Digital Archiving Practice for Camille Claudel’s Sculptures

A digital archive of sculpture is not a matter of putting photographs “in a cloud folder.” It must let a fragile, often misattributed three-dimensional work—one that has been restored and exhibited—still be identified, compared, and cited accurately decades later. The work of the French sculptor Camille Claudel (1864–1943) is a clear case of that difficulty. She practiced in clay from youth, later trained in modeling in Paris, and for a long period took part in large projects as Auguste Rodin’s student, assistant, and co-creator before working independently in later life. Representative works include Shakuntala, The Waltz, The Age of Maturity, and Meditation. Some hands, feet, and small figures were confused with studio works; she also destroyed finished pieces during periods of hardship. A biography can be short. The archival model has to be long.

Separate three kinds of objects

A digital archive must manage at least three kinds of objects at once. They cannot be collapsed into a single “work summary.”

The first is historical people and relationships. The artist, relatives, teachers, studio members, commissioners, and collecting institutions are all entities. Claudel’s documentary relationship with her brother Paul Claudel, and her teacher–student and collaborative relationship with Rodin, can only be recorded as sourced relationship assertions. They must not be written as the system’s sole fact. Family recollections, salon exhibition records, auction catalogues, and later scholarly catalogues carry different grades of evidence.

The second is work versions. One subject may exist as a clay study, a plaster, a bronze cast, and casts of different sizes and periods. The archival primary key should be the physical instance, not the art-historical common name. Titles use a preferred name plus an alias list: The Age of Maturity can link to its usual French title and to Chinese translations, but cast number, dimensions, material, and current location must sit on the instance.

The third is the digital surrogate. Photographs, structured-light point clouds, photogrammetry meshes, pre-conservation images, and exhibition records are observations of one instance at one moment. They are not the work itself. Surrogate files must link back to the instance, capture date, equipment, and a statement of accuracy.

Capture: accuracy for recognition, not a slogan

The differences that matter in studio sculpture often lie in the turn of a hand, overlapping drapery, inscriptions on the base, and traces of the pouring gate. Before capture, record condition: cracks, old repairs, losses, surface coatings, and whether contact is allowed. Plaster that can be touched and bronze that can be molded safely call for different strategies. In a museum setting, non-contact is the default.

Structured light or photogrammetry may both be used. The choice depends on reflectance, scale, and occlusion. Bronze highlights need controlled polarization and exposure sequences. Deep recesses need extra camera positions, or the mesh will open holes between fingers and in folds. Register the point cloud against stable, non-decorative regions so that movable supports or temporary labels are not fitted into the model. Outputs should include at least the original images or raw point cloud, the registered point cloud, a simplified mesh, and one orthographic view with a scale. Simplification is for browsing only. Identification-grade files must not overwrite the raw data.

Quality control uses repeatable measures, not “it looks real”: overlap, reprojection error or point-cloud registration residual, the share of surface that is void, and the smallest resolvable feature (for example, whether strokes of an inscription remain legible). If thresholds are missed, mark the result as a research preview. Do not label it archival grade.

A schematic capture record can be ordinary structured text, unbound to any vendor interface:

instance_id: physical instance identifier
capture_id: this capture identifier
method: photogrammetry | structured_light
date: capture date
scale_ref: scale bar or calibration object
residual_mm: registration residual
void_ratio: void area ratio
grade: archive | preview
operator: person responsible for capture
rights_note: on-site permission and permitted use

Metadata: write “who may have made this” as a traceable assertion

Attribution is where a Claudel archive most easily goes wrong. She modeled parts such as hands and feet for projects including The Gates of Hell, and some works were once treated as Rodin’s. The system should not keep a single author field. It should allow multiple contribution statements: conception, modeling, local carving or modeling, casting, later casts, and restoration. Each statement carries an evidence pointer: archive number, page, image region, and a short quotation summary—a summary, not a long transcription, so copyright and archive agreements are respected.

A minimal field set:

• Identity: preferred title, aliases, date range of creation, subject, material, dimensions, and transcription of signatures and inscriptions.
• Life cycle: place of making, studio, first exhibition, changes of ownership, restoration, damage, or loss.
• Relationships: artist, collaborators, related works (study and finished piece; same model, different casts), and related literature.
• Digital layer: file checksums, derivative level, capture grade, coordinates, and units.
• Rights: photography permission, public scope of the 3D model, and whether commercial derivatives are allowed.

If a date is uncertain, record a range. Do not collapse “they met about 1882” into an exact day. Where sources disagree—1882 or 1883, for example—keep the conflicting sources. Do not force a merge.

Versions, checksums, and long-term preservation

Once an original capture package is ingested, it is read-only. Browse meshes, thumbnails, and simplified models that a web page can load are derivatives. Record the parent checksum and the names of processing steps (registration, denoising, hole filling, decimation). Do not treat one vendor’s proprietary project file as the only master. What must remain readable for the long term is open, documented geometry and image formats. Proprietary projects may be stored alongside; they cannot replace the master.

Storage tiers can use general capabilities: object storage for large raw packages, a database for metadata and relationships, and logs plus monitoring alerts for integrity checks. Recompute checksums on a schedule. If silent corruption appears, restore from a replica. Replicas should cross failure domains, not merely mirror the same disk group. Destruction policy needs special care. The artist herself destroyed finished work in the studio; a digital archive must not imitate that kind of “cleanup.” Taking a record offline changes only its public status. Do not physically delete the master unless law or a gift agreement explicitly requires it, and then keep proof of destruction.

Knowledge organization and retrieval

Readers usually ask: “Is this hand related to a particular gate?” “Was this bronze cast in her lifetime or later?” “What sizes of The Waltz exist?” Search should filter by instance, subject, material, date range, and relationship type, and should show evidence grade by default. A knowledge graph can express part–whole, same model with different casts, and former misattribution, but every edge must carry a source. An edge without a source may enter a queue for review. It must not enter the public timeline.

Full-text search should cover Chinese translated titles and the foreign name Camille Claudel, as well as differences among work titles. A biography page aggregates only reviewed facts: born in Villeneuve-sur-Fère, profession sculptor, committed to a psychiatric hospital in 1913, died in Montdevergues in 1943. Mental condition and involuntary hospitalization are sensitive biographical information. Public display should stay within what art history has already published: do not dwell on clinical detail, and do not turn family decisions into a moral judgment.

Collaborative workflow

1. A cataloguer creates an instance draft from catalogues and marks every gap as unknown.
2. A capture operator binds the capture package and its accuracy grade.
3. A research editor adds attribution and relationships; each item must have evidence.
4. A reviewer checks inscriptions, dimension units, and rights notes.
5. Publication produces a read-only snapshot. Later corrections create a new version; they do not overwrite a snapshot that has already been cited.

Later auctions, newly found casts, or exhibitions then add instances or evidence. They do not rewrite old snapshots. In recent years her work has returned to public view and sold at high prices, which shows that both the market and scholarship depend on images and sources that can be checked. The archive supplies that chain of verification, not a price forecast.

What to check at acceptance

Do not accept a system because the page “looks like a museum website.” Check whether ten sampled instances can be traced from the public model back to the checksum of the original package; whether attribution disputes keep both claims and their sources; whether preview-grade models have been mislabeled as archival grade; whether an empty rights field defaults to no public 3D download; and whether sensitive biography can be displayed at a lower level on its own. When those conditions hold, the digital archive serves identification, teaching, and long-term preservation instead of producing another round of attribution confusion.