model, it adds a parallel branch that processes the masked image and noisy latent separately and injects dense per-pixel control into a frozen pretrained diffusion model. The decomposed design cuts the learning load and keeps the base model's generative power intact, which is why it is described as plug-and-play: the same adapter concept can ride on top of different backbones, including SDXL.
The code is public at github.com/TencentARC/BrushNet with a Hugging Face demo space, and PowerPaint v2 adopted BrushNet as its foundation. The repository license is listed as a custom Tencent license rather than a standard open-source one, so commercial reuse needs a careful read. There is no official hosted API; it is a research release for self-hosting and experimentation.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported results in the ECCV 2024 paper (arXiv 2403.06976); no independent benchmark scores found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics tier, metrics-v1). Evidence: paper arXiv 2403.06976 (Tencent ARC Lab, CUHK), official repo TencentARC/BrushNet (~1.7k stars), Hugging Face demo space linked from project docs. Strengths: elegant adapter design, works with frozen base models, adopted by PowerPaint v2. Docked for: (1) custom Tencent license, not a standard open-source license, so commercial terms are unclear; (2) benchmarks are author-reported only; (3) no official API or maintained hosted product; (4) modest community maintenance. License: custom Tencent license (not open-source commercial).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 5.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · ECCV 2024, arXiv 2403.06976 (Tencent ARC Lab, CUHK); repo TencentARC/BrushNet ~1.7k stars
- · Dual-branch adapter design over frozen base models; SDXL support
- · Adopted by PowerPaint v2 as its foundation
- · Benchmarks author-reported only; no official API