“Cloud Safety” and “Local Safety” refer to two fundamentally different operational and governance architectures within Adobe’s creative ecosystem. Cloud Safety governs features like Generative Fill, Text to Image, and Firefly Web, where prompts and active pixels are transmitted to external Adobe data centers and evaluated by automated content moderation filters. Local Safety refers to on-device machine learning tools—such as the native Remove Tool, local Neural Filters, and smart selection algorithms—which execute entirely on your workstation’s local hardware without transmitting data over the internet or applying content moderation restrictions.
Fast Fix
Cloud Safety runs on Adobe’s external servers, scanning prompts and pixels for policy, NSFW, and copyright violations while consuming monthly generative credits. Local Safety processes tasks entirely on your computer’s local GPU, CPU, or NPU, consuming zero credits, operating without internet connectivity, and running completely free of content moderation filters.
Feature Architecture Comparison
Understanding where image processing takes place helps you choose the right tool for sensitive, confidential, or frequently flagged creative work:
| Operational Dimension | Cloud Safety (Generative Fill / Firefly) | Local Safety (On-Device Remove Tool / Local ML) |
|---|---|---|
| Execution Hardware | External Adobe Cloud GPU Clusters | Local Machine Hardware (GPU, NPU, CPU) |
| Moderation Engine | Automated Multi-Tier Cloud Classifiers | None (User maintains complete local autonomy) |
| Data Transmission | Uploads active selection + context buffer | Zero data leaves the workstation |
| Internet Requirement | Mandatory continuous high-speed connection | Works completely offline |
| Generative Credits | Consumes allocated monthly balance | Zero credits consumed |
| Regulatory Fit | Subject to commercial cloud safety policies | Complies with strict NDAs, HIPAA, and offline air-gaps |
| Metadata Tracking | Attaches C2PA Content Credentials automatically | Standard document metadata (unless configured) |
How Cloud Safety Operates
When you use a cloud-powered tool in Photoshop, your computer acts as a client interface communicating with a remote server:
Cloud Processing Pipeline
┌──────────────────┐ Encrypted HTTPS ┌─────────────────────────────┐
│ Your Workstation │ ────────────────────────► │ Adobe Cloud Servers │
│ (Prompt + Pixels)│ │ │
│ │ │ 1. Text Moderation Filter │
│ │ │ 2. Computer Vision Filter │
│ │ │ 3. Commercial IP Filter │
│ │ │ 4. Diffusion Generation Core│
│ │ ◄──────────────────────── │ │
│ Rendered Output │ Raster Variations └─────────────────────────────┘
└──────────────────┘
Because Adobe provides commercial indemnification for enterprise Firefly assets, the cloud engine must strictly prevent the generation of harmful, explicit, or legally actionable material.
This multi-layer safety check automatically evaluates:
- Text Strings: Blacklisted words, celebrity names, brand trademarks, and suggestive terms.
- Canvas Pixels: Skin-tone distributions, contrast curves, and red hues that could represent violence or exposed anatomy. For deeper insight into why natural lighting trips these filters, read Why Shadows and Curves Trigger the NSFW Filter.
- Reference Ingestion: Uploaded images checked for watermarks and protected design marks.
When any of these tests fail, the server rejects the request, returning a “Content Violation” alert to your screen.
How Local Processing Operates
Local machine learning tools execute entirely within your computer’s local memory and graphics pipeline:
Local Processing Pipeline
┌────────────────────────────────────────────────────────────────────────────┐
│ Your Workstation (Air-Gapped / Offline Capable) │
│ │
│ [ Active Canvas ] ──► [ Local VRAM / NPU ] ──► [ Local Model Execution ] │
│ │
│ *No network packets sent. No content classifiers. No remote logging.* │
└────────────────────────────────────────────────────────────────────────────┘
Because your workstation is private property and no data is broadcast to shared infrastructure, Adobe does not enforce content moderation filters on local execution.
This makes local processing essential for several professional workflows:
- Medical and Forensic Visualization: Clinical photographs involving surgery, pathology, or trauma cannot pass cloud safety filters due to bodily harm policies. Local execution allows medical professionals to edit patient assets safely without policy violations. See Why “Medical” Prompts are Strictly Filtered in 2026.
- Confidential and Air-Gapped Work: Defense contractors, financial institutions, and entertainment studios often operate under non-disclosure agreements that forbid uploading unreleased assets to third-party cloud infrastructure.
- High-End Fashion and Swimwear Retouching: Eliminates false-positive rejections caused by skin folds, lingerie, or tight athletic apparel.
Diagnostic Path: Choosing the Right Engine
Selecting Processing Mode
│
├── Do you need to create entirely new objects or expand canvas dimensions?
│ ├── YES ──► Use Cloud Generative Fill (Requires compliant prompts/pixels).
│ │
│ └── NO ──► Are you removing blemishes, wrinkles, props, or artifacts?
│ │
│ └── YES ──► Use Local Processing (Remove Tool).
│ Bypasses all safety filters and uses zero credits.
To configure your workspace so the Remove Tool always executes on your internal graphics card rather than defaulting to cloud services, follow the instructions in How to Use the “Remove Tool” Locally When Gen-Fill is Blocked.
If you need to optimize memory limits or resolve performance lag when running local models on dedicated GPUs versus Copilot+ NPUs, review How to Switch Between “Cloud” and “Local” AI Engines.
Enterprise Governance and Metadata Implications
Understanding this architecture is also critical for compliance and legal asset delivery:
- Content Authenticity (C2PA): Assets created or modified through Cloud Generative Fill automatically embed provenance metadata recording that synthetic AI tools were applied to the document. To understand how this impacts delivery to clients and social platforms, consult How to Use the “C2PA” Standard in Professional Photography.
- Local Tool Provenance: Edits performed with local tools set to in-device processing do not trigger synthetic asset flags in the same way, as they rely on traditional spatial interpolation and localized contextual machine learning.