A number of freely available models are used by ResourceSpace for AI functions such as image classification, image recognition and text transcription. These models are all open-source and widely adopted in both research and industry. Each has been integrated “as is,” without retraining, and is deployed within our production environment or the customer’s own environment to ensure data remains under their control. The following register provides provenance and transparency information for each model, including its purpose in ResourceSpace, licensing, known limitations, and relevant compliance considerations.

InsightFace ("Faces" Plugin)

Model name
buffalo_l (InsightFace model pack)

Provider / source
InsightFace (open-source) with commercial licensing available.

Purpose in ResourceSpace
Face feature extraction and matching only within the customer’s own corpus (closed set). No external gallery is used or shipped. 

Training data sources (upstream)
InsightFace models are commonly trained on MS1M-ArcFace (a cleaned/refined version of MS-Celeb-1M) and Glint360K; Glint360K itself was built by cleaning large web-scraped celebrity/public-figure datasets (MS-Celeb-1M and Celeb-500K). While InsightFace’s docs don’t publish a per-pack recipe for buffalo_l, their papers and dataset notes indicate MS1M and Glint360K as fixed training sets for their recognition baselines. 

Data rights / licensing (upstream context)

  • InsightFace library/models are distributed under permissive licenses; commercial licenses are also offered. 

  • Upstream datasets (MS-Celeb lineage, Glint360K) were assembled from internet images; they have been criticized for consent/rights issues, and MS-Celeb-1M has been withdrawn. These concerns apply to dataset provenance in the ecosystem, not to ResourceSpace’s operation (which does not identify celebrities).

Known limitations / risks
Typical FR risks apply (demographic performance variation, false matches under poor quality). Dataset lineage implies potential bias reflecting web content; this is broadly documented across FR systems trained on MS1M/Glint-style corpora.

Bias considerations
Advise human verification for high-impact uses; monitor accuracy across demographic groups.

ResourceSpace implementation notes

  • No fine-tuning; buffalo_l used as released.

  • Closed-set matching only: embeddings compared against faces in the customer’s own repository; no prebuilt or third-party identity gallery; no “celebrity mode.”

  • You can disable or scope face features per collection to align with policy.

Compliance notes (GDPR / EU AI Act)

  • Facial recognition for identity/matching is typically high-risk under the EU AI Act; customers must establish a lawful basis (e.g., explicit consent or another valid ground) and conduct DPIAs where appropriate.

  • Since ResourceSpace processes only customer-held images and does not perform open-world/celebrity identification, risk is reduced relative to public-identity services, but governance steps (purpose limitation, access controls, retention) still apply.

  • Provide users with notice that automated similarity is in use and requires human review.

References
InsightFace project & model packs; Glint360K/MS1M lineage and usage in InsightFace papers and docs.

References / documentation links:
InsightFace documentation

CLIP

Model name
ViT-B/32 (CLIP – Contrastive Language-Image Pretraining)

Provider / Source
OpenAI (https://github.com/openai/CLIP)

Version / Release date
CLIP ViT-B/32, released January 2021

Purpose in ResourceSpace
Generates semantic embeddings for images and text to enable features like automated keywording, similarity search, and natural language search queries.

Training data sources
Trained on ~400 million image–text pairs collected from the internet. Specific datasets have not been disclosed; assumed broad web-scraped content.

Data rights / licensing
Weights released under MIT license. Training data sources were scraped without explicit licensing or consent.

Known limitations / risks

  • May embed internet biases (cultural, gender, racial stereotypes).

  • Can produce incorrect or offensive labels if unchecked.

Bias considerations
Tends to reflect patterns and biases of internet content; outputs must be treated as probabilistic suggestions, not factual assertions.

ResourceSpace modifications
Model used directly with no fine-tuning. Preprocessing pipeline unchanged.

Compliance notes (GDPR, EU AI Act, etc.):

  • Considered limited risk under EU AI Act when used for metadata enrichment (not biometric).

  • Must provide transparency to users that tags are AI-generated.

  • Customers should review auto-generated metadata before publication.

References / documentation links:
OpenAI CLIP repository
OpenAI CLIP website

Whisper

Model name
Whisper Turbo (OpenAI speech-to-text model)

Provider / Source
OpenAI (https://github.com/openai/whisper)

Version / Release date
Whisper Turbo (released 2023; optimized faster inference variant of Whisper)

Purpose in ResourceSpace
Performs automatic speech recognition (ASR) on audio/video files to generate transcripts or subtitles, enabling text search and accessibility features.

Training data sources
Trained on ~680,000 hours of multilingual and multitask supervised data collected from the web (publicly available audio with corresponding transcripts).

Data rights / licensing
Weights released under the MIT License. Training data includes large amounts of web-scraped audio, not fully rights-cleared or consented. No private customer data used in training.

Known limitations / risks

  • Accuracy depends on language, accent, and recording quality.

  • May struggle with background noise, overlapping speakers, or domain-specific terminology.

  • Transcripts may include errors that alter meaning (mis-transcriptions).

Bias considerations

  • Performs better on widely represented languages and accents.

  • Lower accuracy expected for underrepresented languages and dialects.

  • Could reinforce biases present in the training corpus.

ResourceSpace modifications:
No fine-tuning. Model used as provided by OpenAI, via command-line interface ("whisper --model turbo"). Only transcription, not translation, enabled by default.

Compliance notes (GDPR, EU AI Act, etc.):

  • Considered limited risk under the EU AI Act when used for transcription.

  • Customers must ensure lawful basis for processing recorded speech (GDPR).

  • ResourceSpace does not retain transcribed audio unless stored by the customer.

  • Transparency: Users should be informed that transcripts are AI-generated and may contain errors.

References / documentation links:
OpenAI Whisper GitHub
Whisper research paper (Radford et al., 2022)

View all policies