日本語READMEはこちら

34,251 images · about 2.07 GB / 1.92 GiB · Release notes · SHA256 checksums

v0.1.2 package refresh — 2026-10-24: The catalog title now opens this repository in a new tab. The ZIP also includes updated English and Japanese READMEs. The 34,158 images and rendering settings are unchanged. The download filename stays the same; use the updated SHA256SUMS.txt on the Release page to verify this refreshed package. Update details.

New — 2026-10-25: eight style categories. All 221 styles are assigned once. Open Styles → Category in the left filter panel; combine categories with search and switch English/Japanese. Thanks to Brie-Wensleydale for Issue #2. The refreshed v0.1.2 ZIP includes the feature; verify it with the current Release checksum. Categories and counts · 314-entry CSV.

Release ZIP refresh (2026-10-27): Added 93 styles: 33,937 characters + 314 styles = 34,251 images. All 314 styles have exactly one of eight categories. The ZIP was selectively updated with 7-Zip; the previous ZIP/checksum are retained as before-style93. Images remain uncurated. The live demo still shows 120 samples. Rendering settings and verification · 314-style category list.

A searchable library of MOSS-Image-generated images and reusable prompts, built from an Anima-oriented character list plus additional characters and picture styles. Browse a subject, inspect its result, and copy a prompt into your own workflow. The interface follows the Krea2 catalog's familiar operations, with English and Japanese available.

Release status — 2026-10-27: v0.1.2 — Download the complete ZIP includes the local application, 33,937 character images + 314 styles, comparison tools and both sample workflow JSON files. The online demo shows only 120 samples. GitHub Code → Download ZIP contains the demo and documentation; use the Release ZIP for the full collection. Model weights and reference libraries are separate downloads.

The collection has not been selected by pass/fail judgments. Incorrect character identities, mismatched prompts and weak style results remain included. Neither inclusion nor resemblance proves training-set coverage or accurate reproduction. The demo is an interface sample, not a list of approved results.

Browse characters · Browse styles

No Python, GPU or installation is needed for the demo. Switch tabs, search names, filter by work or appearance, change the order, click an image to enlarge it, and copy its prompt or original tags. Theme, language and display preferences are remembered in your browser.

The demo uses the same static-view approach as Krea2. Only the sample site is deployed to GitHub Pages. The full collection exceeds Pages' 1 GB site limit. Demo scope and deployment.

This is a character-list-based catalog with additional samples, not an AnimaDex artist list. The source anima_characters.csv has 36,492 rows. Applying user limits removed 2,761 planned entries, and another 41 ungenerated entries were cancelled, leaving 33,690 generated source entries.

Outfits, forms and tests count as separate images, not unique identities. Krea2 images and verdicts are not imported. The 314 styles include 40 earlier samples, 181 subsequent samples and 93 additions generated on 2026-10-26. Their prompts describe visual characteristics in English; painter names, media and animation styles are experiments, not accuracy guarantees.

The complete Release ZIP uses standard DEFLATE compression and is about 2.07 GB / 1.92 GiB. Its exact byte size and SHA256 are listed on the release page. Model weights, personal settings, logs, locks and reference libraries are excluded. The online demo's filter counts describe its sample subset, not the full collection.

Download the ZIP above, extract it, and keep the extracted folder together. To verify the download on Windows, run Get-FileHash .\MOSS-Image-Character-Catalog-v0.1.2.zip -Algorithm SHA256 and compare the result with SHA256SUMS.txt. You need Python 3.10+ and a web browser to view and manage images; GPU hardware is only needed for separate generation.

The preferred address is http://127.0.0.1:8797/. If it is occupied, use the address opened/reported by the launcher. Stop server ends this catalog's process; closing a tab is separate. A compatible Python is reused, or the launcher prepares a private environment. Initial dependency setup may need internet access.

This check concerns the local catalog application, separately from the ComfyUI image-generation environment below. The tested combination was Windows, CPython 3.13.12 (64-bit) and Pillow 12.3.0 with WebP support, in a private test environment. Python discovery was restricted to that 3.13 environment so the launcher could not silently fall back to Python 3.10.

Confirmed in this check:

• Catalog and comparison Python modules imported successfully. The Windows PowerShell bootstrap selected Python 3.13.12 and launched the catalog.
• The catalog / and editor /editor routes returned HTTP 200; /api/info and /api/list returned valid JSON.
• The catalog image inspector decoded one 896 × 1184 WebP sample from the v0.1.0 ZIP.
• The launcher, bootstrap and Pillow requirements used in the test matched the released ZIP. The temporary HTTP service was shut down after testing.

Not covered by this check: browser rendering or interactive operations; Add Image, editing, favorites, Hide/Undo, comparison judgments or review-JSON import end to end; loading or browsing all 34,158 images; macOS/Linux; or every Python/Pillow version and installation type. The HTTP test used an empty isolated catalog fixture, not the complete collection. Successful module imports do not establish that the comparison workflow was tested end to end.

Python 3.10+ is the declared minimum, not a claim that every later version has been fully tested. These results show that the limited checks above pass on 3.13.12; they do not explain an individual user's startup failure. For a failure report, include the actual Python/Pillow versions and the startup error text or app/outputs/review_server.err.log, if created.

A separate Codex Cloud verification report records tests of the official v0.1.0 Release ZIP on Debian 13, x86_64, using isolated environments and Pillow 12.3.0. The downloaded ZIP matched the published SHA256. This records the supplied cloud report; the underlying cloud logs were not independently rechecked in the Windows session.

Reported coverage: the shipped Linux launcher with an explicitly selected Python; syntax checks for 13 Python files and imports of 11 modules; eight real Release images (six characters and two styles); searching, enlargement, clipboard copying, sorting, language/theme switching, editing and saving, registering a test PNG, and persistence after reload and normal stop/restart. Browser operations used headless Chromium 154.0.8037.57. Japanese characters and spaces in application/environment paths and filenames were tested. Only test copies were edited; product code and original data were unchanged.

Not covered by the cloud run: Windows-specific startup, macOS, automatic Python discovery, automatic environment/dependency preparation, automatic desktop-browser opening, all 34,158 images or full-collection performance, end-to-end comparison/review-JSON import, or other Python patch versions/platforms. No application unit tests were bundled, so none were run. No required-function failure was reproduced; the optional /favicon.ico request returned 404, without blocking the tested operations. No product fix was applied.

The Linux results extend the tested combinations; they do not turn the limited Windows check above into full Windows validation or identify the cause of an individual user's startup failure. “Python 3.10+” means 3.10 or newer, not 3.10 only.

Addresses Issue #1: escape & in the Windows launcher title and replace parsing of py.exe -0p output with a query to Python for sys.executable. A working PATH interpreter is checked first; the comparison bootstrap receives the same fix. Discovery does not request automatic Python installation.

Checked for this patch: Windows PowerShell 5.1 / cmd.exe, Miniconda Python 3.13.11 and Pillow 12.2.0. Sixteen startup cases passed; an isolated copy of the real app served the catalog, editor and JSON APIs and shut down normally. The comparison setup entry point passed --help. Py-only and old-default cases used labelled test launchers. The reporter's exact discovery failure, every Python install-manager configuration, full-collection load, all browser interactions and renewed macOS/Linux checks remain unverified. This is not a confirmed general incompatibility with Python 3.13. These checks are separate from the previously documented Python 3.13.12 / Pillow 12.3.0 results. Changes and validation limits.

The initial v0.1.2 release retained the 34,158 images from v0.1.1. The 2026-10-27 refresh adds 93 styles while preserving those existing images and the uncurated state. Extract the entire ZIP into a new folder; preserve your previous folder if it contains personal images, edits or review data.

Copyright (most first) still groups works by their saved image count. Within each work, character images now follow the original source CSV count, descending, instead of alphabetical order. Ties use name, then image key; new user images without source counts appear after counted entries within the same work. Name and generation-order options are unchanged. The source count is a recorded reference value, not a current popularity measurement. The 33,937 character images and 221 styles are unchanged; no pass/fail selection has been added. Release changes.

1. Select Characters or Styles. Characters start in Copyright (most first) order, grouped by series, with characters inside each work ordered by source CSV count (descending). Other works / Original appears last. Works with 20 or fewer saved images are grouped there.
2. Use Generation (newest first) or Generation (oldest first) for the saved MOSS image chronology. Click an image to inspect it. Copy style copies the style description without the shared demonstration scene.
3. Select Add Image, choose Character or Style, then select a WebP, PNG or JPEG. Names, category, reusable prompt, original tags and workflow JSON are optional. Register saves a local catalog entry; blank names use the filename.
4. Web image search opens your entered name/category in a new tab to help with research. Search results are not automatically imported. Use Edit in an image's detail view to correct its metadata later.

Comparison is a separate human review screen, not an automatic quality judge. Configure your own AnimaDex reference installation for reference comparisons. MOSS style images can also be judged without Krea2/AnimaDex images; missing references are labelled and judgments are explicitly MOSS-only.

1. In comparison, choose Accept, Reject or Hold.
2. Select the decisions you want to share and export the confirmed JSON.
3. In the catalog toolbar, choose Import review JSON, select that file and inspect the preview.
4. Choose Apply confirmed decisions. Accept shows/restores an image; Reject hides it. Image files are not deleted. Hold, unselected and unconfirmed rows remain unchanged.

Import checks the actual MOSS image SHA256 and rejects changed/missing images. Share the JSON with people using the same catalog images. CSV list export is a different function; it is not the confirmed-decision import format. Personal references, favorites, hidden flags and review verdicts are excluded from the Release ZIP.

These settings document the actual saved catalog workflows, not universal recommendations. The character batch prioritized practical throughput across many candidates; the style studies used a longer, unaccelerated SFT workflow.

The character workflow uses SFT, not Base. An 8-step acceleration LoRA at 1.0 made the large run practical. It is not a maximum-quality benchmark of unaccelerated SFT or BF16. The first 11,694 character images used flux2-vae.safetensors; the later 22,243 used full_encoder_small_decoder.safetensors. Per-image saved workflows preserve those differences; the portable character sample uses Small Decoder.

Character prompts combine the source name, work and appearance tags with one centered subject, full-body framing, generous margins, a light-gray background and soft lighting. Human subjects receive a clothed standing-pose instruction. Source tags, the reusable copy prompt and actual generation prompt are distinct. The MOSS prompt-enhancer skill was not used for these character batches.

All 314 style images were checked against saved workflows: SFT BF16 diffusion, INT8 encoder, Small Decoder, no LoRA, 25 steps and CFG 2.5. English descriptions specify the style's visual characteristics, then append this common scene:

A single adult woman with short dark brown hair, wearing a cream shirt and a navy apron, carefully arranging a small bouquet of yellow and white flowers on a wooden worktable in a bright flower workshop. Three-quarter view from waist up, natural hand gesture, soft window light, simple uncluttered background. One person.

NovelAI-, Animagine- and niji-inspired entries are descriptive experiments generated with MOSS-Image, not those services/models. Names alone do not establish faithful reproduction. Changing the checkpoint, quantization, LoRA, steps, CFG, wording or seed can change the result; a weak sample does not establish that the subject is impossible to generate.

Using the two samples after release: drag the appropriate JSON into ComfyUI, select your installed model files in the Loaders, replace the character or style/scene prompt, and generate. Both examples use standard Save Image (PNG). The catalog's WebP quality-70 conversion and registration are separate steps; the samples do not register images automatically. Actual style batch saving used SaveImageKJ for an intermediate PNG, followed by WebP conversion with prompt/workflow metadata and a separate workflow JSON.

Use a MOSS-Image-capable ComfyUI. Sample subfolders may differ from yours; choose the installed files in the corresponding Loaders. Models and their licenses are provided by their respective publishers.

RTX 5060 Ti (16 GB VRAM), Ryzen 9 3900X, 64 GB system RAM, Windows 11 Pro. This is the rendering environment, not a minimum requirement for browsing the catalog.

The UI is adapted from Krea2 Character Catalog and aims for familiar AnimaDex-like browsing. This is an independent project, not an official OpenMOSS, Krea2 or AnimaDex product.

Maintainer-authored documentation and the published demo code are covered by MIT. Generated sample images and third-party character/model rights are addressed separately in NOTICE.md. Model weights, reference libraries, credentials and private review state are not published. Report an issue with the image/style name and catalog version; accuracy corrections are welcome.

The refreshed v0.1.2 ZIP and demo fix search carry-over during view navigation and restore filter scroll after reload. Images and catalog data are unchanged. Previous ZIP/checksum assets are retained. See details and validation; use the current Release checksum.