Best Lightroom workflow

[maisun]

Hi @ncoevoet,

I started experimenting with Facet as an Aftershoot replacement. Overall, it's very impressive and has a lot of potential. My main library is in Lightroom Classic, so I have a few questions:

1. Lightroom integration

Do you plan to add deeper Lightroom integration in the future? My ideal workflow would be: select newly imported photos in Lightroom → run Facet → organise bursts into stacks → mark the recommended photo as Pick and rejected photos as Reject → write 1–5 star ratings back to Lightroom.

2. Non-RAW files

For RAW files, the XMP workflow is clear. What is the recommended workflow for JPEG/HEIC? Should Facet create XMP sidecars for these as well, or is embedding the metadata into the original file the intended approach?

3. HDR / panorama

How does Facet distinguish normal bursts from exposure brackets, panoramas and HDR panoramas? I noticed some HDR sequences in my test library being grouped as a normal burst with only one photo recommended. Is that expected, or should HDR sequences be kept together?

Thanks!

[ncoevoet]

Thanks for trying it out, and for the detailed questions. Here is how each part works today, what you can do right now, and what I'd like to add.

1. Lightroom integration

Today there is a Lightroom Classic plug-in in the repository, facet.lrplugin, documented in docs/INTEROP.md. It writes Facet's verdict straight into the catalog: star rating → stars, favourite → Pick, rejected → Reject. It never touches your files, and by default it leaves alone anything you already rated or flagged in Lightroom. Because it matches photos by path, it works the same for ARW as for JPEG.

The closest workflow to what you describe:

1. Import into Lightroom as usual.
2. Scan the import folder: python facet.py /path/to/2026-09-import. Only new files are scored; files already in the database are skipped.
3. Cull the bursts in the viewer's Culling page. Confirming a group marks the frames you didn't keep as rejected.
4. Export the manifest: python facet.py --export-manifest /path/to/2026-09-import.
5. In Lightroom, select the photos, then Library › Plug-in Extras › Facet: Apply ratings and flags…, point it at facet_manifest.json, press Preview, then Apply.
6. For stacks, use Lightroom's own Photo › Stacking › Auto-Stack by Capture Time (a gap of about 1 s groups most bursts).

What's missing, compared with your ideal flow:

• Stacks: Lightroom's plug-in SDK can read stacks but has no call to create them, so no plug-in (Facet's or anyone's) can build them. Auto-stack by capture time is the practical route.
• "Pick the recommended photo": the plug-in turns favourites into Picks. The frame you keep in a burst isn't marked as a favourite, so it doesn't become a Pick yet. The rejected ones do become Rejects.
• Stars from Facet's scores: the plug-in writes the stars you set in Facet. The score-to-stars mapping (--score-to-stars) only exists for XMP sidecars so far.

I'd like to close those last two gaps in the plug-in (see Next steps below).

2. JPEG / HEIC

By default Facet never modifies an original: it writes a sidecar named photo.jpg.xmp. Lightroom doesn't read sidecars for JPEG/HEIC; it reads the XMP embedded in the file. So for Lightroom:

• Recommended: use the plug-in above. It works for every format, RAW included, and writes to the catalog without touching the files.
• If you want the metadata inside the files, for other apps too: python facet.py --export-sidecars /path --embed-originals --score-to-stars, then Metadata › Read Metadata from Files in Lightroom. --embed-originals rewrites the metadata of JPEG/HEIC/TIFF/PNG/DNG files through exiftool; RAW files are never modified. Note that XMP has no field for Lightroom's Pick flag, so a favourite arrives as a Yellow colour label.

One more trap for your ARW files: Facet names RAW sidecars DSC01234.ARW.xmp, while Lightroom looks for DSC01234.xmp, so Lightroom never sees them. Another reason to prefer the plug-in.

3. Bursts, brackets, panoramas

Each scan groups photos in this order:

• Burst: consecutive frames that look alike (perceptual hash) and were shot within a short time window. The best-scored frame is recommended.
• Exposure bracket (HDR): after burst grouping, Facet looks for frames from the same camera, ≤ 3 s apart, with the same framing, whose exposure (worked out from shutter/aperture/ISO in the EXIF) forms an evenly spaced ladder: at least 3 frames, steps of at least 0.5 EV, at least 1 EV in total.
• Panorama: at least 8 frames whose framing drifts steadily (feature matching on the thumbnails). HDR panorama: a panorama whose exposures also span at least 1.5 stops.

Brackets, panoramas and HDR panoramas are kept whole in Culling: all frames stay, because you need all of them to merge. So, no: an HDR sequence showing up as a normal burst with one recommended frame is not expected. It means the bracket test didn't recognise it.

That was a bug, and it's fixed in 1.17.0 "Fusion". Facet's bracket test expected the exposure to go steadily up or down in shooting order (−1, 0, +1). But cameras can also shoot the base exposure first (Sony's Bracket order setting, 0→−→+), so a 0 / −1 / +1 set was rejected and fell back to a plain burst. The test now sorts the frames by exposure first, so any shooting order qualifies. On my own library (mostly Canon, which shoots base-first too) it found 362 brackets it had been missing.

To regroup the photos you already scanned, update and run:

python facet.py --detect-sequences

It works from the EXIF already stored, with no rescan. Most of its run time is the panorama pass: about ten minutes on my 125k photos. Your HDR sets should then show as brackets and be kept whole in Culling. If one still comes out as a burst, please paste the output of this for its frames:

exiftool -T -FileName -SubSecDateTimeOriginal -ExposureTime -FNumber -ISO -ExposureCompensation DSC0xxxx.ARW DSC0xxxy.ARW DSC0xxxz.ARW

Next steps on my side

• Brackets: a manual "mark as bracket" correction in Culling, like the one panoramas already have, for the sets detection still misses.
• Plug-in: an option to Pick the recommended frame of each burst (and optionally Reject the rest), plus stars derived from Facet's scores for photos you haven't rated.
• Plug-in: put each burst / bracket / panorama into a Lightroom collection so a group is one click away from Group into Stack (⌘G).

[maisun]

Hi @ncoevoet , this is really smart design and thanks a lot for the great effort!
It would be nice to flag bursts that looks like HDR or Pano in Lightroom (not sure how maybe using color label?) so it's easier to identify them and merge them in LR. I think it is fair to say that the HDR are easy to spot in Lightroom but I can assure you panos are sometimes harder if don't look closely. So if the model can somehow suggest those photos in this group could be pano, that info should be flagged in Lightroom.
I have a question regarding "Panorama: at least 8 frames whose framing drifts steadily (feature matching on the thumbnails)....": why 8 frames? Sometimes I take just 2 frames, and sometime even with a delay between frames (causal shooting not ideal I know). If there is some smart way to tell: ok those photos you can combine as pano somehow, it would be something great help to Lightroom.

[maisun]

Hi would like to provide some additional reference regarding lightroom integration.
I currently use LrGeniusAI (https://github.com/LrGenius/LrGeniusAI) for AI description of photos, it has seamless integration to LR. I can trigger the analysis from Lightroom with the plugin and think it launches backend process for processing the photos. Can Facet LR plugin add a shortcut to launch the python for analysis and export work?
I also tried the culling features with LrGeniusAI, I think the solution was to group photos in collections (maybe because plugin can't create stacks as you wrote). I think Facet has better pipeline for finding similar photos and tweaking preference for culling. However I think the idea of using LR plugin to start processing and seamless integration to LR could also be done with Facet.
Another cool feature of LrGeniusAI is that I can run any local model, I don't know how Facet use the 2B and 4B models but think my macbook can easily run 27B Qwen Q4, could that improve the culling result in some way?

[ncoevoet]

Thanks again @maisun, both comments turned into real work.

Seeing HDR and panorama sets in Lightroom. The next release does this, though not with colour labels. Facet already writes Yellow for favorites and Red for rejects, and the other colours are yours, so it uses two things the plug-in owns instead:

• Collections. Facet: Apply ratings and flags... can now file every set into Facet › Bursts / Brackets / Panoramas / HDR panoramas, one collection per set, named by date. A Rebuild option refreshes them after a new detection run.
• A "Set kind" metadata field. The plug-in can also write Facet's score, a whole-point score band, the category and the set kind into its own metadata fields. You can filter on those in the Library Filter or in a smart collection, for example "Set kind is HDR panorama".

The same dialog can also pick the recommended frame of each burst, reject the rest, and set stars from the score. There is a new Facet: Export Lightroom State to Facet... item that sends your Lightroom flags and ratings back to Facet, so the two stay in sync. It doesn't have to be a one-way export anymore.

Why 8 frames for a panorama? It's the value that separated real sweeps from look-alikes in the library I tuned on. Every confirmed false positive was 6 frames or fewer: handheld bursts that drift, people walking, pans following a subject. Geometrically, a 2-frame panorama and a 2-frame burst where the camera moved are the same measurement. If I lowered the threshold that far, a lot of ordinary bursts would turn into "panoramas". You can still change it (panorama_detection.min_frames in scoring_config.json), and max_gap_seconds (30 s by default) sets how long a pause between frames can be.

For the short ones, marking by hand is the better tool. In the gallery, select the frames and use Mark as a set → Mark as one panorama. It takes 2 frames or more, can be undone from the snackbar, and survives every later detection run. The next release adds Mark as one bracket for exposure ladders that burst detection missed.

Launching the analysis from Lightroom. Good idea, and the pieces are mostly there: the viewer already has an API for starting a scan and for serving the Lightroom manifest. A plug-in menu item could scan the selected folders, wait for them to finish, then apply the results directly, with no file to save and reopen. What's left is mainly authentication (the plug-in has to log in to the viewer) and making sure Lightroom and Facet see the same paths. It's on the roadmap. The first step will probably be the simpler "fetch and apply the latest results" without triggering the scan.

A 27B Qwen (Q4) on the MacBook. You can try it today. Serve it with Ollama, LM Studio or any OpenAI-compatible server, and point Facet at it with vlm_backend.type: "ollama" or "openai_compatible" (see the VLM backend section of CONFIGURATION.md). But don't expect it to change culling much. The VLM only produces tags, captions, critiques and the "moment" labels. The ranking inside a burst comes from the dedicated aesthetic, face (eyes open, expression) and sharpness models. A bigger VLM would give you better tags and captions, and so better categories and search. It wouldn't pick sharper keepers. On a MacBook it would also be much slower per photo than the built-in models.

And thanks for the kind words about the pipeline compared with LrGeniusAI.

[ncoevoet]

@maisun if you'd like to try all of this before the release, it's on the feature/lightroom-bursts-brackets branch (PR #165). I'd really value a run in a real catalog: the plug-in is tested here against a stubbed Lightroom SDK, so the catalog-writing parts haven't met a real Lightroom yet.

From your source checkout:

git fetch origin && git checkout feature/lightroom-bursts-brackets
pip install -r requirements.txt
cd client && npm install && npx ng build && cd ..   # if you use the viewer
python facet.py --detect-sequences                  # relabel bursts/brackets/panoramas from stored EXIF
python facet.py --export-manifest facet_manifest.json

Then in Lightroom Classic, File → Plug-in Manager: remove the old Facet plug-in and add the branch's facet.lrplugin (zip it first on macOS; Finder treats it as a bundle). Run Library → Plug-in Extras → Facet: Apply ratings and flags... on the manifest. You can also download the manifest from the viewer's Export to editor dialog, in the Lightroom section.

Please try Preview first. It shouldn't change anything in the catalog. It should also suggest the Lightroom/Facet path prefixes if your folders are mounted differently. Then try the new options one at a time, ideally on a copy of the catalog:

• collections (Facet › Bursts / Brackets / Panoramas / HDR panoramas), then Rebuild after running it once
• the Facet metadata fields, including a smart collection on Set kind
• Facet › keywords
• pick the burst lead and reject the rest

To go the other way, Facet: Export Lightroom State to Facet... writes a file that python facet.py --import-lightroom FILE brings back. Change a few flags or ratings in Lightroom first so you can see it working.

Anything odd is useful: an error dialog, a field with no title, a collection in the wrong place, a keyword that ends up in an export. Ticking Write facet-apply.log next to the manifest in the dialog leaves a log beside the manifest, and that log helps a lot. Thanks!

[maisun]

Hi I’d like to try that but my MacBook is under repair and will give a go and I get it back hopefully soon

[ncoevoet]

Hi I’d like to try that but my MacBook is under repair and will give a go and I get it back hopefully soon

Good luck !

Changes has been merged to latest release fyi