This is the multi-page printable view of this section. Click here to print.
RawCull adds optional, locally running AI features to its existing photo-culling workflow. Photographs, prompts, embeddings, masks, and inference results remain on the Mac.
Read:
RawCull AI is planned for release when macOS 27 becomes publicly available. It adds local AI-assisted search and review while retaining the same culling workflow and non-AI functions as the current RawCull version.
The AI features run locally on Apple Silicon. RawCull does not upload photos to an external inference service.
RawCull uses two types of vision model:
AI results are review aids, not automatic decisions. Semantic-search scores describe relative similarity rather than confidence, and the photographer always makes the final selection.
CLIP does not create captions, keywords, or fixed labels while indexing. It converts each image into a normalized numeric embedding that summarizes its overall visual content. RawCull stores that embedding locally and can reuse it for both similarity analysis and semantic search.
Choose a CLIP model in Settings > AI, enable it for similarity, and select Index Similarity or Re-index in the burst workspace. Indexing runs the image encoder once for each photograph that does not already have a compatible cached embedding.
CLIP embeddings are specific to the selected model and its preprocessing configuration. Changing the CLIP model or installing an incompatible model version requires a new index. If CLIP is unavailable, RawCull can use the macOS Vision feature-print fallback for visual grouping, but text-based semantic search requires compatible CLIP embeddings.
Select Analyze Bursts after indexing. RawCull compares the cached image embeddings and groups visually related frames for review. The Similarity control adjusts how tightly frames are grouped: a lower value creates tighter groups, while a higher value admits more visually related frames.
The burst list shows each group and highlights its suggested pick. From here you can open the burst, run Deep Review, mark it reviewed, or defer it until later.
Open a burst to inspect its frames in the filmstrip and compare them at a larger size. The review workspace combines the candidate rank with sharpness, focus-point, saliency, metadata, and subject evidence. You can navigate between frames, assign ratings, pick or reject a photograph, and return to the burst list when the review is complete.
The suggested pick and component scores are starting points, not final judgments. Check expression, pose, timing, framing, and critical focus before deciding which frame to keep.
After a catalog has been indexed with a compatible CLIP model, enter a short description such as puffin, raven, or squirrel. RawCull ranks the catalog using the cached image embeddings, so later searches do not have to reprocess every image.
Search terms work best in English because the supported CLIP models were primarily trained and evaluated with English text.
RawCull supports one selected CLIP model for similarity and semantic search, plus SAM 3 for subject-aware Deep Review.
| RawCull model | Purpose | Upstream model |
|---|---|---|
| OpenAI CLIP ViT-B/32 | Semantic search, similarity, and burst grouping | OpenAI CLIP ViT-B/32 on Hugging Face |
| OpenCLIP ViT-B/32 DataComp | Semantic search, similarity, and burst grouping | DataComp s34B-b86K on Hugging Face |
| Meta SAM 3 | Subject masks and Deep Review | Meta SAM 3 on Hugging Face |
The upstream files on Hugging Face are the source models. RawCull requires model bundles converted and validated for Apple Core AI on macOS 27.
The AI models are not included in the RawCull application or its release download. Compatible model bundles will be available separately. Each RawCull AI release note will identify the supported model versions, download locations, installation steps, and expected checksums.
Follow the RawCull release notes rather than installing an arbitrary conversion. RawCull validates an installed bundle before enabling it and falls back safely when a required model is missing or invalid.
Keeping the models separate makes the application download smaller. It does not remove the upstream model’s licence conditions. Every model archive distributed by RawCull must contain the complete applicable licence, model provenance, conversion information, and checksums.
Before using Download or Accept and Download in RawCull, read AI Model Licences and Downloads.
The RawCull application licence does not replace or extend the licences for the separately downloaded models.
Review the complete licence and model card shown by RawCull before downloading or using a model. RawCull records the exact model revision and licence version applicable to each converted bundle. If a model update changes its licence, RawCull must present the new terms before downloading that update.
RawCull AI requires macOS 27 and Apple Silicon and will be released when the public version of macOS 27 is available.
At that point, active RawCull development and support move to macOS 27. The existing macOS 26 version will no longer receive support or new features, but installed copies will continue to work with their current functionality.
The macOS 27 AI version retains the current RawCull feature set and workflows. AI support is additive: catalogs, ratings, culling, comparison, sharpness tools, previews, and copying continue to work as they do in the current version.
RawCull supports two optional CLIP models for semantic search, visual similarity, and burst grouping. This page explains what CLIP does, how the models were evaluated, and how to compare them during the RawCull beta.
Both models run locally on Apple Silicon. RawCull does not upload photographs, search text, or model results to an external inference service.
CLIP connects text and images by converting both into numeric embeddings. An embedding represents visual and semantic characteristics in a form that can be compared efficiently.
RawCull uses CLIP in two ways:
a dog on a beach, city street at night, or sharp portrait is compared with the indexed photographs.CLIP does not generate captions, modify photographs, rate artistic quality, or make final culling decisions. Its search scores are relative similarity values, not confidence percentages. The photographer remains responsible for every selection and rejection.
| Model | Image input | Background | Practical role |
|---|---|---|---|
| OpenAI CLIP ViT-B/32 | 224 × 224 | The established OpenAI CLIP model | A conservative and well-understood reference model |
| OpenCLIP DataComp ViT-B/32-256 | 256 × 256 | OpenCLIP model trained with DataComp weights | An alternative with different retrieval behavior that may suit some catalogs and searches better |
The models use different weights, image sizes, tokenization details, and converted Core AI graphs. They therefore do not rank every photograph in the same way. A higher raw score from one model cannot be compared directly with a score from the other.
RawCull uses two independent forms of evaluation.
The first test compares embeddings produced by each converted Apple Core AI bundle with embeddings from its original source framework. This detects errors in model conversion, tokenization, preprocessing, output selection, and normalization.
Text and image parity are tested separately. A cosine value of 1.0 means that
the normalized source and Core AI embeddings are identical. Very small
differences are expected from Float16 conversion and differences in image
decoding or resizing.
The second test builds complete, model-specific indexes for the same photo catalog and runs the same 77 text queries. It checks result diversity, repeated “hub” images, query time, paraphrase behavior, and image-similarity neighborhoods.
Conversion parity answers whether Core AI reproduces the source model. Product testing answers whether the model is useful for photographers. One test cannot replace the other.
The current Core AI bundles produced these results on the fixed conversion fixtures and the 77-query RawCull test:
| Result | OpenAI CLIP | DataComp CLIP |
|---|---|---|
| Minimum text parity | approximately 1.0000 | approximately 1.0000 |
| Minimum end-to-end image parity | 0.9988 | 0.9908 |
| Distinct first-ranked images across 77 queries | 53 | 58 |
| Largest repeated first-result hub | 4 of 77 | 4 of 77 |
| Warm mean query time | 67.25 ms | 65.67 ms |
The two models selected the same first-ranked photograph for only 19 of the 77 queries, or approximately 25 percent. This is the most important beta-testing finding: the models offer meaningfully different search behavior.
OpenAI currently passes RawCull’s chosen 0.998 end-to-end conversion-parity
gate. DataComp’s text path is extremely close to its source implementation,
but its image path has a larger difference that remains under investigation.
That difference does not by itself mean that DataComp produces worse semantic
search. DataComp can still retrieve better results for a particular query or
catalog, and the initial product test produced slightly more diverse first
results.
The test set is not yet a large human-labeled accuracy benchmark. Beta feedback is therefore valuable, especially when it describes which model returned a more useful photograph and why.
Downloading both models lets you compare two genuinely different views of the same catalog. OpenAI may work better for one subject or photographic style, while DataComp may work better for another. Testing both also helps RawCull identify weak query categories and choose sensible model-specific defaults.
Each model is optional. Downloading both uses more storage and creating an index for each takes additional time. A user who does not want to participate in comparison testing can install only one model. For this beta, however, installing both is the most useful configuration.
The models are downloaded through macOS Managed Background Assets. macOS stores and manages the asset packs, and RawCull validates a model before enabling it.
Downloaded models run locally. Removing a managed model from the same window removes its RawCull asset pack but does not delete manually installed models.
Use the same photographs and search phrases for both tests.
Useful test phrases include:
a dog, mountains surrounding a lake;yellow flower, red car;person riding a bicycle, bird in flight;sharp portrait, motion blur, backlit subject;Search primarily in English during this beta because the supported models were trained and evaluated mainly with English text.
Useful feedback includes:
Screenshots are helpful, but do not share private photographs unless you are comfortable doing so. A text description of the expected and actual result is enough.
There is no universal winner yet. Choose the model that gives the most useful results for your photographs and vocabulary. OpenAI is the established reference, while DataComp is a promising alternative with noticeably different retrieval behavior.
You can switch later, but changing models requires a compatible index for the new selection. RawCull keeps model identities separate so that embeddings from one model are never silently reused with the other.
For the broader feature overview, see AI Support in RawCull.