DragDiffusion
Interactive point-based image editing for diffusion models: click handle and target points and drag content to a new position.
About
DragDiffusion (arXiv 2306.14435, 2023) brings DragGAN-style drag editing to diffusion models. The workflow: train a small LoRA on the input image, click handle points and matching target points, optionally paint a mask limiting the editable region, and run the algorithm. It optimizes the diffusion latent at an intermediate timestep under motion supervision from UNet features, moving the handle content to the targets while the rest of the image stays put. The authors also released DragBench, the first benchmark for interactive point-based editing.
The code is on GitHub from the NTU team with a Gradio UI, and it works with Stable Diffusion checkpoints. It is a research implementation rather than a product: per-image LoRA training makes each edit slow, results are sensitive to hyperparameters like the optimized timestep, and a 2026 reproducibility study confirmed the core claims while noting that fragility. There is no official API or hosted demo from the authors.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported results on the DragBench benchmark (arXiv 2306.14435); a 2026 reproducibility study confirmed main claims with caveats about hyperparameter sensitivity
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 2306.14435 with code at github.com/Yujun-Shi/DragDiffusion and DragBench benchmark; reproducibility study (2026) corroborates main claims while flagging hyperparameter sensitivity. Strengths: novel capability for diffusion models, public code and benchmark, independent replication attempt. Docked for: (1) no dedicated model weights released (technique applied to SD checkpoints); (2) per-image LoRA training makes it slow and fiddly; (3) no official API or maintained product surface; repo license terms unverified. License: repo license unverified.
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
- 6.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · arXiv 2306.14435 (2023), NTU team; code at github.com/Yujun-Shi/DragDiffusion
- · DragBench benchmark released alongside
- · 2026 reproducibility study confirmed main claims with hyperparameter-sensitivity caveats
- · Per-image LoRA training; no official API or hosted demo