DeviantArt Downloader

Download the full gallery of any DeviantArt profile.

Works are fetched through two routes, so the API quota is spent only on what the API alone can serve:

Each route saves to its own subfolder inside the gallery folder. --force-api restores the old behaviour of routing everything through the official API.

• Downloads the original file when the author allows it, or the highest publicly available resolution image.
• Downloads literature and journals too: text works have no media file, so their full body is saved next to the images as plain text (.txt) or a standalone HTML document (.html), your choice with --literature-format. The body is fetched from the website for no API quota, falling back to the listing excerpt when it is unavailable. Restrict a run with --only, which takes images, literature, mature, ai, no-ai, upscaled and no-upscaled, repeated or comma-separated, and set either flag per user in a _users.json that survives a rename (see below).
• Downloads mature content unblurred when you log in with your account (--login, see below). Without login, --unblur/DA_UNBLUR=true strips the blur where possible: works uploaded since ~mid-2021 have their URL token pinned to the blurred version, so for those the blurred preview is downloaded instead.
• Parallel downloads with retries and API rate-limit handling: every worker draws from one shared budget (DA_API_RATE, 3 requests/second by default), and a 429 holds the whole pool back instead of each thread backing off on its own. The website route needs no OAuth call at all, so a re-sync of an all-ages gallery costs zero API requests.
• Detects duplicates across runs (even if the artwork's title has changed), so it is safe to re-run to sync new works.
• Run it with no arguments to re-sync every user already present in the output folder with their latest works, or with --watching to download every user your account watches.
• Re-syncs are incremental: the gallery listing stops as soon as it reaches a page of already-downloaded works (--full forces a complete walk).
• Files you delete manually stay deleted: the download record (_downloaded.json) is authoritative, so deleted works are not downloaded again unless you pass --redownload-missing.
• Saves the full metadata of every work to _metadata.json, including what the website knows about AI involvement: is_ai_generated (the flag behind the site's own "Suppress AI" filter, true for DreamUp works too), is_upscaled and is_ai_use_disallowed. Only the website listing carries these, so a work listed through the API records them as null — not known, rather than not AI.
• Ends every run with a summary broken down by route and size (items and MB downloaded via the website vs. the API), plus a per-user breakdown when syncing several profiles.

Installation

pip install deviantart-gallery-downloader

Credentials

1. Create a DeviantArt account.
2. Register an application (confidential type) at https://www.deviantart.com/developers/register.
3. Copy the client_id and client_secret from https://www.deviantart.com/developers/apps.

Export them as environment variables or put them in a .env file in the directory you run the command from:

DA_CLIENT_ID=your_client_id
DA_CLIENT_SECRET=your_client_secret
# Optional: simultaneous website downloads (default: 4, recommended not to exceed 8)
DA_WEB_WORKERS=4
# Optional: simultaneous API downloads (default: 2); kept low so parallel API
# requests don't trip the rate limit
DA_API_WORKERS=2
# Optional: API requests per second, shared by every worker (default: 3);
# 0 disables the pacing
DA_API_RATE=3
# Optional: report only results, dropping the per-work progress lines
DA_QUIET=false
# Optional: strip the blur filter the API applies to mature-content previews
# (default: false, images are kept as the API serves them)
DA_UNBLUR=false
# Optional: file format for literature and journals — "txt" (plain text) or
# "html" (a standalone document that keeps the formatting) (default: txt)
DA_LITERATURE_FORMAT=txt
# Optional: keep only the works matching all of images, literature, mature,
# ai, no-ai, upscaled, no-upscaled (default: unset, keeps everything)
DA_ONLY=
# Optional: per-user --only / --literature-format file (default: _users.json in
# the output folder, if it exists)
DA_USER_CONFIG=
# Optional: output folder, absolute or relative ("~" is expanded)
DA_OUTPUT=~/Pictures/deviantart
# Optional: route every work through the API instead of the website listing
DA_FORCE_API=false

Usage

deviantart-downloader https://www.deviantart.com/username
deviantart-downloader username

# Passing the credentials as arguments:
deviantart-downloader username --client-id XXX --client-secret YYY

# Useful options:
deviantart-downloader username --info         # show profile info + galleries, download nothing
deviantart-downloader username -g "Sketches"  # only the named gallery folder (case-insensitive)
deviantart-downloader username -o my_folder   # output folder (default: DA_OUTPUT or downloads)
deviantart-downloader username -w 8           # simultaneous website downloads
deviantart-downloader username --api-workers 3  # simultaneous API downloads (default: 2)
deviantart-downloader username --api-rate 2   # API requests per second, all workers (default: 3)
deviantart-downloader username --redownload-missing  # restore manually deleted files
deviantart-downloader username --redownload-blurred  # replace copies saved blurred before --login
deviantart-downloader username --unblur       # strip the blur on mature-content previews
deviantart-downloader username --literature-format html  # save literature/journals as .html (default: txt)
deviantart-downloader username --only images       # download only images (skip literature/journals)
deviantart-downloader username --only mature      # download only the mature works
deviantart-downloader username --only literature mature  # ...and only the mature literature
deviantart-downloader username -q             # only results: no per-work progress lines
deviantart-downloader username --full         # walk the entire gallery listing
deviantart-downloader username --force-api    # route everything through the API
deviantart-downloader --watching              # download every user you watch (needs --login)
deviantart-downloader --watching --info       # just summarise them, download nothing

Try it on a demo profile

test is a small, long-standing public profile that makes a good first run: one gallery, 18 all-ages works (~4.5 MB), nothing mature, so no --login is needed.

deviantart-downloader test --info     # inspect it first: profile + gallery counts
deviantart-downloader test -o demo     # download all 18 works into ./demo/test/

--only narrows a run to the works you actually want. Its selectors sit on four axes and combine the way filters usually do — a union within an axis, an intersection across them:

mature reads the flag the listing carries. Note that a handful of works are served blurred without carrying it, and those are not selected by it.

The last two axes are separate declarations, not degrees of one: ai / no-ai read what the website's "Suppress AI" setting filters on (DreamUp works included), while upscaled / no-upscaled read whether the author ran the work through an AI upscaler. A hand-drawn piece upscaled that way is not AI-made, so --only no-ai keeps it and only --only no-upscaled turns it away.

Only the website listing carries either one, which is what makes each pair lopsided rather than mirrored: the plain value keeps the works known to be that way, while the no- one keeps everything not known to be, rather than dropping a work over a fact the listing never reported. When the API ends up doing the listing (--force-api, or the website route being unavailable) nothing at all is known, and the run says so, one line per axis, instead of quietly selecting everything or nothing:

WARNING: the API listing does not report whether a work was upscaled with AI, so --only upscaled has nothing to select on and matches no work here.

Different settings per user

--only and --literature-format apply to the whole run, which is the wrong shape for a folder full of galleries that want different things: one artist worth only their images, another followed for the literature. A _users.json in the output folder gives a user their own answer:

{
  "someartist": { "only": "images, no-ai" },
  "aWriter":    { "only": "literature", "literature-format": "html" },
  "everything": { "only": "" }
}

Selectors read exactly as they do on the command line — one string or a list, spaces or commas — and an empty only means everything for that user. Usernames are matched whatever their case, and literature_format is accepted alongside literature-format.

A flag you type outranks the file, which was written before this run. Highest wins:

So --only images with the file above downloads only images from aWriter too, and says which setting the file lost:

--only given on the command line, which settles it for every user: _users.json does not get a say in that.

One flag settles one setting: pass --only images and the file still picks each user's literature-format. The .env variables sit below the file rather than above it, because they are a standing default rather than a decision about this run — DA_ONLY=images with the file above still gets aWriter's literature.

The file is read before the first user is synced, so a typo in it stops the run there instead of halfway through a batch — including in a setting this run overrides, which is wrong today and would bite the first run that leaves the flag off:

_users.json: "someartist" asks --only for sfw, which is not among images, literature, mature, ai, no-ai, upscaled, no-upscaled.

Point somewhere else with --user-config PATH (or DA_USER_CONFIG); a file named that way and not found is an error, while the default location simply has nothing to say when there is no file in it.

A rename does not invalidate it. DeviantArt lets people change their username, and an entry filed under a name nobody answers to any more would quietly stop applying. So each entry records the id the route reports for that user — the one thing a rename does not change — and a run that meets that id under a new name moves the entry across:

_users.json: "oldname" is now "newname"; moving those settings over.

The ids are written by the tool, not by hand, and appear the first time a user is synced. Note the two routes disagree about user ids (the website reports a numeric one, the API a UUID, and for the same user they differ), so each is kept under the name of the route that reported it and neither is ever read as the other:

{ "newname": { "only": "literature", "ids": { "web": "233267" } } }

The gallery folder is still named after the username, so a renamed user's older files stay in the folder under the old name; only the settings follow the rename.

Pass -q/--quiet (or DA_QUIET=true) when the progress is more noise than signal — a long sync prints one line per work, and --watching multiplies that by every user you follow. It drops the lines for works that succeeded; summaries, the works that failed, warnings, errors and the --watching confirmation still print, so a quiet run still tells you what happened and what went wrong. On the demo profile below that is 34 lines of output against 8. It layers on anything else: --watching -q, --info -q, --only images -q.

Progress does not disappear, it moves: in a terminal it goes to the status line pinned at the bottom, next to the keyboard controls, so a quiet run still shows what it is working on.

[running]  12/900  api  Crystal ID  keys: [p] pause  [r] resume  [q] quit

It names the route each work took, since the two behave nothing alike: web costs no quota and runs at -w workers, api is metered and paced by --api-rate. The line is trimmed to the terminal width, and a rate-limit wait adds a second line above it that counts down:

[rate limit]  resuming in 28s                                    429s so far: 3
[running]  12/900  api  Crystal ID                keys: [p] pause  [r] resume  [q] quit

That replaces the line each 429 used to print — several workers, several attempts, per user — with one place that also shows the run is waiting rather than hung. Piped or redirected there is no status line, so a stall announces itself once instead of counting down, and -q simply prints less.

Every run ends with a summary broken down by route, size and (when syncing several users) per user:

Done. Downloaded: 18 | Skipped (already existed): 0 | No file: 0 | Failed: 0
  · via website: 18 item(s), 4.5 MB
  · via API:     0 item(s), 0 B
  · Total downloaded: 4.5 MB in 3.7s (1.2 MB/s), avg 257.5 KB/file
Files saved to: /.../demo/test

Pass -i/--info to print a profile summary — profile URL, avatar and banner URLs, bio, location, birthday, "deviant for X years", links, statistics and every gallery folder with its item count — and exit without downloading anything. Combined with --watching it summarises every user you watch instead of one profile. DeviantArt does not expose pronouns through its endpoints, so those are not shown.

The whole summary comes off the website, so --info costs no API quota at all — which is what makes --watching --info viable over a long watchlist. The website does not publish the real name or a readable artist specialty; add --force-api to fetch the profile through the API and get those two back, at one request per user. (The bio is not a reason to: the API's bio field comes back empty on current profiles, while the website carries the full text.)

Pass -g/--gallery "NAME" to download only one gallery folder instead of the whole gallery (the name is matched case-insensitively; if it doesn't exist the tool lists the folders that do). Files land in the same <output>/<username>/ folder as a full sync, so works are never downloaded twice across runs.

Files are saved to <output>/<username>/web/ or <output>/<username>/api/, depending on the route each work took. Literature and journals land in the same subfolders as a .txt or .html file (see --literature-format). The download record and the metadata live in <output>/<username>/, shared by both routes: a work is never downloaded twice, whichever route lists it.

Galleries downloaded by earlier versions keep their existing flat layout; those files are recognised and left where they are, and only new works land in the route subfolders.

While it is fetching the listing or downloading, and when run in a terminal, you can steer it from the keyboard: p pauses, r resumes, and q quits (like Ctrl+C: it stops and cleans up, and re-running resumes where it left off). A status line pinned to the bottom of the terminal shows the available keys and the current state, and the output scrolls above it. When the output is piped or redirected, these controls are simply inactive.

Replacing copies you saved blurred (--redownload-blurred)

Mature works downloaded before you ran --login are the blurred placeholder the API serves an anonymous visitor. --redownload-blurred fetches those again now that your account can see the real image:

deviantart-downloader --login              # once, if you have not already
deviantart-downloader username --redownload-blurred

Nothing on disk records whether a given file is the blurred one, so the tool works it out from what is on offer now and what you already have:

The summary counts the replacements apart from ordinary downloads, so what the pass achieved is visible rather than buried among works that were simply new:

Done. Downloaded: 1 | Skipped (already existed): 137 | No file: 0 | Failed: 0 | Replaced (were blurred): 42

So it is safe to re-run: the second pass replaces nothing. It requires --login and says so rather than walking every listing to conclude nothing, since without a session the API only ever offers the blur.

It is a repair pass, not a sync. Because it can only ever replace files you already have, a user whose download record holds nothing from the API route has no blur to replace — and is skipped whole, before a single request. That is read off _downloaded.json locally, so it costs nothing:

deviantart-downloader --watching --redownload-blurred

On a 148-user watchlist where 13 had ever downloaded mature content, that walks 13 galleries instead of 148. The other 135 are dismissed in milliseconds. Galleries downloaded by versions old enough to predate the web/ and api/ subfolders are always walked, since their record does not say which route a work took and assuming all-ages would skip works that do need repair.

Note that being a repair pass cuts both ways: a skipped user's new works are not picked up either. Run it without the flag for that.

When re-syncing a user, the gallery listing (newest first) stops at the first page whose works were all downloaded before, so frequent re-runs stay cheap even on huge galleries. Pass --full occasionally to walk the whole listing and pick up older works that became visible later (for example mature content after --login); --redownload-missing implies it.

Sync every downloaded user

With no profile argument, the tool scans the output folder for the users you already downloaded (their subdirectories) and fetches whatever they published since:

deviantart-downloader                # sync everyone under DA_OUTPUT (or ./downloads)
deviantart-downloader -o my_folder   # sync everyone under my_folder

Only subdirectories created by a previous run are considered (they are recognised by the _downloaded.json / _metadata.json files inside), so unrelated folders in the output directory are ignored. Users whose gallery comes back empty (deactivated accounts) are skipped with a notice instead of aborting the run.

Sync everyone you watch (--watching)

--watching reads the watchlist of the account you logged in with and downloads all of those galleries into the output folder, one subfolder per user:

deviantart-downloader --login        # once, to save the session
deviantart-downloader --watching     # then, whenever you want to catch up

Since a watchlist can be long, it tells you how many users it found and asks before starting:

Fetching the users your account watches...
  Page at offset 0: 50 watched user(s) (total: 50)
  Page at offset 50: 48 watched user(s) (total: 98)

You watch 98 user(s).
Download all 98 galleries into downloads? [y/N]

Anything but y/yes cancels. When stdin is piped or redirected the question is skipped and the run goes ahead, so --watching still works unattended from a cron job or a script — which also means a redirected stdin starts the whole download with no confirmation.

It needs that saved session: without it the token belongs to the application, which watches nobody. The list is fetched fresh on each run, so people you started or stopped watching are picked up automatically.

Most options apply per user as usual, so --watching --only images or --watching --full do what you would expect:

Two of those are worth thinking about before you use them on a long watchlist. --redownload-missing and --redownload-blurred both imply --full, so they walk every listing of every watched user from end to end instead of stopping at the first page already downloaded — unavoidably, since the works they look for are exactly the ones an incremental sync skips. --watching --redownload-blurred is the likeliest way you will want the latter, though: it replaces every blurred copy across everyone you follow in one pass.

--unblur is close to pointless here: --watching already requires the logged-in session, and with it mature works arrive unblurred anyway (--unblur exists for runs without --login).

Re-runs are incremental like any other sync, and a watched user whose account has since been deactivated is skipped with a notice. This is not a setting you want in .env: it is an action, and every run downloads every gallery you watch.

Staying under the API rate limit

DeviantArt answers an overrun with user_api_threshold. That limit is per account, not per endpoint, and it reacts to short bursts rather than to a running total — so spreading the work over different endpoints buys nothing, and the only thing that helps is not going too fast.

The cheapest request is the one never made, so two of them are avoided outright:

• The original file is often what you already have. The API charges a request per work to hand out the original, and for most works it hands back exactly what content.src was already serving — the fullview is only re-encoded when the original is too large for it. The listing already says how many bytes the original has (download_filesize) and the CDN says how many the fullview has, so the two are compared first, over a head request that costs no quota. Measured over 23 mature downloadable works: 17 needed no request at all, 73%. Every uncertainty — an unknown size, a CDN that will not say, a blur — spends the request instead, because being wrong the other way would save the fullview as if it were the original.
• An answer already paid for is not bought twice. Reaching a mature work means finding the UUID the API is keyed by, which costs a page of the API's own listing. Those answers are kept in _resolved.json beside the download record, and the CDN URLs they hold carry no expiry, so a repair pass or a retried failure spends nothing on works it has already resolved. A cached answer that fails to download is dropped rather than retried forever, so the next run buys a fresh one. One exception keeps --redownload-blurred honest: an answer cached by a logged-out run holds the blurred placeholder, and once --login could do better it is looked up again instead of reused.

Then two mechanisms keep what is left under the limit:

• One shared budget. Every API request, whichever endpoint it is for and whichever worker makes it, is paced through a single limiter at DA_API_RATE requests per second (3 by default, --api-rate to override, 0 to disable). Workers queue for the next slot instead of racing each other, so the request rate stays flat no matter how you set --api-workers.
• One shared cool-down. When a 429 does arrive, the first worker to see it backs off the whole pool, and the others wait that out rather than each starting their own ladder. Since a 429 means the account is going too fast, that is true of every worker at once; letting the rest keep firing only earns more 429s. The wait doubles from 4 s up to a 5-minute ceiling and resets after any request gets through.

Note that DeviantArt sends no Retry-After header and no rate-limit headers of any kind, so the wait is a blind ladder. The header is honoured if it ever appears.

The default rate was picked by measuring it. Thirty back-to-back API calls, run twice against the same account:

Unpaced, a third of the requests were rejected and had to be retried, so 30 calls cost 45 requests and most of the time went into backoff waits. Paced, the same work cost exactly 30 requests and never tripped the limit. Going slower on purpose finished 17× faster.

Your mileage will differ: the threshold appears to run over a long window, so a run that starts with quota already spent — by an earlier run, or by the same account downloading somewhere else — will see 429s at any rate. If that happens, lower DA_API_RATE; if you never see one, raise it.

The website route is not affected by any of this: it needs no OAuth and costs no quota, which is why a re-sync of an all-ages gallery makes zero API requests.

Unblurred mature content (--login)

Mature works are the reason the API route exists: the website lists them to logged-out visitors as a blurred placeholder whose URL token is pinned to that blurred transformation, so scraping cannot recover them. Without a logged-in user, the API serves mature works as an anonymous visitor would see them: blurred, and with the image URL cryptographically pinned to the blurred version for works uploaded since ~mid-2021 (--unblur cannot help there). To get the real images, log in with your DeviantArt account:

1. In your account settings, enable mature content.
2. In https://www.deviantart.com/developers/apps, edit your application and add http://127.0.0.1:8721/callback to the OAuth2 Redirect URI Whitelist.
3. Run:

deviantart-downloader --login            # one-time browser authorization
deviantart-downloader username           # subsequent runs use the saved session

The browser opens once to authorize the app; the session is stored in ~/.config/deviantart-downloader/token.json and renewed automatically. If it ever expires (about 3 months without use), run --login again. While a session is saved, mature works are downloaded unblurred and --unblur is unnecessary.

Development

git clone https://github.com/Nibblex/DeviantartDownloader
cd DeviantartDownloader
python -m venv .venv && source .venv/bin/activate
pip install -e .[dev]
pytest   # runs the test suite with a coverage report

The package is layered bottom-up, each module depending only on the ones above it:

The test suite mirrors that layout (tests/test_<module>.py), with the fakes and factories in tests/conftest.py.

License

GNU General Public License v3.0 or later — see LICENSE.

Releases up to and including 1.6.1 were published under the MIT license and stay available under those terms; 1.7.0 onwards are GPLv3+.