From cfa54dc8ae0ba2ca1bca3e0038b785142fc0f273 Mon Sep 17 00:00:00 2001
From: Vito Sansevero <vito.sansevero@gmail.com>
Date: Tue, 27 Oct 2026 05:35:22 -0700
Subject: [PATCH] fix(tagging): Tag output images tags each image, not its
 prompt (#379, PR 1)

The native "Tag output images" job now runs the image-tags loop over every
still image the output folders list, linked or not: each image gets its own
model tags (at most "Suggestions per image") in its image tag set, and its
bytes are marked auto-tagged in a new media_image_autotags table. No prompt
tag or prompt text is written.

- database/autotag_schema.py: one forward-only WITHOUT ROWID table, optional
  in old backups, refused without the annotation group.
- add_image_tags/replace_image_tags take mark_autotagged; autotag_state and
  autotagged_locators read the marks through the one current-record rule.
- "Skip already tagged" skips images whose current bytes are marked, counted
  in a stat-only pre-pass; hand tags do not count.
- The review counts images and estimates the time (WD14 on CPU or CUDA).
- System > Auto-tagging review rows and copy, D134 and the Auto-tagging
  contract follow.

Part of #379
---
 TESTING.md                                    |  74 +-
 database/annotations.py                       | 122 +++-
 database/autotag_schema.py                    | 140 ++++
 database/models.py                            |  15 +
 py/api/autotag_routes.py                      | 191 ++----
 py/api/image_tag_jobs.py                      | 410 ++++++++++-
 py/autotag.py                                 |  18 +
 py/tagging_defaults.py                        |  18 +-
 tests/browser/system-tagging-concept.spec.mjs |  46 +-
 .../browser/workspace-system-tagging.spec.mjs |  43 +-
 tests/js/workspace-system-tagging.test.js     |  44 +-
 tests/test_annotation_schema.py               |   4 +-
 tests/test_annotations.py                     | 136 ++++
 tests/test_api_autotag.py                     | 172 +----
 tests/test_autotag_batch.py                   | 634 ++++++++++++++++++
 tests/test_autotag_engine.py                  |  26 +
 tests/test_autotag_schema.py                  | 318 +++++++++
 tests/test_database_backup.py                 |   1 +
 tests/test_review_schema.py                   |   5 +-
 web/js/workspace/system-tagging.mjs           |  58 +-
 20 files changed, 2044 insertions(+), 431 deletions(-)
 create mode 100644 database/autotag_schema.py
 create mode 100644 tests/test_autotag_batch.py
 create mode 100644 tests/test_autotag_schema.py

diff --git a/TESTING.md b/TESTING.md
index 16fb8485..2e3b7184 100644
--- a/TESTING.md
+++ b/TESTING.md
@@ -3103,25 +3103,39 @@ Recorded deviations (plan D60–D75, D79–D82 renumbered D92–D107 and D111–
   the persisted count the Tag image dialog's Suggestions field (D91) starts from. Pinned
   by the comparison's mapping (it inserts the row) and `defaults save through Save
   preferences and the Tag image dialog uses them`.
-- **D134** (Auto-tagging, S11, OQ2/OQ3/OQ7/OQ8): an `.operation` "Tag output images"
-  with "Review" follows `.row-actions`: the legacy prompt-tag batch. Review is
+- **D134** (Auto-tagging, S11, OQ2/OQ3/OQ7/OQ8; amended by #379): an `.operation` "Tag
+  output images" with "Review" follows `.row-actions`: the per-image batch (#379), whose
+  description reads "Tag each output image from its own pixels with the model above.
+  Prompts are not changed. Review the scope first.". Review is
   `aria-disabled` (the press explains) while the chosen engine is not downloaded or a
   value it uses is invalid; otherwise it posts `/autotag/batch/preview` with the drafted
   engine and policy (read-only) and opens "Review tag output images" in the concept's
   record-changing review (its `maintenancePlan` shell): `h3` "Tag output images", the
-  `dl.meta-grid` Model, Existing tags, Images linked, Prompts, Already tagged, Prompts
-  to tag ("up to N"), Settings ("General <g> · Character <c>" or "Template of N
-  characters") and Backup, the upper-bound sentence in the `p.small`, the D55 backup box
-  and Cancel / "Start tagging", which starts the cancellable `autotag` job and opens it.
-  Replace after review, a model not on disk, a short disk, no linked image and nothing to
-  tag each show the rows with only Cancel and the reason ("Replacing tags needs a review
-  of each image. Use Preview tagging review, or choose Skip or Merge for a batch run.",
-  "Every prompt with an output image is already tagged.", "Not enough free disk space:
-  the safety backup needs about X MB and Y MB is free. Nothing is written until there is
-  room.") in the close-only review Integrations shares (`openCloseOnlyReview`). Pinned by
+  `dl.meta-grid` Model, Existing tags, Images, Already auto-tagged, Images to tag
+  ("about N"), Time, Settings ("General <g> · Character <c>" or "Template of N
+  characters", then " · up to L tags per image" with the preview's saved
+  `suggestion_limit`) and Backup, the estimate sentence ("Nothing is written until you
+  start. The number to tag is an estimate: images changed since they were tagged are
+  tagged again, and missing files are skipped. Activity reports what was tagged.") in the
+  `p.small`, the D55 backup box and Cancel / "Start tagging", which starts the
+  cancellable `autotag` job and opens it. Time (DEC11) is `m = max(1,
+  ceil(estimate_seconds / 60))` read "about m min" under 90 and "about H h M min" (or
+  "about H h") from 90, followed by " on a processor; a graphics card can be faster."
+  or, when the preview says WD14 runs on CUDA, " on the graphics card."; the language
+  models read "Language-model engines take longer per image. Activity shows the time
+  left once tagging starts.". The option label "Skip already tagged" is the concept's;
+  for this batch it skips images already auto-tagged (hand tags do not count). Replace
+  after review, a model not on disk, a short disk, no image and nothing to tag each show
+  the rows with only Cancel and the reason ("Replacing tags needs a review of each
+  image. Use Preview tagging review, or choose Skip or Merge for a batch run.", "The
+  output folders contain no images to tag.", "Every output image is already
+  auto-tagged.", "Not enough free disk space: the safety backup needs about X MB and Y
+  MB is free. Nothing is written until there is room.") in the close-only review
+  Integrations shares (`openCloseOnlyReview`). Pinned by
   `tag output images review matches the concept`, its mutation check and `the
   Cancel-only review matches the concept review shell` (system-tagging-concept.spec) and
-  the batch tests of workspace-system-tagging.spec.
+  the batch tests of workspace-system-tagging.spec; the mutation check renames each grid
+  label in turn.
 - **D124** (inspector): with "Load metadata on selection" off, the Images panel and
   the viewer's details show "Metadata loads when you ask for it." and a "Load metadata"
   button instead of reading the file. The concept has no such state, so no concept
@@ -3680,19 +3694,45 @@ images", "Tag output images (legacy page)", "Tagging one image (legacy page)",
   `409 busy`; the legacy SSE route attaches to the same job and keeps its frames.
 - **Unload**: `POST /prompt_manager/autotag/unload` answers `{success, message}` or
   `409 model-busy`.
-- **Tag output images**: `POST /prompt_manager/autotag/batch/preview {model_type,
-  existing}` writes nothing and opens no file; it answers `images`, `prompts`,
-  `already_tagged`, `to_tag` (an upper bound), `model {type, label, size_label,
-  downloaded}`, `existing`, `review_required`, `backup {policy}` and `disk`. `POST
+- **Tag output images** (per image since #379; prompt tags and prompt text are never
+  written): its targets are every still image the output folders list (the Images
+  listing of each root, linked to a prompt or not; videos are not targets; a file under
+  two overlapping roots is one target). `POST /prompt_manager/autotag/batch/preview
+  {model_type, existing}` writes nothing and opens or stats no image file (it may wait
+  for the shared listing walk); it answers `images`, `already_tagged` (images with any
+  auto-tagged byte record), `to_tag` (`images − already_tagged` under Skip, `images`
+  under Merge; both estimates), `estimate_seconds` (WD14: `to_tag` × 1.6 s on the
+  processor or × `WD14_GPU_SECONDS_PER_IMAGE` when ONNX Runtime offers CUDA, see
+  `py/tagging_defaults.py`; `null` for GGUF and HF), `estimate_device` (`cpu`, `gpu`,
+  or `null`), `suggestion_limit` (the saved count the run uses), `model {type, label,
+  size_label, downloaded}`, `existing`, `review_required`, `backup {policy}` and
+  `disk`. `POST
   /prompt_manager/jobs/autotag {model_type, existing, options}` (WD14: both thresholds;
   GGUF/HF: `instructions`, the template) starts the cancellable `autotag` job (202),
   title "Tag output images", admission `records`, a safety backup by the policy first.
+  The job runs the `image-tags` loop (`_tag_each_file`) on each image: a capture, at
+  most `suggestion_limit` model tags, a re-hash, then a union into that image's tags
+  that also marks its bytes auto-tagged (`media_image_autotags`, one row per byte
+  record; an empty result and an unsupported still are marked too; File changed, File
+  not found, Too large to verify, read, inference and tag-limit errors are not). Phases,
+  each checked for Cancel: backup, model load, listing ("Listing output images…"), a
+  pre-pass that stats each file without opening it (gone: "File not found", over 256
+  MiB: "Too large to verify", under Skip a marked current record: "Already
+  auto-tagged"; counted with `record_counted`, rows up to the cap, no log line each),
+  then the loop ("Processing file {k + n} of {T}", the time left from the model's own
+  rate, D56 unchanged). It ends "Tagged N images", "Cancelled. Tags already added were
+  kept.", or failed with "Prompt storage was restored; start the job again.", "The
+  output folders could not be listed. Try again." or the model's refusal.
   Refusals: `400 review-required` (Replace after review never runs as a batch), `409
   model-unavailable` ("<label> is not downloaded. Download it in System ›
   Auto-tagging."), `507 insufficient-space`, `409 busy`, `409 model-busy`. The section
   toasts `model-busy` with its operation, the other three with the server's sentence,
   `busy` with the running job (D57), anything else "<what> could not start. Try again.".
   Outcomes: Tagged / Unchanged / Skipped / Errors (the `image-tags` descriptor, D111).
+  `tests/test_autotag_batch.py` pins it, including two images of one prompt getting
+  different tags with no `prompts` or `prompt_tags` row changed; the marks' schema group
+  is pinned by `tests/test_autotag_schema.py` (forward-only: one empty `WITHOUT ROWID`
+  table, optional in old backups, refused without the annotation group).
 
 No browser test downloads or runs a model: the injected `TestModel` answers, its
 `download_model` raises unless `/__test__/tagging/download {mode: hold | succeed |
diff --git a/database/annotations.py b/database/annotations.py
index 219a5039..6c0ee6b9 100644
--- a/database/annotations.py
+++ b/database/annotations.py
@@ -19,6 +19,11 @@ links (``PromptDatabase._move_tag_links``).
 Cheap reads (spec amendment S1) treat a byte record as *current* only while
 the file's present ``(size, mtime_ns)`` equals the record's verified
 fingerprint; a file without a current record shows no image tags.
+
+A write made from a tagging model's result passes ``mark_autotagged`` (#379):
+the byte record is then marked auto-tagged (``media_image_autotags``) in the
+same transaction, so a mark never exists without the write it records. Hand
+tags never mark.
 """
 
 import re
@@ -34,6 +39,7 @@ from .annotation_schema import (
     MAX_LOCATOR_LENGTH,
     METADATA_TABLE,
 )
+from .autotag_schema import AUTOTAG_TABLE
 from .library_commands import _UUID, LibraryCommandError, _validate_epoch
 from .models import utc_now_iso
 from .vocabulary import validate_vocabulary_tag_name
@@ -96,6 +102,14 @@ _TAGS_OF = (
     "JOIN tags t ON t.id=mit.tag_id WHERE mit.byte_id IN ({marks}) "
     "ORDER BY mit.byte_id,mit.position"
 )
+_MARK = f"INSERT OR IGNORE INTO {AUTOTAG_TABLE}(byte_id) VALUES(?)"  # nosec B608
+_MARKED_OF = (
+    f"SELECT byte_id FROM {AUTOTAG_TABLE} WHERE byte_id IN ({{marks}})"  # nosec B608
+)
+_MARKED_LOCATORS = (
+    f"SELECT DISTINCT f.locator FROM {AUTOTAG_TABLE} m "  # nosec B608
+    f"JOIN {BYTES_TABLE} b ON b.id=m.byte_id JOIN {FILES_TABLE} f ON f.id=b.file_id"
+)
 _RECORDS = (
     "SELECT f.locator,b.id,b.size,b.mtime_ns,t.name "  # nosec B608
     f"FROM {FILES_TABLE} f JOIN {BYTES_TABLE} b ON b.file_id=f.id "
@@ -301,14 +315,17 @@ class AnnotationOperations:
         expected_epoch: str,
         expected_tag_revision: int,
         tags: list,
+        mark_autotagged: bool = False,
     ) -> dict:
         """Replace the set of these exact bytes, guarded by epoch and revision.
 
         Creates the owner and byte record when the bytes are new (an explicit
         empty set is a real save). The identical set at the current revision,
         or one behind it (a lost-response retry), is ``replayed`` and writes
-        nothing; any other revision mismatch is ``tag-conflict`` with the
-        current set and no writes.
+        no tag; any other revision mismatch is ``tag-conflict`` with the
+        current set and no writes. ``mark_autotagged`` marks the bytes
+        auto-tagged on a saved or replayed receipt (the Tag image dialog's
+        save after Generate suggestions).
         """
         _validate_target(locator, sha256, size, mtime_ns)
         _validate_epoch(expected_epoch)
@@ -329,6 +346,8 @@ class AnnotationOperations:
             ):
                 if refingerprinted:
                     conn.execute(_BUMP_INDEX)
+                if mark_autotagged:
+                    conn.execute(_MARK, (byte_id,))
                 return _receipt(epoch, sha256, current, revision, True)
             if expected_tag_revision != revision:
                 raise AnnotationCommandError(
@@ -337,6 +356,8 @@ class AnnotationOperations:
                     {"sha256": sha256, "tags": current, "tag_revision": revision},
                 )
             revision = self._write_image_tags(conn, byte_id, tags)
+            if mark_autotagged:
+                conn.execute(_MARK, (byte_id,))
             return _receipt(epoch, sha256, tags, revision, False)
 
     def add_image_tags(
@@ -348,16 +369,22 @@ class AnnotationOperations:
         mtime_ns: int,
         expected_epoch: str,
         tags: list,
+        mark_autotagged: bool = False,
     ) -> dict:
         """Union ``tags`` into the set of these exact bytes, keeping stored spelling.
 
         ``added`` counts the names that were new (case-insensitively). A union
         over ``MAX_IMAGE_TAGS`` is ``tag-limit`` with the current set and no
         writes; a union that adds nothing to an existing record is ``replayed``.
+        ``mark_autotagged`` (a tagging model's result) also marks the bytes
+        auto-tagged, replayed or not; only such a call may pass no tags (the
+        model found none: the image is marked and its set is unchanged).
         """
         _validate_target(locator, sha256, size, mtime_ns)
         _validate_epoch(expected_epoch)
         tags = normalize_image_tags(tags)
+        if not tags and not mark_autotagged:
+            raise ValueError("tags must name at least one tag")
         with self._pair_transaction() as conn:
             epoch = self._library_epoch(conn, expected_epoch)
             now = utc_now_iso()
@@ -375,6 +402,8 @@ class AnnotationOperations:
                     epoch,
                     {"sha256": sha256, "tags": current, "tag_revision": revision},
                 )
+            if mark_autotagged:
+                conn.execute(_MARK, (byte_id,))
             if not created and not added:
                 if refingerprinted:
                     conn.execute(_BUMP_INDEX)
@@ -403,6 +432,41 @@ class AnnotationOperations:
                 conn.execute(_BUMP_INDEX)
             return cursor.rowcount > 0
 
+    @staticmethod
+    def _validated_stats(stats: object) -> dict:
+        if not isinstance(stats, dict):
+            raise ValueError(
+                "stats must map locators to (size, mtime_ns) integer pairs"
+            )
+        for locator, fingerprint in stats.items():
+            _validate_fingerprint(locator, fingerprint)
+        return stats
+
+    @staticmethod
+    def _current_byte_records(conn: sqlite3.Connection, stats: dict) -> dict:
+        """``{locator: byte_id}`` of each locator's current byte record.
+
+        The one rule every cheap read uses: a record is current while its
+        verified ``(size, mtime_ns)`` equals the file's present fingerprint
+        in ``stats``; the most recently verified wins (ties on ``(verified_at,
+        byte_id)``). A locator without a current record is left out. Lookups
+        go in chunks of 500 locators.
+        """
+        locators = list(stats)
+        chosen: dict[str, tuple[tuple[str, str], str]] = {}
+        for start in range(0, len(locators), _LOOKUP_CHUNK):
+            chunk = locators[start : start + _LOOKUP_CHUNK]
+            marks = ",".join("?" * len(chunk))
+            for locator, byte_id, size, mtime_ns, verified_at in conn.execute(
+                _FINGERPRINTS.format(marks=marks), chunk
+            ):
+                if (size, mtime_ns) != stats[locator]:
+                    continue
+                key = (verified_at, byte_id)
+                if locator not in chosen or key > chosen[locator][0]:
+                    chosen[locator] = (key, byte_id)
+        return {locator: byte_id for locator, (_key, byte_id) in chosen.items()}
+
     def current_image_tags(self, stats: dict) -> dict:
         """Image tags of the files whose current ``(size, mtime_ns)`` is verified.
 
@@ -411,29 +475,11 @@ class AnnotationOperations:
         exactly that verified fingerprint; the most recently verified record
         wins when several do. Lookups go in chunks of 500 locators.
         """
-        if not isinstance(stats, dict):
-            raise ValueError(
-                "stats must map locators to (size, mtime_ns) integer pairs"
-            )
-        for locator, fingerprint in stats.items():
-            _validate_fingerprint(locator, fingerprint)
-        locators = list(stats)
-        chosen: dict[str, tuple[tuple[str, str], str]] = {}
-        result: dict[str, list[str]] = {}
+        stats = self._validated_stats(stats)
         with self._annotation_read() as conn:
-            for start in range(0, len(locators), _LOOKUP_CHUNK):
-                chunk = locators[start : start + _LOOKUP_CHUNK]
-                marks = ",".join("?" * len(chunk))
-                for locator, byte_id, size, mtime_ns, verified_at in conn.execute(
-                    _FINGERPRINTS.format(marks=marks), chunk
-                ):
-                    if (size, mtime_ns) != stats[locator]:
-                        continue
-                    key = (verified_at, byte_id)
-                    if locator not in chosen or key > chosen[locator][0]:
-                        chosen[locator] = (key, byte_id)
-            owners = {byte_id: locator for locator, (_key, byte_id) in chosen.items()}
-            result = {locator: [] for locator in chosen}
+            current = self._current_byte_records(conn, stats)
+            owners = {byte_id: locator for locator, byte_id in current.items()}
+            result: dict[str, list[str]] = {locator: [] for locator in current}
             byte_ids = list(owners)
             for start in range(0, len(byte_ids), _LOOKUP_CHUNK):
                 chunk = byte_ids[start : start + _LOOKUP_CHUNK]
@@ -442,6 +488,34 @@ class AnnotationOperations:
                     result[owners[byte_id]].append(name)
         return result
 
+    def autotag_state(self, stats: dict) -> dict:
+        """``{locator: bool}``: is each file's current byte record auto-tagged?
+
+        ``stats`` as for :meth:`current_image_tags`, whose rule picks the
+        current record; a locator without one is False. One read snapshot.
+        """
+        stats = self._validated_stats(stats)
+        with self._annotation_read() as conn:
+            current = self._current_byte_records(conn, stats)
+            byte_ids = list(dict.fromkeys(current.values()))
+            marked: set[str] = set()
+            for start in range(0, len(byte_ids), _LOOKUP_CHUNK):
+                chunk = byte_ids[start : start + _LOOKUP_CHUNK]
+                marks = ",".join("?" * len(chunk))
+                marked.update(
+                    row[0]
+                    for row in conn.execute(_MARKED_OF.format(marks=marks), chunk)
+                )
+        return {locator: current.get(locator) in marked for locator in stats}
+
+    def autotagged_locators(self) -> set:
+        """Locators with at least one auto-tagged byte record, current or not.
+
+        No fingerprint is checked (the Tag output images review's estimate).
+        """
+        with self._annotation_read() as conn:
+            return {row[0] for row in conn.execute(_MARKED_LOCATORS)}
+
     def iter_image_tag_records(self) -> Iterator[tuple[str, int, int, list[str]]]:
         """Stream every byte record as ``(locator, size, mtime_ns, tags)``.
 
diff --git a/database/autotag_schema.py b/database/autotag_schema.py
new file mode 100644
index 00000000..29e2db1b
--- /dev/null
+++ b/database/autotag_schema.py
@@ -0,0 +1,140 @@
+"""Auto-tagged marks (#379): which image bytes a tagging model has handled.
+
+One table, ``media_image_autotags``: one row per byte record of the image
+annotation group (``media_annotation_bytes``) means "a tagging model looked at
+these exact bytes and its result was stored", or that the bytes were found to
+be untaggable. "Skip already tagged" skips an image whose current byte record
+is marked; hand tags never mark. Rows are written only by the image tag store
+(``database/annotations.py``), in the transaction of the tag write they
+record.
+
+The group installs like ``image_favorites``: optional in old backups,
+required once a database is current, and never backfilled. Like the Needs
+review group it is valid only beside a valid image annotation group, because
+every mark cites one byte record (``ON DELETE CASCADE``). Validation reads
+only: the catalog compared after whitespace compaction, the columns and the
+foreign key through the PRAGMAs, then one bounded value pass and the foreign
+key check.
+"""
+
+import re
+import sqlite3
+from typing import Optional
+
+from .annotation_schema import BYTES_TABLE
+from .annotation_schema import TABLES as ANNOTATION_TABLES
+from .annotation_schema import _catalog as _annotation_catalog
+from .annotation_schema import _uuid, annotation_schema_error
+
+AUTOTAG_TABLE = "media_image_autotags"
+_DDL = (
+    f"CREATE TABLE {AUTOTAG_TABLE} (byte_id TEXT PRIMARY KEY "
+    f"REFERENCES {BYTES_TABLE}(id) ON DELETE CASCADE) WITHOUT ROWID"
+)
+# (cid, name, declared type, not null, default, primary key position): the
+# key of a WITHOUT ROWID table is implicitly NOT NULL.
+_COLUMNS = [(0, "byte_id", "TEXT", 1, None, 1)]
+# (column, parent table, parent column, on delete)
+_FOREIGN_KEYS = {("byte_id", BYTES_TABLE, "id", "CASCADE")}
+
+NEEDS_ANNOTATIONS = "Image auto-tag marks need the image annotation tables"
+INVALID_SCHEMA = "Image auto-tag marks have an invalid schema"
+INVALID_VALUES = "Image auto-tag marks contain invalid values"
+MISSING = "Missing image auto-tag marks schema"
+
+
+def _compact(sql: str) -> str:
+    return re.sub(r"\s+", "", sql).lower()
+
+
+def _catalog(conn: sqlite3.Connection) -> list:
+    return conn.execute(
+        "SELECT type,name,sql FROM sqlite_master WHERE name=? OR tbl_name=?",
+        (AUTOTAG_TABLE, AUTOTAG_TABLE),
+    ).fetchall()
+
+
+def autotag_without_annotations_error(conn: sqlite3.Connection) -> Optional[str]:
+    """Catalog only: the marks table present without its annotation tables.
+
+    Startup asks this before the annotation group installs, which would make
+    a marks table copied without it look valid. It reads no rows.
+    """
+    if not _catalog(conn):
+        return None
+    present = {
+        name for kind, name, _sql in _annotation_catalog(conn) if kind == "table"
+    }
+    if present != set(ANNOTATION_TABLES):
+        return NEEDS_ANNOTATIONS
+    return None
+
+
+def autotag_schema_error(
+    conn: sqlite3.Connection,
+    *,
+    allow_missing: bool = True,
+    annotations_valid: bool = False,
+) -> Optional[str]:
+    """Reject an unsupported marks table or invalid marks; reads only.
+
+    A database without the table is acceptable for backups that predate it
+    unless ``allow_missing`` is false. A present table requires a present,
+    valid image annotation group; ``annotations_valid`` says the caller has
+    just validated it (startup's annotation install), so only its presence is
+    checked again.
+    """
+    objects = _catalog(conn)
+    if not objects:
+        return None if allow_missing else MISSING
+    if annotations_valid:
+        if autotag_without_annotations_error(conn):
+            return NEEDS_ANNOTATIONS
+    elif annotation_schema_error(conn, allow_missing=False):
+        return NEEDS_ANNOTATIONS
+    actual = {(kind, name, _compact(sql or "")) for kind, name, sql in objects}
+    if actual != {("table", AUTOTAG_TABLE, _compact(_DDL))}:
+        return INVALID_SCHEMA
+    columns = [
+        (row[0], row[1], row[2].upper(), row[3], row[4], row[5])
+        for row in conn.execute(f"PRAGMA table_info({AUTOTAG_TABLE})")
+    ]
+    if columns != _COLUMNS:
+        return INVALID_SCHEMA
+    foreign_keys = {
+        (row[3], row[2], row[4], row[6])
+        for row in conn.execute(f"PRAGMA foreign_key_list({AUTOTAG_TABLE})")
+    }
+    if foreign_keys != _FOREIGN_KEYS:
+        return INVALID_SCHEMA
+    # Only this module's constants are interpolated into the SQL below.
+    if conn.execute(
+        f"SELECT 1 FROM {AUTOTAG_TABLE} "  # nosec B608
+        f"WHERE NOT ({_uuid('byte_id')}) LIMIT 1"
+    ).fetchone():
+        return INVALID_VALUES
+    if conn.execute(f"PRAGMA foreign_key_check({AUTOTAG_TABLE})").fetchone():
+        return INVALID_VALUES
+    return None
+
+
+def migrate_autotag_state(conn: sqlite3.Connection) -> None:
+    """Install the table atomically; validate before and after creating it.
+
+    Runs after ``migrate_annotation_state`` (which has just validated the
+    annotation group). Creates one empty table: no row is read or copied.
+    """
+    conn.execute("SAVEPOINT media_autotag_schema")
+    try:
+        error = autotag_schema_error(conn, annotations_valid=True)
+        if error:
+            raise sqlite3.DatabaseError(error)
+        conn.execute(_DDL.replace("CREATE TABLE", "CREATE TABLE IF NOT EXISTS", 1))
+        error = autotag_schema_error(conn, allow_missing=False, annotations_valid=True)
+        if error:
+            raise sqlite3.DatabaseError(error)
+        conn.execute("RELEASE media_autotag_schema")
+    except BaseException:
+        conn.execute("ROLLBACK TO media_autotag_schema")
+        conn.execute("RELEASE media_autotag_schema")
+        raise
diff --git a/database/models.py b/database/models.py
index e49ff94f..dad9cf07 100644
--- a/database/models.py
+++ b/database/models.py
@@ -23,6 +23,11 @@ from . import restore_stage
 from .library_schema import library_schema_error, migrate_library_state
 from .image_favorites import image_favorites_schema_error, migrate_image_favorites
 from .annotation_schema import annotation_schema_error, migrate_annotation_state
+from .autotag_schema import (
+    autotag_schema_error,
+    autotag_without_annotations_error,
+    migrate_autotag_state,
+)
 from .review_schema import (
     migrate_review_state,
     review_schema_error,
@@ -338,8 +343,13 @@ class PromptModel:
         review_error = review_without_annotations_error(conn)
         if review_error:
             raise sqlite3.DatabaseError(review_error)
+        # The auto-tag marks cite annotation bytes the same way (#379).
+        autotag_error = autotag_without_annotations_error(conn)
+        if autotag_error:
+            raise sqlite3.DatabaseError(autotag_error)
         migrate_annotation_state(conn)
         migrate_review_state(conn)
+        migrate_autotag_state(conn)
         # Last: the older generated_images rebuilds run only before file_path.
         migrate_workflow_documents(conn)
 
@@ -1225,6 +1235,11 @@ class PromptModel:
             )
             if review_error:
                 return False, review_error
+            autotag_error = autotag_schema_error(
+                conn, allow_missing=not require_library_state
+            )
+            if autotag_error:
+                return False, autotag_error
             documents_error = workflow_documents_schema_error(
                 conn, allow_missing=not require_library_state, documents=documents
             )
diff --git a/py/api/autotag_routes.py b/py/api/autotag_routes.py
index b69c5210..5d0b7fe0 100644
--- a/py/api/autotag_routes.py
+++ b/py/api/autotag_routes.py
@@ -15,12 +15,11 @@ from urllib.parse import quote
 from .job_routes import busy_response
 from .jobs import (
     AUTOTAG_BATCH_JOB_KIND,
-    AUTOTAG_GROUP_NAMES,
     MODEL_DOWNLOAD_JOB_KIND,
     JobBusy,
 )
-from .admin import _plural, _scan_location, _scan_roots
-from .maintenance import disk_plan, short_space_message
+
+from .maintenance import disk_plan
 from .prompts import thumbnail_url_for
 from . import thumbnail_settings
 from .. import model_gate
@@ -29,7 +28,7 @@ from aiohttp import web
 
 _INTERNAL_ERROR = "An internal error occurred. Check server logs for details."
 
-# Tag output images (#328): the native prompt-tag batch.
+# Tag output images (#328, per image since #379).
 AUTOTAG_JOB_TITLE = "Tag output images"
 AUTOTAG_PREVIEW_LEASE = "Tag output images review"
 AUTOTAG_OWNER_KIND = "autotag-job"
@@ -200,6 +199,18 @@ async def _refusal(title):
     yield {"type": "error", "status": "error", "message": message}
 
 
+def wd14_device(service) -> str:
+    """Worker thread: ``"gpu"`` when WD14 runs (or would run) on CUDA, else
+    ``"cpu"``; a service without the probe is ``"cpu"``."""
+    probe = getattr(service, "wd14_device", None)
+    if not callable(probe):
+        return "cpu"
+    try:
+        return "gpu" if probe() == "gpu" else "cpu"
+    except Exception:
+        return "cpu"
+
+
 class AutotagRoutesMixin:
     """Mixin providing auto-tagging API endpoints."""
 
@@ -802,6 +813,8 @@ class AutotagRoutesMixin:
 
     async def _preview_autotag_batch(self, request):
         from ..autotag import get_autotag_service
+        from ..config import AutotagConfig
+        from ..tagging_defaults import WD14_MODEL_TYPES, wd14_seconds_per_image
 
         data, error = await self._batch_body(
             request, ("model_type", "existing"), BATCH_PREVIEW_BODY_MAX
@@ -813,11 +826,14 @@ class AutotagRoutesMixin:
             return self._bad_request(message)
         model_type, existing = data["model_type"], data["existing"]
         policy = self._safety_backup_policy()
+        service = get_autotag_service()
         try:
-            candidates = await self._run_in_executor(self.db.autotag_batch_candidates)
-            status = await self._run_in_executor(
-                get_autotag_service().get_models_status
-            )
+            targets = await self._autotag_targets()
+            marked = await self._run_in_executor(self.db.autotagged_locators)
+            status = await self._run_in_executor(service.get_models_status)
+            device = None
+            if model_type in WD14_MODEL_TYPES:
+                device = await self._run_in_executor(wd14_device, service)
             disk, _short, _size = await self._run_in_executor(
                 functools.partial(
                     disk_plan,
@@ -831,16 +847,21 @@ class AutotagRoutesMixin:
             return web.json_response(
                 {"success": False, "error": self._public_error(e)}, status=500
             )
-        prompts = {row["prompt_id"]: row["prompt_tags"] for row in candidates}
-        tagged = sum(1 for tags in prompts.values() if self._real_tags(tags))
-        to_tag = len(prompts) - tagged if existing == "skip" else len(prompts)
+        images = len(targets)
+        tagged = sum(1 for item in targets if item.locator in marked)
+        to_tag = images - tagged if existing == "skip" else images
+        estimate = None
+        if device is not None:
+            estimate = round(to_tag * wd14_seconds_per_image(device), 3)
         return web.json_response(
             {
                 "success": True,
-                "images": len(candidates),
-                "prompts": len(prompts),
+                "images": images,
                 "already_tagged": tagged,
                 "to_tag": to_tag,
+                "estimate_seconds": estimate,
+                "estimate_device": device,
+                "suggestion_limit": AutotagConfig.suggestion_limit(),
                 "model": {
                     "type": model_type,
                     "label": config["label"],
@@ -862,9 +883,9 @@ class AutotagRoutesMixin:
         In order: 400 body or options, 400 ``review-required``; 409
         ``model-unavailable``; 507 ``insufficient-space``; 409 ``busy``
         (another records job, the legacy batch's hold included); 409
-        ``model-busy``. No await follows the busy check.
+        ``model-busy``. No await follows the busy check. The job tags each
+        output image (``_run_autotag_batch``), never a prompt (#379).
         """
-        from ..autotag import get_autotag_service
         from ..config import MAX_SUGGESTION_LIMIT, AutotagConfig
         from .annotation_suggest import OptionError, validate_options
 
@@ -873,7 +894,7 @@ class AutotagRoutesMixin:
         )
         if error is not None:
             return error
-        config, message = self._batch_choice(data)
+        _config, message = self._batch_choice(data)
         if message:
             return self._bad_request(message)
         model_type, existing, options = (
@@ -897,144 +918,24 @@ class AutotagRoutesMixin:
                 {"success": False, "code": "review-required", "error": REVIEW_REQUIRED},
                 status=400,
             )
-        service = get_autotag_service()
-        backup_on = self._safety_backup_policy() == "on"
         try:
-            status = await self._run_in_executor(service.get_models_status)
-            candidates = await self._run_in_executor(self.db.autotag_batch_candidates)
-            _disk, short, _size = await self._run_in_executor(
-                functools.partial(
-                    disk_plan, self.db.model.db_path, backup=backup_on, vacuum=False
-                )
+            job, refusal = await self._start_autotag(
+                model_type=model_type,
+                existing=existing,
+                options=options,
+                keep_in_memory=True,
+                title=AUTOTAG_JOB_TITLE,
+                owner_kind=AUTOTAG_OWNER_KIND,
             )
         except Exception as e:
             self.logger.exception("Tag output images could not start")
             return web.json_response(
                 {"success": False, "error": self._public_error(e)}, status=500
             )
-        if not (status.get(model_type) or {}).get("downloaded"):
-            return web.json_response(
-                {
-                    "success": False,
-                    "code": "model-unavailable",
-                    "error": f"{config['label']} is not downloaded. "
-                    "Download it in System › Auto-tagging.",
-                },
-                status=409,
-            )
-        if backup_on and short is not None:
-            return web.json_response(
-                {
-                    "success": False,
-                    "code": "insufficient-space",
-                    "error": short_space_message(short),
-                },
-                status=507,
-            )
-        busy = self._jobs.blocker(AUTOTAG_BATCH_JOB_KIND, admission="records")
-        if busy is not None:
-            return busy_response(JobBusy(busy))
-        owner = model_gate.new_owner(AUTOTAG_OWNER_KIND)
-        if not model_gate.try_acquire(owner):
-            return model_busy_response()
-        opts = {
-            "model_type": model_type,
-            "existing": existing,
-            "keep_in_memory": True,
-            "gate_owner": owner,
-            "general_threshold": options.get("general_threshold"),
-            "character_threshold": options.get("character_threshold"),
-            "prompt": options.get("instructions"),
-        }
-        try:
-            job = self._jobs.start(
-                AUTOTAG_BATCH_JOB_KIND,
-                AUTOTAG_JOB_TITLE,
-                admission="records",
-                cancellable=True,
-                scope={
-                    "model_type": model_type,
-                    "existing": existing,
-                    "images": len(candidates),
-                    "label": config["label"],
-                },
-                groups=AUTOTAG_GROUP_NAMES,
-                runner=functools.partial(
-                    self._run_autotag_job, service, opts, backup_on
-                ),
-            )
-        except BaseException as refused:
-            model_gate.release(owner)
-            if isinstance(refused, JobBusy):
-                return busy_response(refused)
-            raise
-        job.task.add_done_callback(lambda _task: model_gate.release(owner))
+        if refusal is not None:
+            return refusal.response
         return web.json_response({"success": True, "job": job.summary()}, status=202)
 
-    async def _run_autotag_job(self, service, opts, backup_on, job):
-        ok, _backup = await self._safety_backup_step(job, self.db, backup_on)
-        if not ok:
-            return
-        try:
-            await self._tag_output_images(service, opts, job)
-        except Exception:
-            self.logger.exception("Tag output images failed")
-            job.update(status_line=_INTERNAL_ERROR)
-            job.log_line(f"Tagging failed · {_INTERNAL_ERROR}")
-            job.finish("failed")
-        finally:
-            if backup_on:
-                await self._prune_backup_pool(job, self.db)
-
-    async def _tag_output_images(self, service, opts, job):
-        roots = _scan_roots(self._get_all_output_dirs())
-        job.log_line(
-            f"Tagging started · {job.scope['label']} · "
-            + ("skip prompts already tagged" if opts["existing"] == "skip" else "merge")
-        )
-        job.update(status_line="Loading the tagging model")
-
-        def on_item(index, total, group, image_path, detail):
-            path, root, root_index = _scan_location(image_path or "", roots)
-            job.update(
-                processed=index + 1,
-                total=total,
-                progress=100 * (index + 1) // total,
-                status_line=f"Tagging {index + 1} of {total}",
-            )
-            job.record(
-                group,
-                {
-                    "path": path,
-                    "root": root,
-                    "root_index": root_index,
-                    "detail": detail,
-                },
-                f"{path}: {detail}",
-            )
-            job.publish({"type": "progress", "progress": job.progress})
-
-        final = None
-        async for event in self._autotag_batch_events(
-            service,
-            {**opts, "cancelled": lambda: job.cancelled, "on_item": on_item},
-        ):
-            if event.get("type") in ("complete", "error"):
-                final = event
-        if final is not None and final["type"] == "error":
-            job.update(status_line=final["message"])
-            job.log_line(f"Tagging failed · {final['message']}")
-            job.finish("failed")
-        elif job.cancelled:  # a cancel wins over a run that then completed
-            job.update(status_line=BATCH_CANCELLED)
-            job.log_line(BATCH_CANCELLED)
-            job.finish("cancelled")
-        else:
-            done = f"Tagged {_plural(job.counters.get('tagged', 0), 'prompt')}"
-            job.update(status_line=done, progress=100)
-            job.log_line(done)
-            job.finish("finished")
-
     def _gated_runner(self, owner):
         """Executor calls under the model gate *owner* holds (plain when None)."""
         if owner is None:
diff --git a/py/api/image_tag_jobs.py b/py/api/image_tag_jobs.py
index 38fa9b6e..c2596af6 100644
--- a/py/api/image_tag_jobs.py
+++ b/py/api/image_tag_jobs.py
@@ -18,13 +18,21 @@ the suggestion worker and the storage call ``add_image_tags``. Files are
 handled one at a time and the job stops at a file boundary when cancelled,
 keeping what it already wrote. Rows and log lines carry root-relative paths
 and public root labels only.
+
+Auto-tag marks each image it ran the model on as auto-tagged (#379), with the
+union of its tags. The ``autotag`` job ("Tag output images", started by
+``POST /jobs/autotag``) runs the same per-file loop over every still image the
+output folders list (:meth:`ImageTagJobsMixin._run_autotag_batch`): image
+tags only, never a prompt tag or prompt text.
 """
 
 import asyncio
+import dataclasses
 import functools
 import os
 import time
 from dataclasses import dataclass, field
+from typing import NamedTuple
 
 from aiohttp import web
 
@@ -45,10 +53,17 @@ from .annotations import (
     _submit,
     _target,
 )
-from .autotag_routes import model_busy_response
-from .images import _public_path
+from .autotag_routes import BATCH_CANCELLED, MODEL_BUSY_ERROR, model_busy_response
+from .images import _current_fingerprint, _listing_posix, _public_path
 from .job_routes import busy_response
-from .jobs import IMAGE_TAG_GROUP_NAMES, IMAGE_TAG_JOB_KIND, JobBusy
+from .jobs import (
+    AUTOTAG_BATCH_JOB_KIND,
+    AUTOTAG_GROUP_NAMES,
+    IMAGE_TAG_GROUP_NAMES,
+    IMAGE_TAG_JOB_KIND,
+    JobBusy,
+)
+from .maintenance import disk_plan, short_space_message
 from .prompts import bad_request
 
 try:
@@ -107,6 +122,14 @@ READ_FAILED = "The file could not be read."
 SOURCE_CHANGED = "The source changed. Refresh the image collection."
 STORAGE_RESTORED = "Prompt storage was restored; start the job again."
 BUSY_RETRY = "The file could not be read now. Run the job again for it."
+# Tag output images (#379): the per-image batch.
+ALREADY_AUTOTAGGED = "Already auto-tagged"
+LISTING_FAILED = "The output folders could not be listed. Try again."
+LISTING_STATUS = "Listing output images…"
+_INTERNAL_ERROR = "An internal error occurred. Check server logs for details."
+# The pre-pass stats and looks up this many targets at a time, checking for a
+# cancel between chunks.
+PREPASS_CHUNK = 500
 
 _monotonic = time.monotonic  # the job's clock (a test can fake it)
 
@@ -246,6 +269,92 @@ async def _capacity_retried(job, attempt):
     return await attempt()
 
 
+class Refusal(NamedTuple):
+    """Why a Tag output images start was refused.
+
+    ``response`` is the native route's JSON answer; ``message`` the sentence
+    a legacy stream frame carries.
+    """
+
+    status: int
+    code: str
+    message: str
+    response: web.Response
+
+
+@dataclass(frozen=True)
+class AutotagRun:
+    """What one Tag output images job runs with (validated at its start)."""
+
+    model_type: str
+    existing: str
+    options: dict
+    keep_in_memory: bool
+    backup_on: bool
+    owner: object
+
+
+@dataclass(frozen=True)
+class _BatchTarget:
+    """One still image of the batch: the job target and its private keys."""
+
+    target: tuple  # (root_index, root_scope, relative): the image-tags target
+    path: str  # the listed file, for the pre-pass stat (never sent out)
+    locator: str  # the image tag key: normcase(join(realpath(root), relative))
+    label: str  # the root's public label
+
+
+def _still_targets(api, index, root, files, seen: set) -> list:
+    """Worker thread: one root's listed still images as _BatchTargets.
+
+    Newest first (the listing's order). A locator an earlier root already
+    produced (overlapping roots) is dropped; ``seen`` collects them.
+    """
+    scope = api._gallery_root_scope(root, index)
+    label = _public_path(root)
+    real = os.path.realpath(root)
+    prefix = os.path.join(str(root), "")
+    found = []
+    for path, _mtime in files:
+        relative = _listing_posix(root, prefix, path)
+        if is_video_path(relative):
+            continue
+        locator = os.path.normcase(os.path.join(real, *relative.split("/")))
+        if locator in seen:
+            continue
+        seen.add(locator)
+        found.append(_BatchTarget((index, scope, relative), str(path), locator, label))
+    return found
+
+
+def _prepass_chunk(db, chunk, skip_marked: bool, max_file: int):
+    """Worker thread: ``(remaining, skips)`` for a chunk of _BatchTargets.
+
+    Stats each file (no open, no hash): gone is "File not found", over the
+    capture limit "Too large to verify"; under "Skip already tagged" a file
+    whose current bytes are marked is "Already auto-tagged". ``skips`` are
+    ``(target, detail)`` pairs in target order.
+    """
+    stats, skips, alive = {}, [], []
+    for item in chunk:
+        fingerprint = _current_fingerprint(item.path)
+        if fingerprint is None:
+            skips.append((item, FILE_NOT_FOUND))
+        elif fingerprint[0] > max_file:
+            skips.append((item, TOO_LARGE))
+        else:
+            stats[item.locator] = fingerprint
+            alive.append(item)
+    marked = db.autotag_state(stats) if skip_marked and stats else {}
+    remaining = []
+    for item in alive:
+        if marked.get(item.locator):
+            skips.append((item, ALREADY_AUTOTAGGED))
+        else:
+            remaining.append(item)
+    return remaining, skips
+
+
 def _load_model(service, model_type: str, use_gpu: bool) -> None:
     """Worker thread: load (or swap to) *model_type*; never downloads."""
     if service.is_model_loaded() and service.get_loaded_model_type() == model_type:
@@ -254,6 +363,27 @@ def _load_model(service, model_type: str, use_gpu: bool) -> None:
         raise RuntimeError(f"{model_type} could not be loaded")
 
 
+def _refusal(status: int, code: str, message: str) -> Refusal:
+    response = web.json_response(
+        {"success": False, "code": code, "error": message}, status=status
+    )
+    return Refusal(status, code, message, response)
+
+
+def _busy_refusal(busy: JobBusy) -> Refusal:
+    title = getattr(busy.running, "title", "") or "Another job"
+    return Refusal(409, "busy", f"{title} already running", busy_response(busy))
+
+
+def _batch_cancelled(job):
+    """End the batch as cancelled; None (no service to hand back)."""
+    job.update(status_line=BATCH_CANCELLED)
+    job.log_line(BATCH_CANCELLED)
+    job.publish({"type": "cancelled", "progress": job.progress})
+    job.finish("cancelled")
+    return None
+
+
 class ImageTagJobsMixin:
     """The ``image-tags`` job kind: start route and runner."""
 
@@ -301,6 +431,216 @@ class ImageTagJobsMixin:
             job.task.add_done_callback(lambda _task: model_gate.release(owner))
         return web.json_response({"success": True, "job": job.summary()}, status=202)
 
+    # -- Tag output images (#379) ---------------------------------------------
+
+    async def _autotag_targets(self) -> list:
+        """Every still image the output folders list, as _BatchTargets.
+
+        Root by root in configured order, each newest first (the shared
+        listing, ``_get_gallery_files``, used as it is even when stale). Videos
+        are not targets; a file under two overlapping roots is one target.
+        """
+        seen: set = set()
+        found = []
+        for index, root in enumerate(list(self._get_all_output_dirs())):
+            files = await self._get_gallery_files(root)
+            found.extend(
+                await self._run_in_executor(
+                    _still_targets, self, index, root, files, seen
+                )
+            )
+        return found
+
+    async def _start_autotag(
+        self, *, model_type, existing, options, keep_in_memory, title, owner_kind
+    ):
+        """Register a Tag output images job: ``(job, None)`` or ``(None, Refusal)``.
+
+        ``options`` are already validated (``validate_options``). The checks
+        run in this order: 409 ``model-unavailable``; 507
+        ``insufficient-space`` (only when the safety backup is on); 409
+        ``busy`` (another records job); 409 ``model-busy``. No await follows
+        the busy check. The gate is taken before the job registers and
+        released when the task ends, also before its runner ever starts.
+        """
+        from ..autotag import MODELS, get_autotag_service
+
+        backup_on = self._safety_backup_policy() == "on"
+        status = await self._run_in_executor(get_autotag_service().get_models_status)
+        _disk, short, _size = await self._run_in_executor(
+            functools.partial(
+                disk_plan, self.db.model.db_path, backup=backup_on, vacuum=False
+            )
+        )
+        label = MODELS[model_type]["label"]
+        if not (status.get(model_type) or {}).get("downloaded"):
+            message = (
+                f"{label} is not downloaded. Download it in System › Auto-tagging."
+            )
+            return None, _refusal(409, "model-unavailable", message)
+        if backup_on and short is not None:
+            return None, _refusal(507, "insufficient-space", short_space_message(short))
+        busy = self._jobs.blocker(AUTOTAG_BATCH_JOB_KIND, admission="records")
+        if busy is not None:
+            return None, _busy_refusal(JobBusy(busy))
+        owner = model_gate.new_owner(owner_kind)
+        if not model_gate.try_acquire(owner):
+            response = model_busy_response()
+            return None, Refusal(409, "model-busy", MODEL_BUSY_ERROR, response)
+        run = AutotagRun(
+            model_type, existing, options, keep_in_memory, backup_on, owner
+        )
+        try:
+            job = self._jobs.start(
+                AUTOTAG_BATCH_JOB_KIND,
+                title,
+                admission="records",
+                cancellable=True,
+                scope={"model_type": model_type, "existing": existing, "label": label},
+                groups=AUTOTAG_GROUP_NAMES,
+                runner=functools.partial(self._run_autotag_batch, run),
+            )
+        except BaseException as refused:
+            model_gate.release(owner)
+            if isinstance(refused, JobBusy):
+                return None, _busy_refusal(refused)
+            raise
+        job.task.add_done_callback(lambda _task: model_gate.release(owner))
+        return job, None
+
+    async def _run_autotag_batch(self, run: AutotagRun, job) -> None:
+        """The ``autotag`` job: backup, model, listing, pre-pass, per-image loop.
+
+        ``job.cancelled`` is checked between phases (and per pre-pass chunk
+        and per image). Each phase publishes ``{"type": "phase", "phase"}``.
+        With ``keep_in_memory`` false the model is unloaded under the gate
+        before the gate is released, also after a cancel or a failure.
+        """
+        try:
+            await self._autotag_batch_phases(run, job)
+        except Exception:
+            self.logger.exception("Tag output images failed")
+            job.update(status_line=_INTERNAL_ERROR)
+            job.log_line(f"Tagging failed · {_INTERNAL_ERROR}")
+            job.finish("failed")
+        finally:
+            try:
+                if not run.keep_in_memory:
+                    await self._unload_batch_model(run.owner)
+                if run.backup_on:
+                    await self._prune_backup_pool(job, self.db)
+            finally:
+                model_gate.release(run.owner)
+
+    async def _unload_batch_model(self, owner) -> None:
+        from ..autotag import get_autotag_service
+
+        service = get_autotag_service()
+        try:
+            if service.is_model_loaded():
+                await model_gate.run(owner, service.unload_model)
+        except Exception:
+            self.logger.exception("The tagging model could not be unloaded")
+
+    async def _autotag_batch_phases(self, run: AutotagRun, job):
+        job.publish({"type": "phase", "phase": "backup"})
+        ok, _backup = await self._safety_backup_step(job, self.db, run.backup_on)
+        if not ok:
+            return None
+        if job.cancelled:
+            return _batch_cancelled(job)
+        job.publish({"type": "phase", "phase": "model"})
+        job.update(status_line="Loading the tagging model")
+        spec = _Spec("auto", (), model_type=run.model_type, options=run.options)
+        service = await self._prepare_tagging_model(job, spec, run.owner)
+        if service is None:
+            job.finish("failed")
+            return None
+        if job.cancelled:
+            return _batch_cancelled(job)
+        job.publish({"type": "phase", "phase": "listing"})
+        job.update(status_line=LISTING_STATUS)
+        epoch = await self._run_in_executor(self.db.get_execution_epoch)
+        try:
+            targets = await self._autotag_targets()
+        except Exception:
+            self.logger.exception("Tag output images could not list the outputs")
+            job.update(status_line=LISTING_FAILED)
+            job.log_line(f"! {LISTING_FAILED}")
+            job.finish("failed")
+            return None
+        total = len(targets)
+        job.scope["images"] = total
+        job.update(total=total)
+        policy = (
+            "skip images already auto-tagged" if run.existing == "skip" else "merge"
+        )
+        job.log_line(
+            f"Tagging started · {job.scope.get('label') or run.model_type} · "
+            f"{_plural(total, 'image')} · {policy}"
+        )
+        if job.cancelled:
+            return _batch_cancelled(job)
+        remaining = await self._autotag_prepass(job, targets, run.existing == "skip")
+        if remaining is None:
+            return _batch_cancelled(job)
+        done = total - len(remaining)
+        spec = dataclasses.replace(
+            spec, targets=tuple(item.target for item in remaining)
+        )
+        loop_run = _Run(job, spec, epoch, run.owner, service)
+        if not await self._tag_each_file(loop_run, done_before=done):
+            job.update(status_line=STORAGE_RESTORED)
+            job.log_line(f"Tagging stopped · {STORAGE_RESTORED}")
+            job.finish("failed")
+            return service
+        if job.cancelled:
+            return _batch_cancelled(job)
+        finished = f"Tagged {_plural(job.counters.get('tagged', 0), 'image')}"
+        job.update(status_line=finished, progress=100)
+        job.log_line(finished)
+        job.publish({"type": "finished", "progress": 100})
+        job.finish("finished")
+        return service
+
+    async def _autotag_prepass(self, job, targets, skip_marked: bool):
+        """The targets left for the model, or None when cancelled meanwhile.
+
+        Opens and hashes nothing: one stat per file and one indexed read per
+        chunk. Skips are counted with ``record_counted`` (exact counters, rows
+        up to the cap, no log line each).
+        """
+        max_file = self._capture_store().max_file
+        remaining, skipped = [], 0
+        total = len(targets)
+        for start in range(0, total, PREPASS_CHUNK):
+            if job.cancelled:
+                return None
+            chunk = targets[start : start + PREPASS_CHUNK]
+            left, skips = await self._run_in_executor(
+                _prepass_chunk, self.db, chunk, skip_marked, max_file
+            )
+            remaining.extend(left)
+            if skips:
+                job.record_counted(
+                    "skipped",
+                    len(skips),
+                    [
+                        {
+                            "path": item.target[2],
+                            "root": item.label,
+                            "root_index": item.target[0],
+                            "detail": detail,
+                        }
+                        for item, detail in skips
+                    ],
+                )
+                skipped += len(skips)
+                job.update(processed=skipped, progress=skipped * 100 // total)
+        if skipped:
+            job.log_line(f"Skipped {_plural(skipped, 'image')} before tagging")
+        return remaining
+
     # -- runner ----------------------------------------------------------------
 
     def _job_roots(self, targets) -> dict:
@@ -342,17 +682,24 @@ class ImageTagJobsMixin:
             if owner is not None:
                 model_gate.release(owner)
 
-    async def _tag_each_file(self, run: _Run) -> bool:
-        """Tag the targets in order; False when the job had to stop."""
-        job, total = run.job, len(run.spec.targets)
+    async def _tag_each_file(self, run: _Run, *, done_before: int = 0) -> bool:
+        """Tag the targets in order; False when the job had to stop.
+
+        ``done_before`` files were already counted (the batch's pre-pass): file
+        n of this loop is "file done_before + n" of the job, and the time left
+        uses this loop's rate only.
+        """
+        job, count = run.job, len(run.spec.targets)
+        total = done_before + count
         run.roots = await self._run_in_executor(self._job_roots, run.spec.targets)
         started = _monotonic()
         for number, target in enumerate(run.spec.targets, 1):
             if job.cancelled:
                 break
             # The scan's estimate (D56), with each finished file as one batch.
-            left = _time_left(number - 1, total, number - 1, _monotonic() - started)
-            job.update(status_line=f"Processing file {number:,} of {total:,}{left}")
+            left = _time_left(number - 1, count, number - 1, _monotonic() - started)
+            shown = done_before + number
+            job.update(status_line=f"Processing file {shown:,} of {total:,}{left}")
             job.publish({"type": "progress", "progress": job.progress})
             stopped = False
             try:
@@ -363,7 +710,7 @@ class ImageTagJobsMixin:
                 self.logger.exception("Image tagging failed for one file")
                 group, detail = "errors", SAVE_FAILED
             self._record_tag_outcome(run, target, group, detail)
-            job.update(processed=number, progress=number * 100 // total)
+            job.update(processed=shown, progress=shown * 100 // total)
             if stopped:
                 return False
         return True
@@ -439,9 +786,17 @@ class ImageTagJobsMixin:
             store.release(capture.token)
 
     async def _auto_tag_capture(self, run: _Run, store, capture, root, relative):
-        """Suggest on the capture, then union only onto the same current bytes."""
+        """Suggest on the capture, then union only onto the same current bytes.
+
+        The bytes are marked auto-tagged with the union, also when the model
+        returned no names; captured bytes no model can read are marked too
+        (and stay skipped), so a "Skip already tagged" run passes them by.
+        """
         if not capture.inferable:
-            return "skipped", NOT_A_STILL
+            group, detail = await self._union_image_tags(
+                capture, [], run.epoch, mark_autotagged=True
+            )
+            return (group, detail) if group == "errors" else ("skipped", NOT_A_STILL)
         if store.acquire(capture.token) is None:
             raise _JobStopped  # every capture was invalidated (restore)
         spec = run.spec
@@ -467,21 +822,30 @@ class ImageTagJobsMixin:
         if facts.locator != capture.locator or facts.sha256 != capture.sha256:
             return "skipped", FILE_CHANGED
         names = [str(name) for name, _score in pairs]
-        if not names:
-            return "unchanged", NO_NEW_TAGS
-        return await self._union_image_tags(facts, names, run.epoch)
+        return await self._union_image_tags(
+            facts, names, run.epoch, mark_autotagged=True
+        )
 
-    async def _union_image_tags(self, facts, tags: list, epoch: str):
-        """``add_image_tags`` on the verified bytes, as a (group, detail)."""
+    async def _union_image_tags(
+        self, facts, tags: list, epoch: str, *, mark_autotagged: bool = False
+    ):
+        """``add_image_tags`` on the verified bytes, as a (group, detail).
+
+        ``mark_autotagged`` (a model's result) marks the bytes auto-tagged in
+        the same write; then ``tags`` may be empty.
+        """
         try:
             receipt = await self._run_in_executor(
-                self.db.add_image_tags,
-                locator=facts.locator,
-                sha256=facts.sha256,
-                size=facts.size,
-                mtime_ns=facts.mtime_ns,
-                expected_epoch=epoch,
-                tags=tags,
+                functools.partial(
+                    self.db.add_image_tags,
+                    locator=facts.locator,
+                    sha256=facts.sha256,
+                    size=facts.size,
+                    mtime_ns=facts.mtime_ns,
+                    expected_epoch=epoch,
+                    tags=tags,
+                    mark_autotagged=mark_autotagged,
+                )
             )
         except LibraryCommandError as exc:
             if exc.code == "storage-changed":
diff --git a/py/autotag.py b/py/autotag.py
index 9a6d55b1..8938ed5f 100644
--- a/py/autotag.py
+++ b/py/autotag.py
@@ -756,6 +756,24 @@ class AutoTagService:
 
             self.logger.info("Model unloaded, memory freed")
 
+    def wd14_device(self) -> str:
+        """``"gpu"`` when WD14 runs (or would run) on CUDA, else ``"cpu"``.
+
+        A loaded WD14 session answers from its own first provider; otherwise
+        ONNX Runtime's available providers say whether CUDA would be offered
+        (a session can still fall back to the processor). Loads no model.
+        """
+        tagger = self._tagger
+        if tagger is not None and tagger[0] == "wd14":
+            providers = tagger[1][0].get_providers()
+            return "gpu" if providers[:1] == ["CUDAExecutionProvider"] else "cpu"
+        try:
+            import onnxruntime as ort
+        except ImportError:
+            return "cpu"
+        available = ort.get_available_providers()
+        return "gpu" if "CUDAExecutionProvider" in available else "cpu"
+
     def is_model_loaded(self) -> bool:
         """Check if a model is currently loaded."""
         return self._tagger is not None
diff --git a/py/tagging_defaults.py b/py/tagging_defaults.py
index be75c9e7..b9c35f42 100644
--- a/py/tagging_defaults.py
+++ b/py/tagging_defaults.py
@@ -13,7 +13,23 @@ LLM_MODEL_TYPES = frozenset({"gguf", "hf"})
 MODEL_TYPES = WD14_MODEL_TYPES | LLM_MODEL_TYPES
 MAX_INSTRUCTION_CHARS = 4000
 MAX_INSTRUCTION_BYTES = 16 * 1024
-# What a batch does with a prompt that already has tags (#328).
+# Tag output images' review estimate (#379): seconds per image for the WD14
+# engines, by where ONNX Runtime runs them. The processor figure is an upper
+# bound from the maintainer's cancelled CPU run (19 min / 697 images, backup,
+# model load and skips included). The graphics card figure is provisional
+# until PR 1's manual check measures Auto-tag selected on 100 images with CUDA;
+# it is corrected there when the measurement is off by more than a quarter.
+# The language-model engines have no figure.
+WD14_SECONDS_PER_IMAGE = 1.6
+WD14_GPU_SECONDS_PER_IMAGE = 0.25
+
+
+def wd14_seconds_per_image(device: str) -> float:
+    """The estimate's rate for ``"gpu"`` or ``"cpu"`` (anything else)."""
+    return WD14_GPU_SECONDS_PER_IMAGE if device == "gpu" else WD14_SECONDS_PER_IMAGE
+
+
+# What a batch does with an image that already has tags (#328, #379).
 EXISTING_POLICIES = ("skip", "merge", "replace_after_review")
 
 # The language-model tag template every page starts from until one is saved.
diff --git a/tests/browser/system-tagging-concept.spec.mjs b/tests/browser/system-tagging-concept.spec.mjs
index 61b19710..b02f5a83 100644
--- a/tests/browser/system-tagging-concept.spec.mjs
+++ b/tests/browser/system-tagging-concept.spec.mjs
@@ -33,7 +33,7 @@ test.afterEach(async ({ page, request, browserName }) => {
 const PREVIEW_URL = '/prompt_manager/autotag/batch/preview';
 const MODEL_OPTIONS = ['WD14 SwinV2 · ~400 MB', 'WD14 ViT · ~400 MB', 'GGUF JoyCaption · ~4.5 GB', 'HuggingFace JoyCaption · ~16 GB'];
 const SUGGESTIONS = 'From 1 to 100: the most suggestions Generate shows for one image.';
-const BATCH = 'Add tags to each prompt from its linked output images with the model above. Review the scope first.';
+const BATCH = 'Tag each output image from its own pixels with the model above. Prompts are not changed. Review the scope first.';
 
 /** Runs in the page: Auto-tagging's surface under `mainSelector`: rows, field, actions, operation, geometry and tokens. */
 function measureTagging(mainSelector) {
@@ -272,11 +272,19 @@ function measureReview() {
 }
 
 const number = value => value.toLocaleString('en-US');
+/** DEC11 (#379): the Time row the preview's estimate reads as (WD14 on the test server's processor). */
+function timeRow(plan) {
+    const minutes = Math.max(1, Math.ceil(plan.estimate_seconds / 60));
+    const text = minutes < 90 ? `about ${minutes} min`
+        : `about ${Math.floor(minutes / 60)} h${minutes % 60 ? ` ${minutes % 60} min` : ''}`;
+    return `${text} on a processor; a graphics card can be faster.`;
+}
 
 /**
- * D134: the server's batch preview in the concept's review: "Review tag output images" with "Tag output images" as its
- * `h3`, the model, policy, counts and settings in the concept's `dl.meta-grid` (Backup in the concept's wording), the
- * upper-bound sentence in the concept's `p.small`, the D55 backup box and Cancel / "Start tagging".
+ * D134 (#379): the server's batch preview in the concept's review: "Review tag output images" with "Tag output images"
+ * as its `h3`, the model, policy, image counts, time estimate and settings in the concept's `dl.meta-grid` (Backup in
+ * the concept's wording), the estimate sentence in the concept's `p.small`, the D55 backup box and Cancel / "Start
+ * tagging".
  */
 function mapBatchConcept(reference, plan) {
     const [cancel, start] = reference.actions.buttons;
@@ -286,13 +294,14 @@ function mapBatchConcept(reference, plan) {
         tree: ['h3', 'dl.meta-grid', 'p.small', 'div.notice', 'div.dialog-actions'],
         heading: { ...reference.heading, text: 'Tag output images' },
         grid: { ...reference.grid, rows: [['Model', 'WD14 SwinV2 · ~400 MB'], ['Existing tags', 'Merge with existing'],
-            ['Images linked', number(plan.images)], ['Prompts', number(plan.prompts)], ['Already tagged', number(plan.already_tagged)],
-            ['Prompts to tag', `up to ${number(plan.to_tag)}`], ['Settings', 'General 0.35 · Character 0.85'],
+            ['Images', number(plan.images)], ['Already auto-tagged', number(plan.already_tagged)],
+            ['Images to tag', `about ${number(plan.to_tag)}`], ['Time', timeRow(plan)],
+            ['Settings', `General 0.35 · Character 0.85 · up to ${number(plan.suggestion_limit)} tags per image`],
             ['Backup', 'Create and verify before changes']] },
         backup: { tree: ['div', 'div.row-actions'], text: [['h3', 'Back up the database first (recommended)'],
             ['p', 'Downloads a copy of the prompts database before this operation changes any records.']], buttons: ['Back up now'] },
-        note: { ...reference.note, text: 'Nothing is written until you start. Missing files and files outside the output folders '
-            + 'are counted, so the prompts to tag are an upper bound; Activity reports what was tagged.' },
+        note: { ...reference.note, text: 'Nothing is written until you start. The number to tag is an estimate: images changed '
+            + 'since they were tagged are tagged again, and missing files are skipped. Activity reports what was tagged.' },
         actions: { ...reference.actions, buttons: [cancel, { ...start, text: 'Start tagging' }] },
     };
 }
@@ -308,8 +317,8 @@ async function conceptReview(page, theme) {
 
 async function productionReview(page, theme) {
     await productionTagging(page, theme);
-    // Every concept-fixture prompt with an image is already tagged, so under the saved Skip there is nothing to start
-    // (the Cancel-only review). Merge reviews a batch with work: the full review the concept's shell is compared with.
+    // Merge reviews every listed still image (#379): a batch with work, the full review the concept's shell is compared
+    // with.
     await page.locator('#workspace-main #tag-existing').selectOption('merge');
     const answer = page.waitForResponse(response => response.url().endsWith(PREVIEW_URL));
     await page.locator('#workspace-main .operation').getByRole('button', { name: 'Review', exact: true }).click();
@@ -326,8 +335,9 @@ for (const theme of THEMES) for (const [width, height] of SIZES) {
         expect(concept.tree).toEqual(['p.demo-note', 'h3', 'dl.meta-grid', 'p.small', 'div.dialog-actions']);
         await page.screenshot({ path: testInfo.outputPath(`concept-plan-review-${theme}-${width}.png`) });
         const { plan, production } = await productionReview(page, theme);
-        // Under Merge every prompt with a linked image is tagged again: the batch has work to review.
+        // Under Merge every listed still image is tagged again: the batch has work to review.
         expect(plan.to_tag).toBeGreaterThan(0);
+        expect(plan.estimate_device).toBe('cpu');
         await page.screenshot({ path: testInfo.outputPath(`production-tagging-review-${theme}-${width}.png`) });
         expect(diffInventory(mapBatchConcept(concept, plan), production)).toEqual([]);
     });
@@ -372,9 +382,17 @@ test('tag output images review comparison detects mutations', async ({ page }) =
     const expected = mapBatchConcept(concept, plan);
     const paths = async () => diffInventory(expected, await page.evaluate(measureReview)).map(item => item.path);
     expect(await paths()).toEqual([]);
-    await page.evaluate(() => { document.querySelectorAll('dialog#modal[open] dl.meta-grid dt')[5].textContent = 'To tag'; });
-    expect(await paths()).toEqual(['grid.rows.5.0']);
-    await page.evaluate(() => { document.querySelectorAll('dialog#modal[open] dl.meta-grid dt')[5].textContent = 'Prompts to tag'; });
+    // Every label of the grid is compared (#379's rows included): a changed one is named by its row.
+    for (const [index, [term]] of expected.grid.rows.entries()) {
+        await page.evaluate(([at, label]) => {
+            document.querySelectorAll('dialog#modal[open] dl.meta-grid dt')[at].textContent = label;
+        }, [index, `${term}!`]);
+        expect(await paths()).toEqual([`grid.rows.${index}.0`]);
+        await page.evaluate(([at, label]) => {
+            document.querySelectorAll('dialog#modal[open] dl.meta-grid dt')[at].textContent = label;
+        }, [index, term]);
+    }
+    expect(await paths()).toEqual([]);
     const undoGrid = await mutate(page, { css: 'dialog#modal dl.meta-grid { column-gap: 3px; }' });
     expect(await paths()).toEqual(expect.arrayContaining(['grid.columnGap']));
     await undoGrid();
diff --git a/tests/browser/workspace-system-tagging.spec.mjs b/tests/browser/workspace-system-tagging.spec.mjs
index a2720b17..0d9fa58a 100644
--- a/tests/browser/workspace-system-tagging.spec.mjs
+++ b/tests/browser/workspace-system-tagging.spec.mjs
@@ -315,22 +315,28 @@ test('Preview tagging review without a selection toasts the D115 sentence', asyn
     await expect(tagDialog(page)).toHaveCount(0);
 });
 
-test('Tag output images reviews the scope and starts a batch job', async ({ page, request }) => {
+test('Tag output images reviews the images, tags each one and never a prompt (#379)', async ({ page, request }) => {
     const before = await serverJson(request, '/__test__/tagging/snapshot');
     await openTagging(page);
-    // The fixture's one prompt is already tagged: under Merge it is tagged again with what its images give.
-    await page.locator('#tag-existing').selectOption('merge');
+    // The saved policy is Skip: no image is auto-tagged yet, so every listed still image is to tag.
     const previewed = page.waitForResponse(item => new URL(item.url()).pathname === PREVIEW_URL);
     await reviewButton(page).click();
     const answer = await previewed;
-    expect(answer.request().postDataJSON()).toEqual({ model_type: 'wd14-swinv2', existing: 'merge' });
+    expect(answer.request().postDataJSON()).toEqual({ model_type: 'wd14-swinv2', existing: 'skip' });
     const plan = await answer.json();
+    // Four still images (two linked, a copy, an unlinked one); the video is not a target.
+    expect([plan.images, plan.already_tagged, plan.to_tag]).toEqual([4, 0, 4]);
+    expect(plan.prompts).toBeUndefined();
     await expect(dialog(page).locator('#modal-title')).toHaveText('Review tag output images');
     await expect(dialog(page).locator('#modal-body > h3')).toHaveText('Tag output images');
-    expect(await gridOf(page)).toEqual([['Model', 'WD14 SwinV2 · ~400 MB'], ['Existing tags', 'Merge with existing'],
-        ['Images linked', String(plan.images)], ['Prompts', String(plan.prompts)], ['Already tagged', String(plan.already_tagged)],
-        ['Prompts to tag', `up to ${plan.to_tag}`], ['Settings', 'General 0.35 · Character 0.85'],
+    expect(await gridOf(page)).toEqual([['Model', 'WD14 SwinV2 · ~400 MB'], ['Existing tags', 'Skip already tagged'],
+        ['Images', '4'], ['Already auto-tagged', '0'], ['Images to tag', 'about 4'],
+        ['Time', 'about 1 min on a processor; a graphics card can be faster.'],
+        ['Settings', `General 0.35 · Character 0.85 · up to ${plan.suggestion_limit} tags per image`],
         ['Backup', 'Create and verify before changes']]);
+    await expect(dialog(page).locator('#modal-body > p.small')).toHaveText('Nothing is written until you start. The number to '
+        + 'tag is an estimate: images changed since they were tagged are tagged again, and missing files are skipped. '
+        + 'Activity reports what was tagged.');
     await expect(dialog(page).locator('#modal-body > .notice h3')).toHaveText('Back up the database first (recommended)');
     // The review wrote nothing.
     expect(await serverJson(request, '/__test__/tagging/snapshot')).toEqual(before);
@@ -338,15 +344,23 @@ test('Tag output images reviews the scope and starts a batch job', async ({ page
     const started = page.waitForResponse(item => new URL(item.url()).pathname === '/prompt_manager/jobs/autotag');
     await dialog(page).getByRole('button', { name: 'Start tagging', exact: true }).click();
     const start = await started;
-    expect(start.request().postDataJSON()).toEqual({ model_type: 'wd14-swinv2', existing: 'merge',
+    expect(start.request().postDataJSON()).toEqual({ model_type: 'wd14-swinv2', existing: 'skip',
         options: { general_threshold: 0.35, character_threshold: 0.85 } });
     const { job } = await start.json();
     await expect(page).toHaveURL(`${workspace}#/activity/${job.id}`);
     await expect(page.locator('#workspace-main h1')).toHaveText('Tag output images');
     const finished = await ended(request, job.id);
     expect(finished.state).toBe('finished');
-    const after = await serverJson(request, '/__test__/tagging/snapshot');
-    expect(after.prompt_tags.length).toBeGreaterThan(before.prompt_tags.length);
+    expect(finished.status_line).toBe('Tagged 4 images');
+    // Each image got its own tags; no prompt, pair or link row changed.
+    expect(await serverJson(request, '/__test__/tagging/snapshot')).toEqual(before);
+
+    // Under Skip every image is now auto-tagged: nothing to start.
+    await openTagging(page);
+    await reviewButton(page).click();
+    await expect(dialog(page).locator('#modal-body > p.small')).toHaveText('Every output image is already auto-tagged.');
+    expect((await gridOf(page)).slice(2, 5)).toEqual([['Images', '4'], ['Already auto-tagged', '4'], ['Images to tag', 'about 0']]);
+    await expect(dialog(page).locator('.dialog-actions button')).toHaveText(['Cancel']);
 });
 
 test('Replace after review shows only Cancel with the reason', async ({ page, request }) => {
@@ -355,18 +369,13 @@ test('Replace after review shows only Cancel with the reason', async ({ page, re
     await page.locator('#tag-existing').selectOption('replace_after_review');
     await reviewButton(page).click();
     await expect(dialog(page).locator('#modal-title')).toHaveText('Review tag output images');
-    expect((await gridOf(page)).map(([term]) => term)).toEqual(['Model', 'Existing tags', 'Images linked', 'Prompts', 'Already tagged',
-        'Prompts to tag', 'Settings', 'Backup']);
+    expect((await gridOf(page)).map(([term]) => term)).toEqual(['Model', 'Existing tags', 'Images', 'Already auto-tagged',
+        'Images to tag', 'Time', 'Settings', 'Backup']);
     await expect(dialog(page).locator('#modal-body > p.small')).toHaveText('Replacing tags needs a review of each image. '
         + 'Use Preview tagging review, or choose Skip or Merge for a batch run.');
     await expect(dialog(page).locator('.dialog-actions button')).toHaveText(['Cancel']);
     await dialog(page).getByRole('button', { name: 'Cancel', exact: true }).click();
     await expect(dialog(page)).toHaveCount(0);
-    // Under Skip the fixture's only prompt is already tagged: nothing to start either.
-    await page.locator('#tag-existing').selectOption('skip');
-    await reviewButton(page).click();
-    await expect(dialog(page).locator('#modal-body > p.small')).toHaveText('Every prompt with an output image is already tagged.');
-    await expect(dialog(page).locator('.dialog-actions button')).toHaveText(['Cancel']);
     expect(await serverJson(request, '/__test__/tagging/snapshot')).toEqual(before);
     expect((await serverJson(request, '/prompt_manager/jobs')).jobs.filter(item => item.kind === 'autotag')).toEqual([]);
 });
diff --git a/tests/js/workspace-system-tagging.test.js b/tests/js/workspace-system-tagging.test.js
index a8a9c6d0..95c03b4a 100644
--- a/tests/js/workspace-system-tagging.test.js
+++ b/tests/js/workspace-system-tagging.test.js
@@ -360,19 +360,23 @@ test('Preview tagging review opens the Tag image dialog with the section’s val
     } finally { scope.dispose(); dom.restore(); }
 });
 
-const PLAN = { images: 3, prompts: 2, already_tagged: 1, to_tag: 1,
+const PLAN = { images: 3, already_tagged: 1, to_tag: 2, estimate_seconds: 3.2, estimate_device: 'cpu', suggestion_limit: 10,
     model: { type: 'wd14-swinv2', label: 'WD14 SwinV2', size_label: '~400 MB', downloaded: true }, existing: 'skip',
     review_required: false, backup: { policy: 'on' }, disk: { free_bytes: 10 * 2 ** 30, required_bytes: 2 ** 20, sufficient: true } };
 
-test('the batch review grid, body and Cancel-only reasons (D134, CUJ 4–5)', async () => {
-    const { batchGrid, batchBody, batchCloseNote, existingLabel, REPLACE_NOTE, NOTHING_TO_TAG, NO_LINKED_IMAGES } = await tagging;
+test('the batch review grid, body and Cancel-only reasons (D134, #379, CUJ 4–5)', async () => {
+    const { batchGrid, batchBody, batchCloseNote, existingLabel, REPLACE_NOTE, NOTHING_TO_TAG, NO_IMAGES } = await tagging;
     assert.deepEqual(batchGrid(PLAN, HOST), [['Model', 'WD14 SwinV2 · ~400 MB'], ['Existing tags', 'Skip already tagged'],
-        ['Images linked', '3'], ['Prompts', '2'], ['Already tagged', '1'], ['Prompts to tag', 'up to 1'],
-        ['Settings', 'General 0.35 · Character 0.85'], ['Backup', 'Create and verify before changes']]);
+        ['Images', '3'], ['Already auto-tagged', '1'], ['Images to tag', 'about 2'],
+        ['Time', 'about 1 min on a processor; a graphics card can be faster.'],
+        ['Settings', 'General 0.35 · Character 0.85 · up to 10 tags per image'], ['Backup', 'Create and verify before changes']]);
     const llm = { ...HOST, autotag_model_type: 'gguf' };
-    assert.deepEqual(batchGrid({ ...PLAN, model: { type: 'gguf' }, backup: { policy: 'off' } }, llm).slice(-2),
-        [['Settings', 'Template of 27 characters'], ['Backup', 'No safety backup (the policy is off)']]);
+    assert.deepEqual(batchGrid({ ...PLAN, model: { type: 'gguf' }, backup: { policy: 'off' }, estimate_seconds: null,
+        estimate_device: null, suggestion_limit: 4 }, llm).slice(-3),
+    [['Time', 'Language-model engines take longer per image. Activity shows the time left once tagging starts.'],
+        ['Settings', 'Template of 27 characters · up to 4 tags per image'], ['Backup', 'No safety backup (the policy is off)']]);
     assert.equal(batchGrid({}, HOST)[2][1], '0');
+    assert.equal(batchGrid({}, HOST)[6][1], 'General 0.35 · Character 0.85');
     assert.equal(existingLabel('odd'), 'odd');
     assert.deepEqual(batchBody(HOST), { model_type: 'wd14-swinv2', existing: 'skip',
         options: { general_threshold: 0.35, character_threshold: 0.85 } });
@@ -385,9 +389,31 @@ test('the batch review grid, body and Cancel-only reasons (D134, CUJ 4–5)', as
     assert.equal(batchCloseNote({ ...PLAN, model: { ...PLAN.model, downloaded: false } }), 'WD14 SwinV2 is not downloaded. Download it first.');
     assert.equal(batchCloseNote({ ...PLAN, disk: { free_bytes: 2 ** 20, required_bytes: 2 ** 21, sufficient: false } }),
         'Not enough free disk space: the safety backup needs about 2.0 MB and 1.0 MB is free. Nothing is written until there is room.');
-    assert.equal(batchCloseNote({ ...PLAN, prompts: 0, to_tag: 0 }), NO_LINKED_IMAGES);
+    assert.equal(batchCloseNote({ ...PLAN, images: 0, to_tag: 0 }), NO_IMAGES);
+    assert.equal(batchCloseNote({ ...PLAN, images: undefined }), NO_IMAGES);
+    assert.equal(NO_IMAGES, 'The output folders contain no images to tag.');
     assert.equal(batchCloseNote({ ...PLAN, to_tag: 0 }), NOTHING_TO_TAG);
-    assert.equal(NOTHING_TO_TAG, 'Every prompt with an output image is already tagged.');
+    assert.equal(NOTHING_TO_TAG, 'Every output image is already auto-tagged.');
+});
+
+test('the review Time row (DEC11, #379): minutes, hours and the engine it runs on', async () => {
+    const { durationText, timeText, TIME_UNKNOWN } = await tagging;
+    assert.equal(durationText(1), 'about 1 min');
+    assert.equal(durationText(0), 'about 1 min');
+    assert.equal(durationText(89), 'about 89 min');
+    assert.equal(durationText(90), 'about 1 h 30 min');
+    assert.equal(durationText(120), 'about 2 h');
+    assert.equal(durationText(631), 'about 10 h 31 min');
+    assert.equal(durationText(Number.NaN), 'about 1 min');
+    // 23,645 images at 1.6 s on a processor (the design's CUJ 1).
+    assert.equal(timeText({ estimate_seconds: 37832, estimate_device: 'cpu' }),
+        'about 10 h 31 min on a processor; a graphics card can be faster.');
+    assert.equal(timeText({ estimate_seconds: 59, estimate_device: 'gpu' }), 'about 1 min on the graphics card.');
+    assert.equal(timeText({ estimate_seconds: 0, estimate_device: 'cpu' }), 'about 1 min on a processor; a graphics card can be faster.');
+    for (const plan of [{ estimate_seconds: null }, {}, { estimate_seconds: -1 }, { estimate_seconds: Infinity }, null]) {
+        assert.equal(timeText(plan), TIME_UNKNOWN);
+    }
+    assert.equal(TIME_UNKNOWN, 'Language-model engines take longer per image. Activity shows the time left once tagging starts.');
 });
 
 test('showBatchPlan opens the review with the backup box, or the Cancel-only review', async () => {
diff --git a/tests/test_annotation_schema.py b/tests/test_annotation_schema.py
index 49611179..ce5c5027 100644
--- a/tests/test_annotation_schema.py
+++ b/tests/test_annotation_schema.py
@@ -35,6 +35,8 @@ REVIEW_TABLES = (
     "media_review_items",
     "media_review_metadata",
 )
+# The auto-tag marks (#379) cite annotation bytes too; an old backup lacks them.
+CITING_TABLES = REVIEW_TABLES + ("media_image_autotags",)
 # The exact DDL the plan pins; the module must produce precisely these objects.
 DDL = {
     METADATA: (
@@ -214,7 +216,7 @@ class AnnotationSchemaTests(unittest.TestCase):
         """
         with closing(sqlite3.connect(path)) as raw, raw:
             raw.execute("PRAGMA foreign_keys=OFF")
-            for table in REVIEW_TABLES + tuple(reversed(TABLES)):
+            for table in CITING_TABLES + tuple(reversed(TABLES)):
                 raw.execute(f"DROP TABLE {table}")
 
     def test_fresh_install_and_reopen_is_idempotent(self):
diff --git a/tests/test_annotations.py b/tests/test_annotations.py
index a5aa6002..a70e2198 100644
--- a/tests/test_annotations.py
+++ b/tests/test_annotations.py
@@ -561,6 +561,142 @@ class CurrentImageTagsTests(AnnotationTestCase):
         )
 
 
+class AutotagMarkTests(AnnotationTestCase):
+    """The auto-tagged mark (#379): per byte record, written with the tags."""
+
+    def marked(self):
+        conn = self.db.model.get_connection()
+        return sorted(
+            row[0] for row in conn.execute("SELECT * FROM media_image_autotags")
+        )
+
+    def byte_id(self, loc, digest=DIGEST_A):
+        conn = self.db.model.get_connection()
+        return conn.execute(
+            "SELECT b.id FROM media_annotation_bytes b JOIN media_annotation_files f "
+            "ON f.id=b.file_id WHERE f.locator=? AND b.sha256=?",
+            (loc, digest),
+        ).fetchone()[0]
+
+    def test_hand_tags_never_mark(self):
+        loc = self.locator("a.png")
+        self.replace(loc, tags=["keeper"])
+        self.add(loc, tags=["more"])
+        self.assertEqual(self.marked(), [])
+        self.assertEqual(self.db.autotag_state({loc: (SIZE, MTIME)}), {loc: False})
+        self.assertEqual(self.db.autotagged_locators(), set())
+
+    def test_model_union_marks_including_replayed_and_empty_results(self):
+        loc = self.locator("a.png")
+        receipt = self.add(loc, tags=["bird"], mark_autotagged=True)
+        self.assertEqual((receipt["added"], receipt["replayed"]), (1, False))
+        self.assertEqual(self.marked(), [self.byte_id(loc)])
+        # A replayed union and an empty result still mark (idempotently).
+        other = self.locator("b.png")
+        self.replace(other, DIGEST_B, ["keeper"])
+        replayed = self.add(other, DIGEST_B, ["KEEPER"], mark_autotagged=True)
+        self.assertTrue(replayed["replayed"])
+        empty = self.locator("c.png")
+        nothing = self.add(empty, tags=[], mark_autotagged=True)
+        self.assertEqual((nothing["tags"], nothing["added"]), ([], 0))
+        self.add(empty, tags=[], mark_autotagged=True)
+        self.assertEqual(
+            self.marked(),
+            sorted(
+                [self.byte_id(loc), self.byte_id(other, DIGEST_B), self.byte_id(empty)]
+            ),
+        )
+        self.assertEqual(
+            self.db.autotag_state(
+                {loc: (SIZE, MTIME), other: (SIZE, MTIME), empty: (SIZE, MTIME)}
+            ),
+            {loc: True, other: True, empty: True},
+        )
+        self.schema_clean()
+
+    def test_empty_union_without_mark_is_refused(self):
+        with self.assertRaises(ValueError):
+            self.add(self.locator("a.png"), tags=[])
+        self.assertEqual(self.counts()["media_annotation_files"], 0)
+
+    def test_tag_limit_union_writes_no_tag_and_no_mark(self):
+        loc = self.locator("a.png")
+        self.replace(loc, tags=[f"t{i}" for i in range(100)])
+        before = self.rows(*ANNOTATION_TABLES, "tags")
+        with self.assertRaises(AnnotationCommandError) as ctx:
+            self.add(loc, tags=["one more"], mark_autotagged=True)
+        self.assertEqual(ctx.exception.code, "tag-limit")
+        self.assertEqual(self.rows(*ANNOTATION_TABLES, "tags"), before)
+        self.assertEqual(self.marked(), [])
+
+    def test_replace_marks_on_saved_and_replayed_receipts_not_on_conflict(self):
+        loc = self.locator("a.png")
+        saved = self.replace(loc, tags=["a"], mark_autotagged=True)
+        self.assertFalse(saved["replayed"])
+        self.assertEqual(self.marked(), [self.byte_id(loc)])
+        other = self.locator("b.png")
+        self.replace(other, DIGEST_B, ["x"])
+        replayed = self.replace(
+            other, DIGEST_B, ["x"], revision=1, mark_autotagged=True
+        )
+        self.assertTrue(replayed["replayed"])
+        third = self.locator("c.png")
+        self.replace(third, tags=["y"])
+        with self.assertRaises(AnnotationCommandError):
+            self.replace(third, tags=["z"], revision=5, mark_autotagged=True)
+        self.assertEqual(
+            self.marked(), sorted([self.byte_id(loc), self.byte_id(other, DIGEST_B)])
+        )
+
+    def test_mark_follows_the_bytes_not_the_file(self):
+        loc = self.locator("a.png")
+        self.add(loc, tags=["bird"], mark_autotagged=True)
+        # The file's bytes changed: the new record is current and unmarked.
+        self.replace(loc, DIGEST_B, ["hand"], mtime_ns=MTIME + 1)
+        self.assertEqual(self.db.autotag_state({loc: (SIZE, MTIME + 1)}), {loc: False})
+        self.assertEqual(self.db.autotag_state({loc: (SIZE, MTIME)}), {loc: True})
+        # The review's estimate counts any marked record of the locator.
+        self.assertEqual(self.db.autotagged_locators(), {loc})
+        missing = self.locator("none.png")
+        self.assertEqual(self.db.autotag_state({missing: (1, 1)}), {missing: False})
+
+    def test_state_and_tags_choose_the_same_current_record(self):
+        loc = self.locator("a.png")
+        # Two records with the same verified fingerprint: the newest wins.
+        self.replace(loc, DIGEST_A, ["old bytes"])
+        with patch.object(
+            annotations_module, "utc_now_iso", return_value="2099-01-01T00:00:00"
+        ):
+            self.add(loc, DIGEST_B, ["new bytes"], mark_autotagged=True)
+        stats = {loc: (SIZE, MTIME)}
+        self.assertEqual(self.db.current_image_tags(stats), {loc: ["new bytes"]})
+        self.assertEqual(self.db.autotag_state(stats), {loc: True})
+        with patch.object(
+            annotations_module, "utc_now_iso", return_value="2199-01-01T00:00:00"
+        ):
+            self.add(loc, DIGEST_A, ["older bytes again"])
+        self.assertEqual(
+            self.db.current_image_tags(stats), {loc: ["old bytes", "older bytes again"]}
+        )
+        self.assertEqual(self.db.autotag_state(stats), {loc: False})
+
+    def test_marks_never_touch_prompt_rows(self):
+        conn = self.db.model.get_connection()
+        prompt_id = self.db.save_prompt("a beach", tags=["beach"])
+        before = [
+            [tuple(row) for row in conn.execute(f"SELECT * FROM {table} ORDER BY 1")]
+            for table in ("prompts", "prompt_tags")
+        ]
+        self.add(self.locator("a.png"), tags=["bird"], mark_autotagged=True)
+        self.replace(self.locator("b.png"), DIGEST_B, ["waves"], mark_autotagged=True)
+        after = [
+            [tuple(row) for row in conn.execute(f"SELECT * FROM {table} ORDER BY 1")]
+            for table in ("prompts", "prompt_tags")
+        ]
+        self.assertIsNotNone(prompt_id)
+        self.assertEqual(after, before)
+
+
 class AnnotationIsolationTests(AnnotationTestCase):
     def test_cascade_from_vocabulary_delete_keeps_byte_records(self):
         loc = self.locator("a.png")
diff --git a/tests/test_api_autotag.py b/tests/test_api_autotag.py
index 444594c7..da420590 100644
--- a/tests/test_api_autotag.py
+++ b/tests/test_api_autotag.py
@@ -1893,13 +1893,15 @@ class TestTagOutputImagesPreview(BatchJobTestCase):
         self.assertEqual(self.service.generate_calls, [])
         self.assertEqual(self.service.load_calls, [])
 
-    async def test_preview_counts_skip_and_merge(self):
+    async def test_preview_counts_listed_still_images_not_prompts(self):
         self._seed()
         skip = await (await self._preview(existing="skip")).json()
+        # t, f1, f2 and m are listed; gone.png is missing, o.png is outside.
         self.assertEqual(
-            {k: skip[k] for k in ("images", "prompts", "already_tagged", "to_tag")},
-            {"images": 6, "prompts": 5, "already_tagged": 1, "to_tag": 4},
+            {k: skip[k] for k in ("images", "already_tagged", "to_tag")},
+            {"images": 4, "already_tagged": 0, "to_tag": 4},
         )
+        self.assertNotIn("prompts", skip)
         self.assertEqual(
             skip["model"],
             {
@@ -1916,7 +1918,7 @@ class TestTagOutputImagesPreview(BatchJobTestCase):
             set(skip["disk"]), {"free_bytes", "required_bytes", "sufficient"}
         )
         merge = await (await self._preview(existing="merge")).json()
-        self.assertEqual(merge["to_tag"], 5)
+        self.assertEqual(merge["to_tag"], 4)
         leaked = [s for s in _strings(skip) if _absolute(s)]
         self.assertEqual(leaked, [])
 
@@ -1924,7 +1926,7 @@ class TestTagOutputImagesPreview(BatchJobTestCase):
         self._seed()
         data = await (await self._preview(existing="replace_after_review")).json()
         self.assertTrue(data["review_required"])
-        self.assertEqual(data["prompts"], 5)
+        self.assertEqual(data["images"], 4)
 
     async def test_preview_refuses_bad_bodies(self):
         for body in (
@@ -1979,119 +1981,6 @@ class TestTagOutputImagesJob(BatchJobTestCase):
             "WD14 ViT is not downloaded. Download it in System › Auto-tagging.",
         )
 
-    async def test_job_writes_what_legacy_writes(self):
-        ids = self._seed()
-        await self._run_job(existing="skip")
-        native = {name: sorted(self._tags(pid)) for name, pid in ids.items()}
-        # The same library through the legacy batch, in a fresh database.
-        for name, pid in ids.items():
-            self.api.db.update_prompt_metadata(
-                pid,
-                tags=(
-                    ["existing"]
-                    if name == "tagged"
-                    else (["auto-scanned"] if name == "marked" else [])
-                ),
-            )
-        resp = await self.client.post(
-            "/prompt_manager/autotag/start?model_type=wd14-vit&skip_tagged=true"
-        )
-        self.assertEqual(self._sse_events(await resp.text())[-1]["type"], "complete")
-        legacy = {name: sorted(self._tags(pid)) for name, pid in ids.items()}
-        self.assertEqual(native, legacy)
-        self.assertEqual(native["tagged"], ["existing"])
-        self.assertEqual(native["fresh"], ["1girl", "smile"])
-
-    async def test_job_outcomes_follow_the_mapping(self):
-        ids = self._seed()
-        self.api.db.update_prompt_metadata(
-            ids["marked"], tags=["auto-scanned", "1girl", "smile"]
-        )
-
-        def details(job):
-            return {
-                group: sorted(row["detail"] for row in rows)
-                for group, rows in job.outcomes.items()
-            }
-
-        merged = await self._run_job(existing="merge")
-        self.assertEqual(merged.state, "finished")
-        self.assertEqual(
-            details(merged),
-            {
-                "tagged": ["Added 2 tags", "Added 2 tags"],
-                "unchanged": ["No new tags"],
-                "skipped": [
-                    "File not found",
-                    "Outside the output folders",
-                    "Prompt handled by another image",
-                ],
-                "errors": [],
-            },
-        )
-        self.assertEqual(merged.status_line, "Tagged 2 prompts")
-
-        skipped = await self._run_job(existing="skip")
-        self.assertEqual(details(skipped)["tagged"], [])
-        self.assertEqual(details(skipped)["skipped"].count("Prompt already tagged"), 3)
-        self.assertEqual(skipped.status_line, "Tagged 0 prompts")
-
-    async def test_tagging_errors_are_fixed_strings(self):
-        self._seed()
-        self.service.generate_error = RuntimeError("/srv/secret/model exploded")
-        job = await self._run_job(existing="merge")
-        self.assertTrue(job.outcomes["errors"])
-        self.assertEqual(
-            {r["detail"] for r in job.outcomes["errors"]}, {"Tagging failed"}
-        )
-        self.assertNotIn("/srv/secret", json.dumps(job.report()))
-
-    async def test_job_progress_is_determinate(self):
-        self._seed()
-        seen = []
-        original = self.service.generate_tags
-
-        def generate(*args, **kwargs):
-            job = self.api._jobs.active("autotag")
-            seen.append((job.processed, job.total))
-            return original(*args, **kwargs)
-
-        self.service.generate_tags = generate
-        job = await self._run_job(existing="merge")
-        self.assertEqual((job.processed, job.total, job.progress), (6, 6, 100))
-        self.assertTrue(all(total == 6 for _p, total in seen[1:]), seen)
-        self.assertTrue(job.cancellable)
-
-    async def test_job_cancel_stops_before_the_next_image_and_keeps_writes(self):
-        ids = self._seed()
-        started = threading.Event()
-        proceed = threading.Event()
-        calls = []
-
-        def slow(image_path, **kwargs):
-            calls.append(image_path)
-            started.set()
-            proceed.wait(30)
-            return ["slow"]
-
-        self.service.generate_tags = slow
-        resp = await self._start_job(existing="merge")
-        job = self.api._jobs.get((await resp.json())["job"]["id"])
-        try:
-            await asyncio.get_running_loop().run_in_executor(None, started.wait, 30)
-            cancel = await self.client.post(f"/prompt_manager/jobs/{job.id}/cancel")
-            self.assertTrue((await cancel.json())["cancellation_requested"])
-        finally:
-            proceed.set()
-        await asyncio.wait_for(asyncio.shield(job.task), 30)
-        self.assertEqual(job.state, "cancelled")
-        self.assertEqual(job.status_line, "Cancelled. Tags already added were kept.")
-        self.assertEqual(len(calls), 1)
-        tagged = [n for n, pid in ids.items() if "slow" in self._tags(pid)]
-        self.assertEqual(len(tagged), 1)
-        self.assertIsNone(model_gate.holder())
-        self.assertEqual(self.service.unload_calls, 0)
-
     async def test_job_refused_while_legacy_batch_holds(self):
         async with self.api._jobs.hold(
             "autotag", "Tag output images (legacy page)", admission="records"
@@ -2103,41 +1992,6 @@ class TestTagOutputImagesJob(BatchJobTestCase):
         self.assertEqual(data["running"]["title"], "Tag output images (legacy page)")
         self.assertIsNone(model_gate.holder())
 
-    async def test_legacy_refused_while_job_runs(self):
-        self._seed()
-        started, proceed = threading.Event(), threading.Event()
-
-        def slow(image_path, **kwargs):
-            started.set()
-            proceed.wait(30)
-            return ["slow"]
-
-        self.service.generate_tags = slow
-        resp = await self._start_job(existing="merge")
-        job = self.api._jobs.get((await resp.json())["job"]["id"])
-        try:
-            await asyncio.get_running_loop().run_in_executor(None, started.wait, 30)
-            self.assertTrue(model_gate.holder().startswith("autotag-job:"))
-            legacy = await self.client.post("/prompt_manager/autotag/start")
-            # Both legacy pages read a stream: the refusal is one error frame
-            # that names the running job.
-            self.assertEqual(legacy.status, 200)
-            self.assertEqual(
-                self._sse_events(await legacy.text()),
-                [
-                    {
-                        "type": "error",
-                        "status": "error",
-                        "message": "Tag output images already running",
-                    }
-                ],
-            )
-            unload = await self.client.post("/prompt_manager/autotag/unload")
-            self.assertEqual((await unload.json())["operation"], "Tag output images")
-        finally:
-            proceed.set()
-        await asyncio.wait_for(asyncio.shield(job.task), 30)
-
     async def test_job_releases_gate_when_start_is_refused(self):
         class Running:
             id = "other"
@@ -2185,7 +2039,7 @@ class TestTagOutputImagesJob(BatchJobTestCase):
             ("database folder", "the safety backup", 10, 1),
             10,
         )
-        with patch("py.api.autotag_routes.disk_plan", return_value=short):
+        with patch("py.api.image_tag_jobs.disk_plan", return_value=short):
             resp = await self._start_job()
         data = await resp.json()
         self.assertEqual(resp.status, 507)
@@ -2203,14 +2057,6 @@ class TestTagOutputImagesJob(BatchJobTestCase):
         inside = [row for row in rows if row["root_index"] == 0]
         self.assertTrue(inside)
 
-    async def test_native_batch_leaves_custom_prompt_unchanged(self):
-        self._seed()
-        self.service.custom_prompt = "legacy page prompt"
-        await self._run_job(existing="merge", model_type="gguf")
-        self.assertEqual(self.service.custom_prompt, "legacy page prompt")
-        prompts = {call[1] for call in self.service.generate_calls}
-        self.assertEqual(prompts, {"Tags please."})
-
 
 # The legacy batch's per-image frames over BatchJobTestCase's library, newest
 # first, with every throttle check passing (pinned from main's generator).
@@ -2342,7 +2188,7 @@ class TestTagOutputImagesEdges(BatchJobTestCase):
     async def test_unexpected_error_fails_the_job_with_a_reason(self):
         self._seed()
         with patch.object(
-            self.api.db, "autotag_batch_candidates", side_effect=[[], RuntimeError("x")]
+            self.api, "_autotag_prepass", side_effect=RuntimeError("/srv/x")
         ):
             job = await self._run_job(existing="merge")
         self.assertEqual(job.state, "failed")
diff --git a/tests/test_autotag_batch.py b/tests/test_autotag_batch.py
new file mode 100644
index 00000000..ba557d39
--- /dev/null
+++ b/tests/test_autotag_batch.py
@@ -0,0 +1,634 @@
+"""Tag output images tags each image, never its prompt (#379, PR 1).
+
+``POST /prompt_manager/jobs/autotag`` runs the ``image-tags`` loop over every
+still image the output folders list, linked to a prompt or not: each image
+gets its own model tags in its image tag set and is marked auto-tagged.
+"Skip already tagged" skips images whose current bytes are marked, counted
+before the loop without opening them. Prompt tags and prompt text never
+change. No real model is loaded: the engine is a fake that answers per image.
+"""
+
+import asyncio
+import json
+import ntpath
+import os
+import posixpath
+import re
+import sqlite3
+import threading
+from pathlib import Path
+from unittest.mock import patch
+
+try:
+    from test_api_images import make_png
+    from test_image_tag_jobs import HookedTagger, ImageTagJobTestCase
+except ImportError:
+    from tests.test_api_images import make_png
+    from tests.test_image_tag_jobs import HookedTagger, ImageTagJobTestCase
+
+from database.execution_history_schema import rotate_history_epoch
+from py import model_gate
+from py.api import admin as admin_module
+from py.api import image_tag_jobs as jobs_module
+from py.config import PromptManagerConfig
+from utils.hashing import generate_prompt_hash
+
+JOB = "/prompt_manager/jobs/autotag"
+PREVIEW = "/prompt_manager/autotag/batch/preview"
+WAIT = 30.0
+WD14 = {"general_threshold": 0.35, "character_threshold": 0.85}
+CANCELLED = "Cancelled. Tags already added were kept."
+_DRIVE = re.compile(r"^[A-Za-z]:[\\/]")
+
+
+class PerImageTagger(HookedTagger):
+    """Answers each image from its own bytes: ``by_bytes`` maps exact file
+    contents to the names the "model" sees; any other image gets SCORED."""
+
+    def __init__(self):
+        super().__init__()
+        self.by_bytes = {}
+
+    def generate_tag_suggestions(self, image_path, **kwargs):
+        result = super().generate_tag_suggestions(image_path, **kwargs)
+        names = self.by_bytes.get(self.read_bytes[-1])
+        if names is None:
+            return result
+        return [(name, 0.9) for name in names][: kwargs["limit"]]
+
+
+def _strings(value):
+    if isinstance(value, str):
+        yield value
+    elif isinstance(value, dict):
+        for item in value.values():
+            yield from _strings(item)
+    elif isinstance(value, (list, tuple)):
+        for item in value:
+            yield from _strings(item)
+
+
+def _absolute(text):
+    return posixpath.isabs(text) or ntpath.isabs(text) or bool(_DRIVE.match(text))
+
+
+class BatchTestCase(ImageTagJobTestCase):
+    async def get_application(self):
+        app = await super().get_application()
+        self.tagger = PerImageTagger()
+        patcher = patch("py.autotag.get_autotag_service", return_value=self.tagger)
+        patcher.start()
+        self.addCleanup(patcher.stop)
+        backup = PromptManagerConfig.SAFETY_BACKUP
+        self.addCleanup(setattr, PromptManagerConfig, "SAFETY_BACKUP", backup)
+        PromptManagerConfig.SAFETY_BACKUP = False
+        return app
+
+    # -- helpers ---------------------------------------------------------
+
+    def image(self, name, color, sees=None):
+        path = make_png(self.output_dir / name, size=(8, 6), color=color)
+        if sees is not None:
+            self.tagger.by_bytes[path.read_bytes()] = list(sees)
+        return path
+
+    async def relist(self):
+        """Walk the output folder again (a cached listing is used as it is)."""
+        await self.api._listing(self.output_dir, fresh=True)
+
+    def prompt(self, text, tags=()):
+        return self.api.db.save_prompt(
+            text=text,
+            category=None,
+            tags=list(tags),
+            rating=None,
+            prompt_hash=generate_prompt_hash(text),
+        )
+
+    def link(self, prompt_id, path):
+        return self.api.db.link_image_to_prompt(prompt_id, str(path))
+
+    def prompt_rows(self):
+        conn = sqlite3.connect(self._temp_db.name)
+        try:
+            return {
+                table: conn.execute(f"SELECT * FROM {table} ORDER BY 1,2").fetchall()
+                for table in ("prompts", "prompt_tags")
+            }
+        finally:
+            conn.close()
+
+    def marked(self):
+        conn = sqlite3.connect(self._temp_db.name)
+        try:
+            return {
+                row[0]
+                for row in conn.execute(
+                    "SELECT f.locator FROM media_image_autotags m JOIN "
+                    "media_annotation_bytes b ON b.id=m.byte_id JOIN "
+                    "media_annotation_files f ON f.id=b.file_id"
+                )
+            }
+        finally:
+            conn.close()
+
+    async def preview(self, existing="skip", model_type="wd14-vit"):
+        response = await self.client.post(
+            PREVIEW, json={"model_type": model_type, "existing": existing}
+        )
+        return response.status, await self.json_of(response)
+
+    async def start_batch(self, existing="skip", model_type="wd14-vit", options=None):
+        if options is None:
+            options = (
+                dict(WD14)
+                if model_type.startswith("wd14")
+                else {"instructions": "Tags please."}
+            )
+        response = await self.client.post(
+            JOB,
+            json={"model_type": model_type, "existing": existing, "options": options},
+        )
+        return response.status, await self.json_of(response)
+
+    async def run_batch(self, **kwargs):
+        status, data = await self.start_batch(**kwargs)
+        self.assertEqual(status, 202, data)
+        job = self.api._jobs.get(data["job"]["id"])
+        await asyncio.wait_for(asyncio.shield(job.task), WAIT)
+        return job
+
+
+class TestPerImageWrites(BatchTestCase):
+    async def test_two_images_of_one_prompt_get_their_own_tags_and_no_prompt_changes(
+        self,
+    ):
+        prompt_id = self.prompt("a girl at the beach", tags=["beach"])
+        a = self.image("a.png", (10, 20, 30), sees=["1girl", "bird"])
+        b = self.image("b.png", (40, 50, 60), sees=["1girl", "waves"])
+        self.link(prompt_id, a)
+        self.link(prompt_id, b)
+        before = self.prompt_rows()
+
+        job = await self.run_batch(existing="skip")
+
+        self.assertEqual(job.state, "finished", job.status_line)
+        self.assertEqual(self.stored("a.png")["tags"], ["1girl", "bird"])
+        self.assertEqual(self.stored("b.png")["tags"], ["1girl", "waves"])
+        self.assertEqual(self.prompt_rows(), before)
+        self.assertEqual(self.api.db.get_prompt_by_id(prompt_id)["tags"], ["beach"])
+        self.assertEqual(self.marked(), {self.locator("a.png"), self.locator("b.png")})
+        self.assertEqual(
+            job.counters, {"tagged": 2, "unchanged": 0, "skipped": 0, "errors": 0}
+        )
+        self.assertEqual(job.status_line, "Tagged 2 images")
+        self.assertEqual((job.processed, job.total, job.progress), (2, 2, 100))
+        self.assertEqual(job.scope["images"], 2)
+        # Inference read the private captures, one per image.
+        self.assertEqual(len(self.tagger.calls), 2)
+
+    async def test_unlinked_images_are_tagged_and_videos_are_not_targets(self):
+        self.image("free.png", (1, 2, 3), sees=["tree"])
+        (self.output_dir / "clip.mp4").write_bytes(b"\x00\x00\x00\x18ftypmp42" * 8)
+
+        job = await self.run_batch()
+
+        self.assertEqual(self.stored("free.png")["tags"], ["tree"])
+        self.assertEqual(job.total, 1)
+        self.assertEqual(job.status_line, "Tagged 1 image")
+        rows = [row["path"] for rows in job.outcomes.values() for row in rows]
+        self.assertNotIn("clip.mp4", rows)
+
+    async def test_hand_tags_do_not_block_a_skip_run(self):
+        path = self.image("c.png", (7, 7, 7), sees=["bird"])
+        self.saved_set("c.png", path.read_bytes(), ["keeper"])
+
+        job = await self.run_batch(existing="skip")
+
+        self.assertEqual(self.stored("c.png")["tags"], ["keeper", "bird"])
+        self.assertEqual(job.counters["tagged"], 1)
+
+    async def test_skip_counts_marked_images_without_opening_them_and_merge_retags(
+        self,
+    ):
+        self.image("a.png", (1, 1, 1), sees=["bird"])
+        self.image("b.png", (2, 2, 2), sees=["cat"])
+        first = await self.run_batch()
+        self.assertEqual(first.counters["tagged"], 2)
+        self.tagger.calls.clear()
+
+        captures = []
+        store = self.api._capture_store()
+        real_capture = store.capture
+
+        def spy(*args, **kwargs):
+            captures.append(args)
+            return real_capture(*args, **kwargs)
+
+        with patch.object(store, "capture", side_effect=spy):
+            skipped = await self.run_batch(existing="skip")
+        self.assertEqual(captures, [])
+        self.assertEqual(self.tagger.calls, [])
+        self.assertEqual(
+            skipped.counters, {"tagged": 0, "unchanged": 0, "skipped": 2, "errors": 0}
+        )
+        self.assertEqual(
+            sorted(self.rows(skipped, "skipped")),
+            [("a.png", "Already auto-tagged"), ("b.png", "Already auto-tagged")],
+        )
+        self.assertFalse(any("Already auto-tagged" in line for line in skipped.log))
+        self.assertEqual((skipped.processed, skipped.total), (2, 2))
+        self.assertEqual(skipped.status_line, "Tagged 0 images")
+
+        self.tagger.by_bytes[(self.output_dir / "a.png").read_bytes()] = ["bird", "sky"]
+        merged = await self.run_batch(existing="merge")
+        self.assertEqual(len(self.tagger.calls), 2)
+        self.assertEqual(self.stored("a.png")["tags"], ["bird", "sky"])
+        self.assertEqual(self.rows(merged, "tagged"), [("a.png", "Added 1 tag")])
+        self.assertEqual(self.rows(merged, "unchanged"), [("b.png", "No new tags")])
+
+    async def test_changed_bytes_are_tagged_again(self):
+        self.image("a.png", (1, 1, 1), sees=["bird"])
+        await self.run_batch()
+        self.image("a.png", (9, 9, 9), sees=["cat"])
+        job = await self.run_batch(existing="skip")
+        self.assertEqual(self.rows(job, "tagged"), [("a.png", "Added 1 tag")])
+        self.assertEqual(self.stored("a.png")["tags"], ["cat"])
+
+    async def test_an_image_with_no_model_tags_is_marked_and_skipped_next_time(self):
+        self.image("plain.png", (3, 3, 3), sees=[])
+        job = await self.run_batch()
+        self.assertEqual(self.rows(job, "unchanged"), [("plain.png", "No new tags")])
+        self.assertEqual(self.marked(), {self.locator("plain.png")})
+        again = await self.run_batch()
+        self.assertEqual(
+            self.rows(again, "skipped"), [("plain.png", "Already auto-tagged")]
+        )
+
+    async def test_unsupported_still_is_marked_and_skipped_by_the_next_run(self):
+        (self.output_dir / "broken.png").write_bytes(b"not an image")
+        job = await self.run_batch()
+        self.assertEqual(
+            self.rows(job, "skipped"), [("broken.png", "Not a supported still image")]
+        )
+        self.assertEqual(self.marked(), {self.locator("broken.png")})
+        again = await self.run_batch()
+        self.assertEqual(
+            self.rows(again, "skipped"), [("broken.png", "Already auto-tagged")]
+        )
+        self.assertEqual(self.tagger.calls, [])
+
+    async def test_too_large_is_skipped_before_any_copy(self):
+        self.image("big.png", (5, 5, 5))
+        self.api._capture_store().max_file = 16
+        store = self.api._capture_store()
+        with patch.object(store, "capture", side_effect=AssertionError("copied")):
+            job = await self.run_batch()
+        self.assertEqual(
+            self.rows(job, "skipped"), [("big.png", "Too large to verify")]
+        )
+        self.assertEqual(self.marked(), set())
+
+    async def test_overlapping_roots_list_a_file_once(self):
+        self.image("sub/a.png", (4, 4, 4), sees=["bird"])
+        sub = self.output_dir / "sub"
+        self.api._get_all_output_dirs = lambda: [self.output_dir, sub]
+        job = await self.run_batch()
+        self.assertEqual(job.total, 1)
+        self.assertEqual(len(self.tagger.calls), 1)
+        self.assertEqual(self.rows(job, "tagged"), [("sub/a.png", "Added 1 tag")])
+        status, data = await self.preview()
+        self.assertEqual((status, data["images"]), (200, 1))
+
+    async def test_rows_log_and_status_carry_no_absolute_paths(self):
+        self.image("a.png", (1, 1, 1))
+        (self.output_dir / "broken.png").write_bytes(b"not an image")
+        job = await self.run_batch()
+        leaked = [s for s in _strings(job.report()) if _absolute(s)]
+        self.assertEqual(leaked, [])
+        self.assertNotIn(str(self.output_dir), json.dumps(job.report()))
+
+
+class TestBatchLifecycle(BatchTestCase):
+    async def test_cancel_after_one_image_keeps_it_and_the_rerun_skips_it(self):
+        for n in range(3):
+            self.image(f"{n}.png", (n, n, n), sees=[f"t{n}"])
+        entered, proceed = threading.Event(), threading.Event()
+        self.tagger.block = (entered, proceed)
+        status, data = await self.start_batch()
+        self.assertEqual(status, 202, data)
+        job = self.api._jobs.get(data["job"]["id"])
+        try:
+            await self.wait_for(entered)
+            cancel = await self.client.post(f"/prompt_manager/jobs/{job.id}/cancel")
+            self.assertEqual(cancel.status, 200)
+        finally:
+            self.tagger.block = None
+            proceed.set()
+        await asyncio.wait_for(asyncio.shield(job.task), WAIT)
+        self.assertEqual(job.state, "cancelled")
+        self.assertEqual(job.status_line, CANCELLED)
+        self.assertEqual(job.counters["tagged"], 1)
+        self.assertEqual(len(self.marked()), 1)
+        self.assertIsNone(model_gate.holder())
+
+        lines = []
+        real_update = type(job).update
+
+        def spy(self_job, **kwargs):
+            if kwargs.get("status_line"):
+                lines.append(kwargs["status_line"])
+            return real_update(self_job, **kwargs)
+
+        with patch.object(type(job), "update", spy):
+            rerun = await self.run_batch()
+        self.assertEqual(rerun.counters["skipped"], 1)
+        self.assertEqual(rerun.counters["tagged"], 2)
+        self.assertIn("Processing file 2 of 3", lines)
+        self.assertIn("Processing file 3 of 3", lines)
+        self.assertNotIn("Processing file 1 of 3", lines)
+        self.assertEqual((rerun.processed, rerun.total), (3, 3))
+
+    async def test_cancel_in_each_phase_writes_nothing(self):
+        self.image("a.png", (1, 1, 1))
+        for phase in (
+            "_safety_backup_step",
+            "_prepare_tagging_model",
+            "_autotag_targets",
+        ):
+            with self.subTest(phase):
+                real = getattr(self.api, phase)
+
+                async def cancelling(*args, _real=real, **kwargs):
+                    result = await _real(*args, **kwargs)
+                    self.api._jobs.active("autotag").cancelled = True
+                    return result
+
+                with patch.object(self.api, phase, side_effect=cancelling):
+                    job = await self.run_batch()
+                self.assertEqual(job.state, "cancelled")
+                self.assertEqual(job.status_line, CANCELLED)
+                self.assertEqual(self.marked(), set())
+                self.assertEqual(self.tagger.calls, [])
+                self.assertIsNone(model_gate.holder())
+
+    async def test_restore_mid_run_fails_the_job(self):
+        self.image("a.png", (1, 1, 1))
+        self.image("b.png", (2, 2, 2))
+
+        def restore(_path):
+            conn = sqlite3.connect(self._temp_db.name)
+            try:
+                rotate_history_epoch(conn)
+            finally:
+                conn.close()
+
+        self.tagger.on_generate = restore
+        job = await self.run_batch()
+        self.assertEqual(job.state, "failed")
+        self.assertEqual(
+            job.status_line, "Prompt storage was restored; start the job again."
+        )
+
+    async def test_listing_error_fails_the_job(self):
+        self.image("a.png", (1, 1, 1))
+        with patch.object(
+            self.api, "_get_gallery_files", side_effect=OSError("/secret/out")
+        ):
+            job = await self.run_batch()
+        self.assertEqual(job.state, "failed")
+        self.assertEqual(
+            job.status_line, "The output folders could not be listed. Try again."
+        )
+        self.assertNotIn("/secret", json.dumps(job.report()))
+
+    async def test_model_that_cannot_load_fails_the_job(self):
+        self.image("a.png", (1, 1, 1))
+        self.tagger.load_error = RuntimeError("no CUDA")
+        job = await self.run_batch()
+        self.assertEqual(job.state, "failed")
+        self.assertEqual(
+            job.status_line, "Auto-tagging failed. Check server logs for details."
+        )
+        self.assertIsNone(model_gate.holder())
+
+    async def test_zero_targets_finish_with_tagged_zero(self):
+        job = await self.run_batch()
+        self.assertEqual(job.state, "finished")
+        self.assertEqual(job.status_line, "Tagged 0 images")
+        self.assertEqual((job.processed, job.total), (0, 0))
+
+    async def test_runner_unloads_after_a_cancel_when_not_kept_in_memory(self):
+        self.image("a.png", (1, 1, 1))
+        owner = model_gate.new_owner("legacy-batch")
+        self.assertTrue(model_gate.try_acquire(owner))
+        run = jobs_module.AutotagRun(
+            model_type="wd14-vit",
+            existing="skip",
+            options={**WD14, "use_gpu": True, "suggestion_limit": 10},
+            keep_in_memory=False,
+            backup_on=False,
+            owner=owner,
+        )
+
+        from py.api.jobs import AUTOTAG_GROUP_NAMES, Job as RealJob
+
+        job = RealJob(
+            "autotag",
+            "Tag output images",
+            admission="records",
+            groups=AUTOTAG_GROUP_NAMES,
+        )
+        real_targets = self.api._autotag_targets
+
+        async def cancel_after_listing():
+            targets = await real_targets()
+            job.cancelled = True
+            return targets
+
+        with patch.object(
+            self.api, "_autotag_targets", side_effect=cancel_after_listing
+        ):
+            await self.api._run_autotag_batch(run, job)
+        self.assertEqual(job.state, "cancelled")
+        self.assertEqual(self.tagger.unload_calls, 1)
+        self.assertIsNone(model_gate.holder())
+
+    async def test_time_left_uses_the_models_rate_only(self):
+        for n in range(4):
+            self.image(f"{n}.png", (n, n, n))
+        await self.run_batch()
+        for n in range(4, 8):
+            self.image(f"{n}.png", (n, n, n))
+        await self.relist()
+        calls = []
+        real = jobs_module._time_left
+
+        def spy(*args):
+            calls.append(args)
+            return real(*args)
+
+        with patch.object(jobs_module, "_time_left", side_effect=spy):
+            job = await self.run_batch()
+        self.assertEqual(job.counters["skipped"], 4)
+        # (done, total, batches, elapsed): over the 4 images the model runs on.
+        self.assertEqual(
+            [(c[0], c[1], c[2]) for c in calls], [(n, 4, n) for n in range(4)]
+        )
+        # The shared D56 text is unchanged.
+        self.assertEqual(admin_module._time_left(5, 10, 5, 300), " · about 5 min left")
+
+
+class TestStartRefusals(BatchTestCase):
+    async def test_unavailable_model_and_bad_bodies(self):
+        self.tagger.downloaded["wd14-vit"] = False
+        status, data = await self.start_batch()
+        self.assertEqual((status, data["code"]), (409, "model-unavailable"))
+        status, data = await self.start_batch(existing="replace_after_review")
+        self.assertEqual((status, data["code"]), (400, "review-required"))
+        status, _data = await self.start_batch(options={**WD14, "suggestion_limit": 3})
+        self.assertEqual(status, 400)
+        self.assertEqual(self.api._jobs.list(), [])
+        self.assertIsNone(model_gate.holder())
+
+    async def test_model_busy_when_the_gate_is_held(self):
+        owner = model_gate.new_owner("image-annotations")
+        self.assertTrue(model_gate.try_acquire(owner))
+        try:
+            status, data = await self.start_batch()
+        finally:
+            model_gate.release(owner)
+        self.assertEqual((status, data["code"]), (409, "model-busy"))
+
+
+class TestBatchBeside(BatchTestCase):
+    async def test_legacy_batch_and_unload_are_refused_while_the_job_runs(self):
+        self.image("a.png", (1, 1, 1))
+        entered, proceed = threading.Event(), threading.Event()
+        self.tagger.block = (entered, proceed)
+        status, data = await self.start_batch()
+        self.assertEqual(status, 202, data)
+        job = self.api._jobs.get(data["job"]["id"])
+        try:
+            await self.wait_for(entered)
+            self.assertTrue(model_gate.holder().startswith("autotag-job:"))
+            legacy = await self.client.post("/prompt_manager/autotag/start")
+            self.assertEqual(legacy.status, 200)
+            frames = [
+                json.loads(line[len("data: ") :])
+                for line in (await legacy.text()).splitlines()
+                if line.startswith("data: ")
+            ]
+            self.assertEqual(
+                frames,
+                [
+                    {
+                        "type": "error",
+                        "status": "error",
+                        "message": "Tag output images already running",
+                    }
+                ],
+            )
+            unload = await self.client.post("/prompt_manager/autotag/unload")
+            self.assertEqual((await unload.json())["operation"], "Tag output images")
+        finally:
+            self.tagger.block = None
+            proceed.set()
+        await asyncio.wait_for(asyncio.shield(job.task), WAIT)
+        self.assertEqual(job.state, "finished")
+
+    async def test_language_model_batch_passes_instructions_per_call(self):
+        self.image("a.png", (1, 1, 1))
+        job = await self.run_batch(model_type="gguf")
+        self.assertEqual(job.state, "finished", job.status_line)
+        self.assertEqual(self.tagger.calls[0]["prompt"], "Tags please.")
+        self.assertEqual(self.tagger.custom_prompt, "legacy prompt")
+
+    async def test_inference_errors_are_fixed_strings_and_not_marked(self):
+        self.image("a.png", (1, 1, 1))
+
+        def explode(_path):
+            raise RuntimeError("/srv/secret/model exploded")
+
+        self.tagger.on_generate = explode
+        job = await self.run_batch()
+        self.assertEqual(
+            self.rows(job, "errors"),
+            [("a.png", "Auto-tagging failed. Check server logs for details.")],
+        )
+        self.assertNotIn("/srv/secret", json.dumps(job.report()))
+        self.assertEqual(self.marked(), set())
+
+
+class TestPreview(BatchTestCase):
+    async def test_preview_counts_images_and_estimates_time(self):
+        prompt_id = self.prompt("p")
+        a = self.image("a.png", (1, 1, 1))
+        self.image("b.png", (2, 2, 2))
+        self.image("c.png", (3, 3, 3))
+        self.link(prompt_id, a)
+        (self.output_dir / "clip.mp4").write_bytes(b"\x00\x00\x00\x18ftypmp42" * 8)
+        status, fresh = await self.preview()
+        self.assertEqual(status, 200, fresh)
+        self.assertEqual(
+            {k: fresh[k] for k in ("images", "already_tagged", "to_tag")},
+            {"images": 3, "already_tagged": 0, "to_tag": 3},
+        )
+        self.assertNotIn("prompts", fresh)
+        self.assertEqual(fresh["suggestion_limit"], 10)
+        self.assertEqual(fresh["estimate_device"], "cpu")
+        self.assertAlmostEqual(fresh["estimate_seconds"], 3 * 1.6)
+
+        self.tagger.block = None
+        await self.run_batch()
+        self.image("d.png", (4, 4, 4))
+        await self.relist()
+        skip = (await self.preview("skip"))[1]
+        self.assertEqual(
+            {k: skip[k] for k in ("images", "already_tagged", "to_tag")},
+            {"images": 4, "already_tagged": 3, "to_tag": 1},
+        )
+        merge = (await self.preview("merge"))[1]
+        self.assertEqual(merge["to_tag"], 4)
+        for model_type in ("gguf", "hf"):
+            llm = (await self.preview(model_type=model_type))[1]
+            self.assertIsNone(llm["estimate_seconds"])
+            self.assertIsNone(llm["estimate_device"])
+        self.assertEqual([s for s in _strings(skip) if _absolute(s)], [])
+
+    async def test_preview_uses_the_graphics_card_rate_when_wd14_runs_on_cuda(self):
+        self.image("a.png", (1, 1, 1))
+        self.tagger.wd14_device = lambda: "gpu"
+        data = (await self.preview())[1]
+        self.assertEqual(data["estimate_device"], "gpu")
+        from py.tagging_defaults import WD14_GPU_SECONDS_PER_IMAGE
+
+        self.assertAlmostEqual(data["estimate_seconds"], WD14_GPU_SECONDS_PER_IMAGE)
+
+    async def test_preview_writes_nothing_and_opens_no_media_file(self):
+        self.image("a.png", (1, 1, 1))
+        await self.run_batch()
+        self.image("b.png", (2, 2, 2))
+        await self.relist()  # the listing, primed
+        files = [self._temp_db.name + suffix for suffix in ("", "-wal")]
+        before = [Path(f).read_bytes() if os.path.exists(f) else None for f in files]
+        media = str(self.output_dir)
+        real_open, real_stat = open, os.stat
+
+        def guarded_open(file, *args, **kwargs):
+            if str(file).startswith(media):
+                raise AssertionError(f"opened {file}")
+            return real_open(file, *args, **kwargs)
+
+        def guarded_stat(path, *args, **kwargs):
+            if str(path).startswith(media) and str(path) != media:
+                raise AssertionError(f"stat {path}")
+            return real_stat(path, *args, **kwargs)
+
+        with patch("builtins.open", guarded_open), patch("os.stat", guarded_stat):
+            status, data = await self.preview()
+        self.assertEqual(status, 200, data)
+        after = [Path(f).read_bytes() if os.path.exists(f) else None for f in files]
+        self.assertEqual(after, before)
+        self.assertEqual(data["to_tag"], 1)
+        self.assertEqual(self.tagger.calls[1:], [])
diff --git a/tests/test_autotag_engine.py b/tests/test_autotag_engine.py
index 53e0bcac..7f7faa3f 100644
--- a/tests/test_autotag_engine.py
+++ b/tests/test_autotag_engine.py
@@ -884,3 +884,29 @@ class TestImageConversion(EngineTestCase):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class TestWd14Device(EngineTestCase):
+    """Where WD14 runs, for the Tag output images estimate (#379)."""
+
+    def test_loaded_session_answers_from_its_first_provider(self):
+        cpu = FakeOnnxSession([0.5] * len(WD14_TAGS))
+        service = self._service(("wd14", (cpu, WD14_TAGS)))
+        _touch_model_files(self.models_dir, "wd14-vit")
+        service.load_model("wd14-vit")
+        self.assertEqual(service.wd14_device(), "cpu")
+        cpu.get_providers = lambda: ["CUDAExecutionProvider", "CPUExecutionProvider"]
+        self.assertEqual(service.wd14_device(), "gpu")
+
+    def test_without_a_session_the_available_providers_answer(self):
+        service = self._service(None)
+        fake = types.SimpleNamespace(
+            get_available_providers=lambda: ["CUDAExecutionProvider"]
+        )
+        with patch.dict(sys.modules, {"onnxruntime": fake}):
+            self.assertEqual(service.wd14_device(), "gpu")
+        fake.get_available_providers = lambda: ["CPUExecutionProvider"]
+        with patch.dict(sys.modules, {"onnxruntime": fake}):
+            self.assertEqual(service.wd14_device(), "cpu")
+        with patch.dict(sys.modules, {"onnxruntime": None}):
+            self.assertEqual(service.wd14_device(), "cpu")
diff --git a/tests/test_autotag_schema.py b/tests/test_autotag_schema.py
new file mode 100644
index 00000000..c7f3c2c3
--- /dev/null
+++ b/tests/test_autotag_schema.py
@@ -0,0 +1,318 @@
+"""Auto-tagged marks schema group (#379): install, validation and restore."""
+
+import hashlib
+import re
+import shutil
+import sqlite3
+import tempfile
+import unittest
+from contextlib import closing
+from pathlib import Path
+
+try:
+    from tests.open_handles import assert_closed
+except ImportError:  # discovered with tests/ as the top-level directory
+    from open_handles import assert_closed
+
+from database.models import PromptModel
+
+try:  # guarded so a missing module fails assertions, not test loading
+    from database.autotag_schema import (
+        AUTOTAG_TABLE,
+        autotag_schema_error,
+        autotag_without_annotations_error,
+        migrate_autotag_state,
+    )
+except ImportError:
+    AUTOTAG_TABLE = None
+    autotag_schema_error = None
+    autotag_without_annotations_error = None
+    migrate_autotag_state = None
+
+TABLE = "media_image_autotags"
+DDL = (
+    f"CREATE TABLE {TABLE} (byte_id TEXT PRIMARY KEY "
+    "REFERENCES media_annotation_bytes(id) ON DELETE CASCADE) WITHOUT ROWID"
+)
+ANNOTATION_TABLES = (
+    "media_annotation_metadata",
+    "media_annotation_files",
+    "media_annotation_bytes",
+    "media_image_tags",
+)
+REVIEW_TABLES = (
+    "media_review_metadata",
+    "media_review_items",
+    "media_review_decisions",
+)
+FILE_ID = "0f8e4b2a-1c3d-4e5f-8a9b-0c1d2e3f4a5b"
+BYTE_ID = "6a1b2c3d-4e5f-4a6b-8c9d-0e1f2a3b4c5d"
+OTHER_ID = "7b2c3d4e-5f6a-4b7c-9d0e-1f2a3b4c5d6e"
+DIGEST = hashlib.sha256(b"img").hexdigest()
+NOW = "2026-10-27T00:00:00.000000+00:00"
+NEEDS_ANNOTATIONS = "Image auto-tag marks need the image annotation tables"
+INVALID_VALUES = "Image auto-tag marks contain invalid values"
+INVALID_SCHEMA = "Image auto-tag marks have an invalid schema"
+
+
+def _compact(sql):
+    return re.sub(r"\s+", "", sql or "").lower()
+
+
+def catalog(conn):
+    return {
+        (kind, name, _compact(sql))
+        for kind, name, sql in conn.execute(
+            "SELECT type,name,sql FROM sqlite_master WHERE name=? OR tbl_name=?",
+            (TABLE, TABLE),
+        )
+    }
+
+
+def marks(conn):
+    return [tuple(row) for row in conn.execute(f"SELECT byte_id FROM {TABLE}")]
+
+
+def insert_byte(conn, file_id=FILE_ID, byte_id=BYTE_ID, locator="/out/img.png"):
+    conn.execute(
+        "INSERT INTO media_annotation_files VALUES(?,?,?)", (file_id, locator, NOW)
+    )
+    conn.execute(
+        "INSERT INTO media_annotation_bytes VALUES(?,?,?,?,?,?,?,0)",
+        (byte_id, file_id, DIGEST, 3, 1_700_000_000_000_000_000, NOW, NOW),
+    )
+
+
+class AutotagSchemaTests(unittest.TestCase):
+    def setUp(self):
+        self.temp = tempfile.TemporaryDirectory()
+        self.addCleanup(self.temp.cleanup)
+        self.addCleanup(assert_closed, self, self.temp.name)
+        self.addCleanup(PromptModel.reset_schema_cache)
+        self.models = []
+        self.addCleanup(self.close_models)
+        self.path = str(Path(self.temp.name) / "live.db")
+
+    def close_models(self):
+        for model in self.models:
+            model.close_all()
+
+    def model(self, path=None):
+        model = PromptModel(path or self.path)
+        self.models.append(model)
+        return model
+
+    def require_module(self):
+        self.assertIsNotNone(
+            autotag_schema_error, "database.autotag_schema is not implemented"
+        )
+
+    def raw(self, path, *statements, foreign_keys=False):
+        with closing(sqlite3.connect(path)) as raw, raw:
+            raw.execute(f"PRAGMA foreign_keys={'ON' if foreign_keys else 'OFF'}")
+            for statement in statements:
+                if isinstance(statement, tuple):
+                    raw.execute(*statement)
+                else:
+                    raw.execute(statement)
+
+    def closed_current(self, name="current.db", *, with_mark=True):
+        """A closed, current database with one byte record (and its mark)."""
+        path = str(Path(self.temp.name) / name)
+        model = self.model(path)
+        with model.get_connection() as conn:
+            insert_byte(conn)
+            if with_mark:
+                conn.execute(f"INSERT INTO {TABLE} VALUES(?)", (BYTE_ID,))
+        model.close_all()
+        return path
+
+    def test_fresh_install_has_exactly_the_pinned_table(self):
+        self.require_module()
+        self.assertEqual(AUTOTAG_TABLE, TABLE)
+        model = self.model()
+        with model.get_connection() as conn:
+            # WITHOUT ROWID with a primary key: no sqlite_autoindex row.
+            self.assertEqual(catalog(conn), {("table", TABLE, _compact(DDL))})
+            self.assertEqual(
+                [tuple(row) for row in conn.execute(f"PRAGMA table_info({TABLE})")],
+                [(0, "byte_id", "TEXT", 1, None, 1)],
+            )
+            self.assertIsNone(autotag_schema_error(conn, allow_missing=False))
+            self.assertEqual(marks(conn), [])
+        self.assertEqual(
+            PromptModel.verify_database_file(self.path, require_library_state=True),
+            (True, "ok"),
+        )
+
+    def test_reopen_and_rerun_are_idempotent(self):
+        self.require_module()
+        path = self.closed_current()
+        for _ in range(2):
+            PromptModel.reset_schema_cache()
+            model = self.model(path)
+            with model.get_connection() as conn:
+                migrate_autotag_state(conn)
+                self.assertEqual(marks(conn), [(BYTE_ID,)])
+                self.assertEqual(catalog(conn), {("table", TABLE, _compact(DDL))})
+            model.close_all()
+
+    def test_old_database_without_the_table_gains_it_and_keeps_its_rows(self):
+        self.require_module()
+        path = self.closed_current(with_mark=False)
+        self.raw(path, f"DROP TABLE {TABLE}")
+        with closing(sqlite3.connect(path)) as raw:
+            before = raw.execute("SELECT * FROM media_annotation_bytes").fetchall()
+            self.assertIsNone(autotag_schema_error(raw))
+            self.assertEqual(
+                autotag_schema_error(raw, allow_missing=False),
+                "Missing image auto-tag marks schema",
+            )
+        self.assertTrue(PromptModel.verify_database_file(path)[0])
+        self.assertEqual(
+            PromptModel.verify_database_file(path, require_library_state=True),
+            (False, "Missing image auto-tag marks schema"),
+        )
+        PromptModel.reset_schema_cache()
+        model = self.model(path)
+        with model.get_connection() as conn:
+            self.assertEqual(catalog(conn), {("table", TABLE, _compact(DDL))})
+            self.assertEqual(
+                [
+                    tuple(row)
+                    for row in conn.execute("SELECT * FROM media_annotation_bytes")
+                ],
+                before,
+            )
+        model.close_all()
+
+    def test_table_without_annotation_group_is_refused(self):
+        self.require_module()
+        path = self.closed_current(with_mark=False)
+        self.raw(
+            path,
+            *(
+                f"DROP TABLE {t}"
+                for t in REVIEW_TABLES + tuple(reversed(ANNOTATION_TABLES))
+            ),
+        )
+        with closing(sqlite3.connect(path)) as raw:
+            self.assertEqual(autotag_without_annotations_error(raw), NEEDS_ANNOTATIONS)
+            for allow_missing in (True, False):
+                self.assertEqual(
+                    autotag_schema_error(raw, allow_missing=allow_missing),
+                    NEEDS_ANNOTATIONS,
+                )
+        self.assertEqual(
+            PromptModel.verify_database_file(path), (False, NEEDS_ANNOTATIONS)
+        )
+        PromptModel.reset_schema_cache()
+        with self.assertRaises(sqlite3.DatabaseError) as raised:
+            self.model(path)
+        self.assertIn(NEEDS_ANNOTATIONS, str(raised.exception))
+
+    def test_invalid_values_are_refused_without_repair(self):
+        self.require_module()
+        cases = {
+            "not-a-uuid": (f"INSERT INTO {TABLE} VALUES('not-a-uuid')",),
+            "upper-uuid": ((f"INSERT INTO {TABLE} VALUES(?)", (OTHER_ID.upper(),)),),
+            "dangling": ((f"INSERT INTO {TABLE} VALUES(?)", (OTHER_ID,)),),
+        }
+        for name, statements in cases.items():
+            with self.subTest(name):
+                path = self.closed_current(f"{name}.db", with_mark=False)
+                self.raw(path, *statements)
+                with closing(sqlite3.connect(path)) as raw:
+                    self.assertEqual(autotag_schema_error(raw), INVALID_VALUES)
+                    before = marks(raw)
+                self.assertEqual(
+                    PromptModel.verify_database_file(path), (False, INVALID_VALUES)
+                )
+                PromptModel.reset_schema_cache()
+                with self.assertRaises(sqlite3.DatabaseError):
+                    self.model(path)
+                with closing(sqlite3.connect(path)) as raw:
+                    self.assertEqual(marks(raw), before)
+
+    def test_altered_definitions_are_refused(self):
+        self.require_module()
+        cases = {
+            "rowid": f"CREATE TABLE {TABLE} (byte_id TEXT PRIMARY KEY "
+            "REFERENCES media_annotation_bytes(id) ON DELETE CASCADE)",
+            "restrict": f"CREATE TABLE {TABLE} (byte_id TEXT PRIMARY KEY "
+            "REFERENCES media_annotation_bytes(id) ON DELETE RESTRICT) WITHOUT ROWID",
+            "extra column": f"CREATE TABLE {TABLE} (byte_id TEXT PRIMARY KEY "
+            "REFERENCES media_annotation_bytes(id) ON DELETE CASCADE, engine TEXT) "
+            "WITHOUT ROWID",
+        }
+        for name, ddl in cases.items():
+            with self.subTest(name):
+                path = self.closed_current(f"{name}.db", with_mark=False)
+                self.raw(path, f"DROP TABLE {TABLE}", ddl)
+                with closing(sqlite3.connect(path)) as raw:
+                    self.assertEqual(autotag_schema_error(raw), INVALID_SCHEMA)
+                self.assertFalse(PromptModel.verify_database_file(path)[0])
+        path = self.closed_current("index.db", with_mark=False)
+        self.raw(path, f"CREATE INDEX idx_extra ON {TABLE}(byte_id)")
+        with closing(sqlite3.connect(path)) as raw:
+            self.assertEqual(autotag_schema_error(raw), INVALID_SCHEMA)
+
+    def test_deleting_the_byte_record_cascades_to_its_mark(self):
+        self.require_module()
+        model = self.model()
+        with model.get_connection() as conn:
+            insert_byte(conn)
+            conn.execute(f"INSERT INTO {TABLE} VALUES(?)", (BYTE_ID,))
+            conn.execute(f"INSERT OR IGNORE INTO {TABLE} VALUES(?)", (BYTE_ID,))
+            self.assertEqual(marks(conn), [(BYTE_ID,)])
+            conn.execute("DELETE FROM media_annotation_files")
+            self.assertEqual(marks(conn), [])
+
+    def test_restore_of_a_pre_change_backup_migrates(self):
+        self.require_module()
+        src = self.closed_current("backup.db", with_mark=False)
+        self.raw(src, f"DROP TABLE {TABLE}")
+        live = self.model()
+        live.restore_from_file(src)
+        with closing(sqlite3.connect(self.path)) as raw:
+            self.assertEqual(catalog(raw), {("table", TABLE, _compact(DDL))})
+            self.assertEqual(
+                raw.execute("SELECT id FROM media_annotation_bytes").fetchall(),
+                [(BYTE_ID,)],
+            )
+        self.assertEqual(
+            PromptModel.verify_database_file(self.path, require_library_state=True),
+            (True, "ok"),
+        )
+        with closing(sqlite3.connect(src)) as raw:  # the source is untouched
+            self.assertEqual(catalog(raw), set())
+
+    def test_install_failure_rolls_back(self):
+        self.require_module()
+        src = self.closed_current("copy.db", with_mark=False)
+        self.raw(src, f"DROP TABLE {TABLE}")
+        path = str(Path(self.temp.name) / "rollback.db")
+        shutil.copyfile(src, path)
+        with closing(sqlite3.connect(path)) as raw:
+            raw.execute("BEGIN")
+            raw.execute("CREATE TABLE marker(x)")
+            # An invalid group already present: nothing created, outer work kept.
+            raw.execute(f"CREATE TABLE {TABLE}(byte_id TEXT)")
+            with self.assertRaises(sqlite3.DatabaseError):
+                migrate_autotag_state(raw)
+            self.assertEqual(
+                raw.execute(
+                    f"SELECT sql FROM sqlite_master WHERE name='{TABLE}'"
+                ).fetchone()[0],
+                f"CREATE TABLE {TABLE}(byte_id TEXT)",
+            )
+            self.assertIsNotNone(
+                raw.execute(
+                    "SELECT 1 FROM sqlite_master WHERE name='marker'"
+                ).fetchone()
+            )
+            raw.rollback()
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/tests/test_database_backup.py b/tests/test_database_backup.py
index e4aa9a88..2437220b 100644
--- a/tests/test_database_backup.py
+++ b/tests/test_database_backup.py
@@ -468,6 +468,7 @@ class TestRestoreFromFile(BackupTestCase):
 
     def test_restore_old_backup_without_annotations(self):
         tables = (
+            "media_image_autotags",
             "media_review_decisions",
             "media_review_items",
             "media_review_metadata",
diff --git a/tests/test_review_schema.py b/tests/test_review_schema.py
index 1ee935e0..b97780ac 100644
--- a/tests/test_review_schema.py
+++ b/tests/test_review_schema.py
@@ -371,7 +371,10 @@ class ReviewSchemaTests(unittest.TestCase):
         self.require_module()
         for label, dropped in (
             ("annotations only", TABLES),
-            ("predates annotations", ANNOTATION_TABLES + TABLES),
+            (
+                "predates annotations",
+                ("media_image_autotags",) + ANNOTATION_TABLES + TABLES,
+            ),
         ):
             with self.subTest(label):
                 path = str(Path(self.temp.name) / f"{label.replace(' ', '-')}.db")
diff --git a/web/js/workspace/system-tagging.mjs b/web/js/workspace/system-tagging.mjs
index aaea5422..d9b369fa 100644
--- a/web/js/workspace/system-tagging.mjs
+++ b/web/js/workspace/system-tagging.mjs
@@ -15,10 +15,11 @@
  *   value (the toast names it). D132: the concept's two simulated job buttons are "Download" (or
  *   "Download again" for a model on disk), which starts the `model_download` Activity job and opens it, and "Unload",
  *   a direct `POST /autotag/unload` (enabled while a model is loaded; a busy model's refusal names the operation).
- * - D134: the `.operation` "Tag output images" › "Review" previews the prompt-tag batch
- *   (`POST /autotag/batch/preview`, read-only) and shows it in the shared review dialog with the D55 backup box; "Start
- *   tagging" runs the cancellable `autotag` Activity job. A review that cannot start (Replace after review, a short
- *   disk, nothing to tag) offers only Cancel and says why.
+ * - D134: the `.operation` "Tag output images" › "Review" previews the per-image batch (#379,
+ *   `POST /autotag/batch/preview`, read-only: image counts and a time estimate) and shows it in the shared review dialog
+ *   with the D55 backup box; "Start tagging" runs the cancellable `autotag` Activity job, which tags each output image
+ *   from its own pixels and never a prompt. A review that cannot start (Replace after review, a short disk, no image,
+ *   nothing to tag) offers only Cancel and says why.
  *
  * Every value is a host draft saved by the one Save row (system.mjs); every listener sits on the section's scope.
  */
@@ -44,7 +45,7 @@ export const REVIEW_TITLE = 'Review tag output images';
 export const BATCH_TITLE = 'Tag output images';
 export const START_LABEL = 'Start tagging';
 export const INTRO = 'One tagging configuration for images and prompt-library actions. Every result can be reviewed before applying.';
-export const BATCH_DESCRIPTION = 'Add tags to each prompt from its linked output images with the model above. Review the scope first.';
+export const BATCH_DESCRIPTION = 'Tag each output image from its own pixels with the model above. Prompts are not changed. Review the scope first.';
 export const SUGGESTIONS_DESCRIPTION = 'From 1 to 100: the most suggestions Generate shows for one image.';
 
 /** D132: the engines in the concept's order, with the labels and sizes the server states when it cannot be read. */
@@ -97,8 +98,12 @@ export const SETTING_PROBLEMS = Object.freeze({
 export const characters = text => [...String(text ?? '')].length;
 
 export const REPLACE_NOTE = 'Replacing tags needs a review of each image. Use Preview tagging review, or choose Skip or Merge for a batch run.';
-export const NOTHING_TO_TAG = 'Every prompt with an output image is already tagged.';
-export const NO_LINKED_IMAGES = 'No output image is linked to a prompt yet, so there is nothing to tag.';
+export const NOTHING_TO_TAG = 'Every output image is already auto-tagged.';
+export const NO_IMAGES = 'The output folders contain no images to tag.';
+/** The Time row's tails (#379): where WD14 runs, and the language models, which have no estimate. */
+export const TIME_ON_CPU = ' on a processor; a graphics card can be faster.';
+export const TIME_ON_GPU = ' on the graphics card.';
+export const TIME_UNKNOWN = 'Language-model engines take longer per image. Activity shows the time left once tagging starts.';
 export const STATUS_UNREAD = 'The model status is still being read. Try again in a moment.';
 export const SETTINGS_UNREAD = 'The settings are still being read. Try again in a moment.';
 export const UNLOAD_FAILED = 'The tagging model could not be unloaded. Try again.';
@@ -254,19 +259,40 @@ export async function startJob({ activity, scope, kind, body, what, notify = sho
 /** The existing-tags policy's label. */
 export const existingLabel = value => EXISTING_OPTIONS.find(([id]) => id === value)?.[1] ?? String(value ?? '');
 
-/** D134: the review's rows from the server's preview and the drafts it was made with. */
+/**
+ * DEC11 (#379): a whole number of minutes as the review's Time text: "about m min" under 90, then "about H h M min",
+ * or "about H h" on the hour.
+ */
+export function durationText(minutes) {
+    const m = Math.max(1, Math.ceil(Number.isFinite(minutes) ? minutes : 1));
+    if (m < 90) return `about ${m} min`;
+    const hours = Math.floor(m / 60);
+    const rest = m % 60;
+    return rest === 0 ? `about ${hours} h` : `about ${hours} h ${rest} min`;
+}
+
+/** The review's Time row: the preview's `estimate_seconds` where WD14 runs (`estimate_device`), else the sentence. */
+export function timeText(plan) {
+    const seconds = plan?.estimate_seconds;
+    if (typeof seconds !== 'number' || !Number.isFinite(seconds) || seconds < 0) return TIME_UNKNOWN;
+    return durationText(Math.ceil(seconds / 60)) + (plan.estimate_device === 'gpu' ? TIME_ON_GPU : TIME_ON_CPU);
+}
+
+/** D134: the review's rows from the server's preview and the drafts it was made with (image counts, #379). */
 export function batchGrid(plan, values) {
     const id = plan.model?.type ?? values.autotag_model_type;
-    const settings = CLASSIFIERS.has(id)
+    let settings = CLASSIFIERS.has(id)
         ? `General ${values.autotag_general_threshold} · Character ${values.autotag_character_threshold}`
         : `Template of ${number(characters(values.autotag_template))} characters`;
+    // The saved count the run uses (the preview's), not an unsaved draft.
+    if (Number.isSafeInteger(plan.suggestion_limit)) settings += ` · up to ${number(plan.suggestion_limit)} tags per image`;
     return [
         ['Model', `${modelLabel(id, plan.model)} · ${modelSize(id, plan.model)}`],
         ['Existing tags', existingLabel(plan.existing ?? values.autotag_existing)],
-        ['Images linked', number(plan.images)],
-        ['Prompts', number(plan.prompts)],
-        ['Already tagged', number(plan.already_tagged)],
-        ['Prompts to tag', `up to ${number(plan.to_tag)}`],
+        ['Images', number(plan.images)],
+        ['Already auto-tagged', number(plan.already_tagged)],
+        ['Images to tag', `about ${number(plan.to_tag)}`],
+        ['Time', timeText(plan)],
         ['Settings', settings],
         ['Backup', backupText(plan.backup)],
     ];
@@ -291,13 +317,13 @@ export function batchCloseNote(plan) {
     if (plan.review_required === true) return REPLACE_NOTE;
     if (plan.model?.downloaded === false) return notDownloaded(modelLabel(plan.model.type, plan.model));
     if (plan.disk?.sufficient === false) return diskShortNote(plan.disk);
-    if (!Number.isSafeInteger(plan.prompts) || plan.prompts === 0) return NO_LINKED_IMAGES;
+    if (!Number.isSafeInteger(plan.images) || plan.images === 0) return NO_IMAGES;
     if (plan.to_tag === 0) return NOTHING_TO_TAG;
     return null;
 }
 
-export const BATCH_SENTENCE = 'Nothing is written until you start. Missing files and files outside the output folders are '
-    + 'counted, so the prompts to tag are an upper bound; Activity reports what was tagged.';
+export const BATCH_SENTENCE = 'Nothing is written until you start. The number to tag is an estimate: images changed since '
+    + 'they were tagged are tagged again, and missing files are skipped. Activity reports what was tagged.';
 
 /** The Cancel-only review: the counts and why the batch cannot start (CUJ 5), in the shared close-only shell. */
 export function openBatchCloseOnly({ scope, trigger, grid, note }) {