Opening a file, resizing, saving PNG/JPG, creating the four-file ZIP, and ordinary SVG tracing are local browser operations. An AI request begins only when you choose “Run AI edit” or “AI-assisted SVG.”
Separate local processing from AI processing
| Task | Default processing location | Image sent externally | Notes |
|---|---|---|---|
| Open PNG/JPG/WebP | Current browser | No | Decoded into a canvas |
| Open PDF or PDF-compatible AI file | Current browser | No | First page rendered with PDF.js |
| Resize, watermark, or create ZIP | Current browser | No | Output files are also created on the device |
| Ordinary SVG conversion | Current browser | No | Paths generated with ImageTracer |
| AI image editing | Browser + DrowPack Worker + AI provider | Yes | Current canvas and prompt are sent |
| AI-assisted SVG | AI simplifies the image, then the browser converts it to paths | Yes | The AI output becomes the input for SVG tracing |
The phrase “AI file” can have two meanings here. Opening an Adobe Illustrator .ai file through its PDF-compatible data is not a generative-AI request. A generative image model is used only when you run a separate AI editing command.
The actual AI editing request flow
- You enter an instruction and run it. An empty prompt is not sent; the current server limit is 1,200 characters.
- The current preview canvas becomes a PNG. When its long edge exceeds 511px, it is reduced with aspect-ratio-preserving, high-quality resampling.
- The request goes to the same-origin DrowPack API. The Worker checks the request format, image type and size, and origin, and applies a per-minute rate limit.
- The configured provider performs inference. The current deployment uses an image editing model on Cloudflare Workers AI.
- The returned image is displayed on the canvas. You review the result, then decide whether to save it or use ordinary or AI-assisted SVG conversion.
The 511px limit applies to the copy used for the AI request; it does not overwrite the dimensions of your source file. Once an AI result is applied to the canvas, later saves and SVG conversions use that result. This is also why tiny lettering and precise logos may change after AI editing.
Current provider and model
On July 19, 2026, the live AI status endpoint reported provider cloudflare, model @cf/black-forest-labs/flux-2-klein-4b, and available: true. Deployment settings can change, so check the current status response before relying on the feature.
- Cloudflare Workers AI model documentation
- Cloudflare Workers AI data-use documentation
- Cloudflare Privacy Policy
Cloudflare's current documentation describes Workers AI inputs and outputs as Customer Content and says that, without explicit consent, it does not use that content to train models or improve Cloudflare or third-party services. Models supplied by third parties can have additional licensing conditions, so the provider's latest documentation takes priority.
Storage and logging boundaries
DrowPack's AI processing code has no step that stores an uploaded source or AI result in R2, KV, Durable Objects, or a separate application database. API responses also use Cache-Control: no-store. This does not mean no external system processes the data: Cloudflare still handles the request as part of network and AI service operations.
Do not put data into the AI feature when external AI processing would be inappropriate, including identity documents, bank or address details, medical information, confidential contracts, or faces used without consent. Use local features only, or remove the sensitive areas first.
What AI does not guarantee
An image editing model interprets an instruction and generates new pixels. Even when the input and output appear semantically similar, every element is not guaranteed to remain exact.
- Faces, fingers, gaze direction, and body proportions
- Logo geometry, trademark shapes, and QR codes
- Sentences, numbers, signatures, and small type
- Color values, transparency, and line weight
- Permission to use the result, including copyright, publicity, and trademark rights
- Factual accuracy or an exact depiction of real people and events
Overlay the source and AI result or enlarge faces, hands, text, and logos for review. Keep an independent copy of the source before important edits. Treat AI output as a suggestion or draft; responsibility for final publication remains with the user.
Responsible-use standards
- Use only images you created or have permission to edit
- When editing an identifiable person, confirm consent and context
- Do not use the feature for impersonation, fraud, harassment, or defamation
- Do not create non-consensual sexual imagery, child sexual abuse material, or illegal content
- Do not distribute manipulated political or social-event imagery as authentic evidence
- Disclose AI use where hiding it would create a misleading impression
See the DrowPack Terms of Use for further prohibited uses and rights responsibilities.
Provenance markers in AI-assisted SVG
The AI-assisted SVG flow retraces an AI-simplified image into paths. To distinguish it from an ordinary vector result, DrowPack names the file artwork_ai_vector.svg and records data-drowpack-ai-assisted="true" on the SVG root. When model information is available, it is placed in data-drowpack-ai-model, and internal metadata records the pipeline type.
These markers are not signatures certifying quality or rights. They are provenance information that helps you determine which processing flow created a file and prevents it from being confused with an ordinary SVG result.
Checklist before running AI
- Does this task truly require AI editing, or are local resize and SVG conversion sufficient?
- Does the image contain personal information that should not go to an external service?
- Do you have the rights to edit and distribute the image and any identifiable people in it?
- Have you avoided passwords, contact details, and confidential information in the prompt?
- Have you kept a separate copy of the source?
- Do you plan to review faces, hands, text, logos, and colors in the result yourself?