Xenosaga Episode I Python Extractor

Unpack, browse and mod Xenosaga Episode I — Der Wille zur Macht (PS2, USA release SLUS-20469) with nothing but Python. Point it at your own disc image and you get every file on the disc in a form your computer can open: textures as PNG, models as glTF for Blender, voices and music as WAV, cutscenes as MP4, game text as UTF-8, and the cutscene scripts as Java you can read. Then write your changes back to a bootable ISO.

The kit ships no game data. You need your own copy of the disc.

The game running from a disc the kit rebuilt, captured in PCSX2 (nothing composited):

What's new (October 2026)

• Translation without length limits. Cutscene and field dialogue now exports and imports like the rest of the text (evt-export / evt-import, any length, all three chains), mails may grow, and an object that no longer fits its slot on the disc can be relocated (--grow). The second layer's chain of 89 cutscenes, 58 movies and 44 voice streams is read like the other two.
• Scripts the game never had. The fonts are decoded and rebuilt byte-exactly, and lang scan / lang encode draw a translation's Arabic, Hebrew, Devanagari (or any HarfBuzz-shaped script) glyphs from a TrueType font into the game's proportional font and encode the text shaped, wrapped and right-to-left where it should be. Two scripts can share one font pair (--extend), mixed-script words keep their Latin halves, and the disc's own symbols and control codes pass through. Menu and battle words, which live as C strings in the code overlays, are a third tree (ovl-export / ovl-import, any length), and the battle screen's own bitmap font gets new letters with statfont. Movie subtitles are rendered by the kit in any script and burned on a black bar over the English with plain ffmpeg (subs-burn --font), timed from the original with subs-template --detect. All of it seen running on PCSX2, drawn by the game itself: Hindi dialogue and menus, Georgian battle bars, a Hindi movie subtitle, an Arabic line, an Amharic mail. docs/SCRIPTS.md. Several of the formats behind this were first worked out by poketony for the Korean localisation XS1KOR.
• Models open in Blender, rigged. Every character, enemy, weapon, mech, prop and map on the disc comes out as a .glb with its texture inside. File ▸ Import ▸ glTF 2.0 and you have the mesh, its UVs, vertex colours, skin weights and a posable armature built from the game's own skeleton files, every bone on the mesh: select it, Pose Mode, grab a bone and rotate. Faces are merged into their bodies. How-to and what to expect: docs/BLENDER.md.
• Faces are complete. Eyes, mouths and eyelashes on every character, the lashes being the last piece: they are a small true-colour alpha strip the lash polygons stretch into strokes, and the decoder had been reading it through a palette. The gallery above is rendered by Blender straight from the kit's output.
• Character textures look right, eyelashes included. Opening models in Blender showed four texture bugs that had been there since the first release, and all are fixed: 42 character atlases that came out as beige noise (Cherenkov, Jr., MOMO, Shion and Virgil variants among them), sleeves and trims drawn with the wrong palette, hair sheets half-missing, and the texture scale on narrower atlases. 756 of the 1,019 model textures on the disc changed, all for the better.
• The casino is fixed. Card faces, the slot machine and the concept-art gallery were the last textures that came out wrong. The game's menu code says which palette colours which part of a sheet, and the kit now reads that. This closes GitHub issue #1.
• More to look at and listen to. Battle shouts and attack sounds, the casino art and gallery sketches, the title and help screens from the menu containers, the memory-card icon, and every string table the kit can sniff out of binary files, including the casino dialogue.
• The soundtrack plays. The sequenced music is rendered to WAV and FLAC, with MIDI and SoundFont files alongside, at the right tempo. The story of how: docs/MUSIC.md.

Still on the list: animation (the motion packs are located but not decoded, so bones pose but do not play) and a few map meshes in formats the model parser skips. The full change log is in docs/UPDATES.md.

Running it

There is a point-and-click app and a command line. Both do the same things.

The app. Download the bundle for your system from the Releases page and double-click it. Nothing to install.

Windows	Xenosaga-I-Extractor-windows/Xenosaga-I-Extractor.exe
macOS	Xenosaga-I-Extractor.app (first time: right-click ▸ Open, because of Gatekeeper)

From a source checkout (Python 3.9 or newer, no packages to install): double-click launch.bat (Windows), launch.command (macOS) or launch.sh (Linux), or run python gui.py. The app opens as a page in your browser, finds your disc image, and shows one card per job with a live log.

From a terminal:

python cli.py extract --iso "Xenosaga Episode I (USA).iso" --out out/ --code
python cli.py classes --iso "Xenosaga Episode I (USA).iso" --out out/
python cli.py browse  --out out/

extract unpacks the disc into out/dump/. browse converts it into out/browse/, which is the folder you actually open: PNGs, GLBs, WAVs, MP4s and text. A guided tour of what is in there, and which folders are worth opening first, is in docs/BROWSING.md.

What you get

Everything on the disc, converted to something you can open.

Pictures and models

• Every texture as PNG: characters, enemies, mechs, props, maps, menus, the card game's 117 card faces, and the casino's slot reels, poker tables and the pencil concept sketches hidden in its gallery.
• 1,400 more textures pulled out of effect, scene and menu containers: battle effects, title and help screens, room backdrops.
• Every model as a .glb with its texture inside, ready for Blender.
• The 58 full cutscene movies the disc hides on its second layer, as MP4 with sound, plus 189 smaller in-game clips. The boot logos and title art. The memory-card icon.

Sound

• Every voice line and cutscene mix as WAV.
• Every battle shout and attack sound: 535 banks, one per character and boss, about 1,500 WAVs.
• The whole soundtrack, rendered from the sequenced music the disc actually carries (Yasunori Mitsuda's name is literally in the files), as WAV and FLAC, with the MIDI and SoundFont alongside so you can remix it. docs/MUSIC.md tells the story.

Hear it. Two short excerpts straight out of the kit, so you know what to expect before you run it: the opening of Battle 1, rendered from the sequenced music (15 s), a voice line from the streamed cutscene audio (8 s), and a string of battle shouts from Jr., KOS-MOS, MOMO and Ziggy's attack and boost banks (16 s). Click to play in the browser. Everything else, the full tracks and all 103 voice streams, comes from your own disc.

Words and code

• All in-game text as UTF-8: menus, text boxes, the U.M.N. mails, item descriptions, the casino dialogue, and the developers' own Japanese comments left in the scene scripts.
• The cutscene scripts as Java class files you can decompile and read (docs/JAVA.md).
• The executables with every function name intact.
• Eight developers' personal working folders that shipped on the retail disc, test maps and all (docs/FINDS.md).

	Where	Format
Textures	browse/textures_png/	PNG
Models	browse/models/	glTF binary (.glb)
Voices and cutscene audio	browse/audio/	WAV
Battle voices and sound effects	browse/battle_audio/	WAV
Music	browse/music/ (the Render the soundtrack card, or music-export)	WAV, FLAC, MIDI, SoundFont
Movies	browse/movies/	MP4
Game text	browse/text/	UTF-8
Cutscene scripts	browse/classes/	Java class files
Executables	browse/code/	ELF and IRX

Translating the game

The kit is built to carry a fan translation end to end, and the pipeline has been proven in-game.

• Menus, text boxes and mails. text-export writes all 914 text objects as plain UTF-8 files, one per object, each marked with how many bytes it may use. Edit them in anything. text-import checks every file fits, re-encodes it, and writes a patched ISO only when all of them pass.
• Cutscene dialogue. Spoken lines in cutscenes live inside the Java scripts. evt-export writes them out, one line per string, and evt-import rebuilds the scripts around translations of any length. Objects that outgrow their place on the disc are relocated with --grow. Same-length byte swaps still work for a quick in-game check; the opening tutorial has been played through in French that way.
• Other alphabets. lang scan builds the glyphs your translation uses from a TrueType font into the game's font files; lang encode shapes, wraps and orders the text (right-to-left for Arabic and Hebrew) and writes the codes the importers take. docs/SCRIPTS.md.
• Movie subtitles. The cutscene movies have no subtitle track at all, so the kit gives you a timing template, lets you write and time the lines in any SRT editor, and burns them into the video with the original audio untouched. docs/SUBTITLES.md.
• Guides. A step-by-step French walkthrough in docs/FRENCH-TRANSLATION-GUIDE.md, the multi-script pipeline in docs/SCRIPTS.md, and the original right-to-left design notes in docs/ARABIC-TRANSLATION-SPEC.md.

Modding

The disc is unusually friendly to patching. The kit writes changes back in place, so everything you did not touch stays byte-for-byte the same as the retail disc.

• Replace any file: python cli.py patch --iso GAME.iso --out MODDED.iso --set 'chain0:char\pc\kosmos.xtx=my_kosmos.xtx'
• Recolour KOS-MOS's hair with one click (pinkhair). It is the worked example because the texture is stored in twelve places on the disc and the kit finds all of them.
• Translate it. See the section above; the same patcher writes the translated text and subtitled movies back.

How all of that works, in enough detail to build on: docs/MODDING.md.

How the disc is put together

Three things explain most of what the kit does.

• The disc has seven big files and nothing else. XENOSAGA.00 to .13 hold all the game data. Read .00+.01+.02 end to end and you have "chain 0", the game world: 8,573 models, textures, scripts and sound banks. .10 to .13 are chain 1, the streamed voice and cutscene data. A table of contents at the front of each chain maps game paths like char\pc\kosmos.xtx to where the bytes are.
• Half the disc is invisible. The second layer of the DVD is not in the filesystem at all. It is 58 full cutscene movies that the game plays by raw sector number. The kit finds them by scanning.
• About 2,100 files are compressed with Monolith's own ARX scheme. The kit's compressor is a byte-perfect clone of the original, which is what makes clean patching possible.

Two bonuses for the curious: every executable still has its function names, and the cutscenes are compiled Java 1.1, debug info and all. The disc also shipped eight developers' personal working folders; docs/FINDS.md and docs/HISTORY.md have those stories.

The Japanese release is expected to work (nothing is tied to the USA serial) but has not been tested on a real JP disc. If something breaks, please open an issue with the exact error.

Docs

Read this	When you want to
BROWSING.md	know what is in the output and where to look first
BLENDER.md	bring the models into Blender
MODDING.md	change the game: patching, recolours, translation, verification
SUBTITLES.md	add subtitles to the cutscenes
FRENCH-TRANSLATION-GUIDE.md, SCRIPTS.md, ARABIC-TRANSLATION-SPEC.md	translate the game, step by step; into other alphabets
MUSIC.md	understand and render the soundtrack
JAVA.md	read the cutscene scripts
FORMATS.md	see the byte-level details of every file format
CODEBASE.md	find your way around the code
UPDATES.md	see what changed and when
FINDS.md, HISTORY.md	enjoy the leftovers and the history the disc records

Building the app yourself

python build.py (needs pip install pyinstaller) makes the self-contained bundle. Build on Windows for the .exe and on macOS for the .app; the GitHub Actions workflow builds both when a version tag is pushed.

Credits

ARX decompression, the texture canvas layout and the model format were learned from Lakuwu's xenotool and re-implemented here, with credit where each piece is used in docs/FORMATS.md. The font files' GS layout, the FL00 shift rules for resized classes, the mail objects' pointer table and the idea of relocating grown objects rather than rebuilding the disc were first worked out by poketony for XS1KOR, the Korean localisation of the Japanese release, and re-implemented here for the USA disc; the engine-side cell addressing, the proportional-slot permutation and the metrics semantics were recovered from the executable for this kit. The ARX compressor, the table of contents, the hidden-layer movie carver, the Java class carver, the audio and music decoders, the overlay-palette and model-material texture work, and the whole repack layer were reverse-engineered for this kit.