Privacy Policy

1. Who we are

NSS Background Remover (“the tool,” “we,” “us”) is a product of Novus Stream Solutions, based in London, Ontario, Canada. This policy governs use of the tool at bgremover.novusstreamsolutions.com. For questions, see Section 9.

2. Media we cannot access

The editing architecture keeps the following content on your device. We do not collect, process on our servers, or have access to:

• The images or videos you upload or process
• The images or videos you export or download
• Filenames of any images or videos
• Media content entered into the brush editor, timeline, masks, caption tracks, or settings
• Your editing work — the tools are one-time-use and save nothing, not even locally in your browser

All AI processing — background removal, video processing, upscaling, filtering, and all other tool operations — runs exclusively in your browser using WebAssembly and WebGPU. There is no server that receives your image or video data. This is an architectural constraint, not merely a policy promise.

3. What we do process (minimally)

4. Third-party services

5. Your rights (PIPEDA)

Canada’s Personal Information Protection and Electronic Documents Act (PIPEDA) applies where required to personal information handled in our commercial activities. Other privacy laws may apply based on where you live or how a service provider processes data.

We intentionally limit personal information to the request, consent, analytics, advertising, and correspondence data described in this policy. You have the right to:

• Request confirmation of what personal information (if any) we hold about you
• Request correction of any inaccurate personal information
• Withdraw consent to analytics or advertising at any time via “Manage cookies” in the footer
• Clear all locally stored data by clearing your browser storage for this site

To exercise these rights or make a privacy complaint, contact our Privacy Officer at bgremover@novusstreamsolutions.com. We will acknowledge the request, investigate it, explain our decision, and describe any available review or appeal route. If our response does not resolve a PIPEDA concern, you may contact the Office of the Privacy Commissioner of Canada.

5a. Additional rights for EU/EEA users (GDPR)

If you are located in the European Union or European Economic Area, the General Data Protection Regulation (GDPR) grants you the following additional rights with respect to any personal data we process:

• Right of access (Art. 15): Request a copy of any personal data we hold about you and information about how we process it.
• Right to erasure (Art. 17): Request deletion of personal data we control, subject to legal exceptions. This may concern correspondence or data associated with analytics or advertising identifiers.
• Right to data portability (Art. 20): Receive any personal data you have provided to us in a structured, machine-readable format and transfer it to another controller.
• Right to object (Art. 21):Object to processing of your personal data for analytics or advertising at any time via “Manage cookies” in the footer.
• Right to restrict processing (Art. 18): Request restriction of processing in certain circumstances.
• Right to lodge a complaint: You have the right to lodge a complaint with your local supervisory authority (Data Protection Authority).

Our legal basis for processing analytics data (when consented) is your explicit consent (Art. 6(1)(a) GDPR). You may withdraw consent at any time. Our legal basis for infrastructure-level data processing (Vercel CDN logs) is legitimate interest (Art. 6(1)(f) GDPR) in delivering the Service securely.

To exercise these rights, contact bgremover@novusstreamsolutions.com. We will respond within 30 days.

5b. Additional rights for California residents (CCPA/CPRA)

California residents have the following rights under the California Consumer Privacy Act (CCPA) and its amendment (CPRA):

• Right to know: Request disclosure of the categories and specific pieces of personal information we have collected about you, the categories of sources, the business purpose, and the categories of third parties we share it with.
• Right to delete: Request deletion of personal information we have collected. The limited categories are described in Sections 3 and 4; requests will be acknowledged within 45 days.
• Right to opt-out of sale or sharing:We do not sell your personal information for money. However, the use of advertising cookies by Adsterra to serve personalized ads may be considered “sharing” for cross-context behavioral advertising under the CPRA. You can opt out of interest-based ads at aboutads.info and manage advertising cookies via “Manage cookies” in the footer; we honor Global Privacy Control (GPC) signals as a valid opt-out request.
• Right to non-discrimination: We will not discriminate against you for exercising your CCPA rights.

To submit a CCPA request, contact bgremover@novusstreamsolutions.com with the subject “CCPA Request.” We will respond within 45 days.

5c. AI processing disclosure

Model-powered processing on this Service occurs exclusively on your device using WebAssembly (WASM) and WebGPU/WebGL. No image, video, or audio data is transmitted to any AI inference server, cloud API, or external service. There is no endpoint of ours that receives your media, and adding the tools listed below created none. A model is downloaded only when a tool needs it; browser-only utilities that do not use a model download nothing.

Most managed model files are fetched from immutable Hugging Face URLs. The video background-removal, interactive-selection, and face-landmark assets are served from this site, and the separate optional WebLLM assistant uses its own model delivery. Those file requests carry ordinary network metadata such as an IP address and user agent, but never the image, video, audio, mask, prompt, or canvas content being processed.

A saved per-model permission is required for the large add-on models listed below, including face restoration, deblurring, image captioning, colourisation, 4× upscaling, LaMa inpainting, replacement-asset cut-out, and Whisper transcription. Background-removal weights instead load when you select that quality mode. Smaller task-specific assets, including interactive selection, face assistance, and privacy detection, load only after an explicit action such as Build AI mask, Detect, or Suggest framing. The tool shows the relevant trigger and download disclosure before it runs.

Current model-powered and local-analysis features:

• Background removal (/background-remover) — ORMBG Fast and BiRefNet Best Quality, with a local fallback. 45 MB Fast or up to 192 MB Best Quality depending on runtime. Selecting the quality mode requests that model; size and cache state are shown before or during the download. A compatible local model path is used when the requested runtime is unavailable.
• Video background removal (/video-background-remover) — RVM MobileNetV3. Approximately 15 MB, served from this site. Requested only when the video background-removal workflow runs. No cloud inference fallback.
• Subject selection — SlimSAM. Approximately 40 MB. Requested only when the subject-selection operation runs. Manual selection remains available.
• Upscaling (/upscale) — Swin2SR and optional Real-ESRGAN 4x. Swin2SR is approximately 47 MB; Real-ESRGAN is approximately 66 MB. Real-ESRGAN requires an explicit per-model yes/no choice. A two-pass Swin2SR cascade remains available.
• Face restoration (/tools/face-restore) — GFPGAN v1.4. Approximately 340 MB. Explicit per-model yes/no choice. A classical sharpening path runs when declined or unavailable.
• Deblurring (/tools/deblur) — NAFNet. Approximately 92 MB. Explicit per-model yes/no choice. A classical sharpening path runs when declined or unavailable.
• Image captioning and alt-text drafting (/tools/alt-text) — Local image-captioning model. Size is shown by the tool before consent. Explicit per-model yes/no choice. The tool reports that it cannot run; it does not fabricate a caption.
• AI colourisation (/tools/colorize) — DDColor. Size is shown by the tool before consent. Explicit per-model yes/no choice. Classical colour styles remain available.
• Video-editor object removal (/video-editor) — LaMa. Size is shown before consent. Explicit per-model yes/no choice. No model request occurs after a decline.
• Object Remove & Replace (/tools/remove-object) — Self-hosted MediaPipe MagicTouch selection, LaMa removal, and optional BiRefNet replacement-asset cut-out. About 17.2 MB for first-use selection, 200 MB for LaMa, and approximately 99 MB on WebGPU or up to 192 MB on WASM for replacement cut-out. Building an AI mask explicitly starts the selector; LaMa and replacement cut-out each require their own saved permission. Manual mask refinement and a limited neighbour-fill path remain available; already-transparent replacement assets need no BiRefNet download.
• Privacy Blur (/tools/privacy-blur) — Self-hosted MediaPipe face assistance and pinned Apache-2.0 RT-DETRv2 licence-plate detection. About 15 MB for face assistance or about 81 MB for licence-plate detection, plus potentially uncached shared ONNX Runtime assets. Nothing downloads until you press Detect for the selected detector; this action is not stored as a reusable model-consent decision. All detected regions remain editable, manual regions are always available, and the detector has a classical compatibility fallback.
• Smart Crop (/tools/smart-crop) — Self-hosted MediaPipe face assistance plus local contrast saliency. About 15 MB when face assistance is first requested; contrast saliency has no model download. The pinned face task loads only after you press Suggest framing. Local contrast and rule-of-thirds framing remains available if face assistance cannot run.
• Auto Subtitles (/tools/auto-subtitles) — Pinned Xenova Whisper Base q8. Approximately 82 MB. Explicit revision-specific approval is required before transcription. Declining or detecting no audio leaves the clip and caption track unchanged.
• Auto Highlights (/tools/auto-highlights) — No downloaded model; local histogram, motion, visual-prominence, RMS, and voice-activity scoring. No model download. Analysis begins only after you start the workflow. The review-first scene and range editor remains non-destructive until you apply an approved timeline.
• Optional local AI assistant — User-selected WebLLM language model. 500 MB, 800 MB, or 2.2 GB depending on the selected model. The separate assistant downloads only after the user presses Enable AI assistant. Deterministic commands and registered tool staging continue without WebLLM.

Model weights are downloaded once per browser profile, cached locally (see Section 3.2), and deletable anytime from Downloads & storage. They contain no personal data. You can verify on-device processing using your browser’s Network tab — after a model downloads, no requests containing your image, video, or audio data are made.

AI output is probabilistic and may be imperfect or contain errors. You are responsible for reviewing all AI-processed and AI-generated output before use.

5d. Data breach notification

Our architecture keeps source media away from our servers, but limited personal information may still exist in infrastructure logs, consented analytics or advertising, and correspondence. If a breach of security safeguards involving personal information occurs, we will:

• Investigate and contain the incident and assess the sensitivity and probability of misuse
• Report the breach to the Office of the Privacy Commissioner of Canada and notify affected individuals as soon as feasible when PIPEDA’s real risk of significant harm threshold is met
• Notify another organization or government institution when required to reduce or mitigate harm
• Keep a record of every breach of security safeguards for at least 24 months, whether or not notification is required

To report a security vulnerability, see our Security Policy.

6. Data retention

The tools are one-time-use and retain no session data: your images, masks, and edits exist only in memory while the tab is open and are discarded the moment you close or reload it. Nothing is written to our servers, and — apart from the cached model weights and preferences described in Section 3.2 — nothing is written to your browser either.

Cached model weights and consent preferences persist locally until you clear them via Downloads & storage, “Manage cookies” in the footer, or by clearing your browser’s site data. We have no ability to access or delete any of this on your behalf.

Vercel, Google, Hugging Face, and other disclosed delivery providers retain request, analytics, or advertising data under their own documented settings and retention schedules. We retain support, privacy, security, and legal correspondence only for as long as reasonably necessary to respond, maintain required records, resolve a dispute, or comply with law. Security-breach records are retained for at least 24 months.

7. Children’s privacy

This service is not directed to children under 13. We do not knowingly collect personal information from children. If you believe a child has provided personal information, contact us at bgremover@novusstreamsolutions.com and we will take appropriate steps.

8. Changes to this policy

We may update this policy periodically. Material changes will be noted in the version history below and communicated through an appropriate notice. If a change introduces a new purpose that requires consent, we will request that consent before the new processing begins; continued use alone is not treated as consent to a materially new purpose.

9. Privacy Officer and contact

The Privacy Officer for Novus Stream Solutions is accountable for this policy, privacy requests, and complaints. Contact the Privacy Officer at bgremover@novusstreamsolutions.com. For general questions, contact bgremover@novusstreamsolutions.com. Please do not send sensitive media or identity documents by ordinary email unless we have asked for them and provided suitable instructions.

Version history