PowerPaint¶

Development line: project:powerpaint · thread powerpaint
Last event: 2024-07-21 · 2 dated since 2023-12-12 · Researched: 2026-09-04 · confidence: medium

What it is¶

PowerPaint is an inpainting system based on Stable Diffusion for editing masked image areas.

• Text-guided object insertion to add elements from text prompts.
• Prompt-free object removal with optional negative prompts to erase objects.
• Image outpainting to extend canvas boundaries.
• Shape-guided object insertion to fit added elements to a mask outline.

Development line¶

• 2023-12-12 — PowerPaint became publicly available. On 2023-12-12, the project launched across its website, an MMagic implementation, and an OpenXLab application page.
• 2024-07-21 — Standalone OpenMMLab repository link. On 2024-07-21, PowerPaint pointed users to the standalone open-mmlab/PowerPaint repository. The links mark this move without showing code or model changes.

What changed¶

• 2023-12-12 — PowerPaint's paper introduced learnable task prompts so one model handles context filling, object insertion, removal, and shape guidance.
• 2023-12-18 — Stable v1 weights replaced the earlier build.
• 2023-12-22 — The maintainers fixed the ControlNet-loading error in the demo path.
• 2024-04-06 — Retraining with BrushNet produced PowerPaint v2 while keeping task-prompt cross-attention layers.
• 2024-04-07 — PowerPaint v2 code and weights were released.
• 2024-05-22 — PowerPaint v2-1 weights replaced initial v2 weights to fix training issues.
• 2024-07-21 — The public repository was available, but its changelog records no separate July 21 model release.

How to use this¶

From 2023-12-12, the project ran through its site, MMagic, and OpenXLab; from 2024-07-21, use the standalone OpenMMLab repository.

1. Clone the repository, create a Conda environment with Python 3.9, and install the pinned requirements. — https://github.com/open-mmlab/PowerPaint
2. Install Git LFS, then download the PowerPaint v2 checkpoint into a local checkpoint directory. — https://github.com/open-mmlab/PowerPaint
3. Start the v2 Gradio UI with python app.py --share --version ppt-v2 --checkpoint_dir checkpoints/ppt-v2. — https://github.com/open-mmlab/PowerPaint
4. Upload an image, paint the edit mask, pick the task tab, and supply text only for object insertion or shape guidance. — https://github.com/open-mmlab/PowerPaint
5. Leave the text prompt empty for removal or outpainting; set horizontal and vertical expansion ratios for outpainting. — https://github.com/open-mmlab/PowerPaint

Best practices¶

• Leave the prompt empty for object removal; add a negative prompt and raise Guidance Scale to 10 or higher if unwanted objects appear. — https://github.com/open-mmlab/PowerPaint
• Set fitting degree to 0.5–0.6 when the object can diverge from the mask, and 0.8–0.95 to match the mask shape closely. — https://github.com/open-mmlab/PowerPaint
• Use v2-1 weights instead of the initial v2 release, as maintainers shipped v2-1 specifically to correct training issues. — https://github.com/open-mmlab/PowerPaint

Superseded by this¶

• 2023-12-22 — Early ControlNet demo guidance is obsolete because the loading logic was fixed.
• 2024-05-22 — Initial PowerPaint v2 weights are superseded by v2-1, which corrected training issues.

Still unknown¶

• The repository has no tagged release for a distinct 2024-07-21 release; that date may reflect a repository move rather than a model update.
• Documentation covers local setup and the Gradio UI, but omits GPU memory limits, operating system support, and Diffusers compatibility.
• Neither the original project page nor the OpenXLab demo endpoint returned usable technical details during this check.

Sources¶

Agent brief¶

• Subject: project:powerpaint, thread powerpaint, 2 dated events 2023-12-12 → 2024-07-21.
• Practical note: From 2023-12-12, practitioners could find PowerPaint through its site, MMagic, and OpenXLab; from 2024-07-21, consult the standalone OpenMMLab repository.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.