A 3D model pipeline, pushed by GPT-6-Astra

agent (cont.)

+0:20
  • exec_commandpython3 - <<'PY' (+1 more)

    command

    python3 - <<'PY'
    from html.parser import HTMLParser
    class Text(HTMLParser):
     def __init__(self): super().__init__(); self.skip=False
     def handle_starttag(self,t,a):
      if t in ('style','script'): self.skip=True
      if t in ('p','li','h1','h2','h3','h4','tr','figcaption'): print()
     def handle_endtag(self,t):
      if t in ('style','script'): self.skip=False
     def handle_data(self,d):
      if not self.skip and d.strip(): print(d.strip(),end=' ')
    p=Text(); p.feed(open('~/context/reflect/lowmarch.html').read())
    PY
    
    wc -l ~/context/lowmarch.md docs/portrait-guide.md docs/vroid.md && cat docs/map.md && git status --short --branch

    result, 4 lines

    Script completed
    Wall time 0.1 seconds
    Output:
    {"chunk_id":"7da89e","wall_time_seconds":0.00000496,"exit_code":0,"original_token_count":6850,"output":"Lowmarch, Looking Back \nLowmarch, Looking Back \nFrom the lowmarch session, 2026-11-08. Reflection and research only: nothing was built, rendered or changed. Dates are your local dates. First session 1. What happened 2. How professionals do it 3. Tools 4. How to move forward Sources \nStart here: the first session \nYou want to start over with one character, from scratch, so we can work out together how I should work. Nothing is built yet, and I'll wait for you to start it. Everything below this box is background: you don't need to read it or comment on it first. \nThe one thing I'd ask first \nShow me one picture you wish was a screenshot of our game. It can come from anything. It gives me the look, the camera, how big the character is on screen and what they're doing, without you having to put any of that into words. We'd study it only, never trace or ship it. \nThe session, about an hour \nI say back what I see in your picture, in a few lines. You correct me. \nI set up that frame: same camera, same size, a plain grey figure standing in. We judge only there from now on, not on T-poses on black. \nRow 1, big shapes: hair and outfit as rough flat-coloured blocks, 3 or 4 options. You pick one or mark it up. \nRow 2, light and dark: skin tone and the pattern of shadow on your pick, 3 options. \nI write down what you picked and why, in your words. Then we spend five minutes on what was slow or annoying. That's what this session is really testing. \nLater sessions stay in the same frame: the face, then clothes, then one motion. That gets us one whole character, and a record of what each step cost. \nHow I'd show options quickly \nOne link, updated in place, so you never have to hunt for a file. \nYour picture on the left, options beside it, labelled A, B, C. Answer \"B\", or draw on it. \nEach row changes one thing, with under 100 words beneath it. \nRough first. Blocks and colour changes should take minutes, and only your pick gets built properly. If a row will take more than 10 minutes, I'll say so before starting. \nI time every round, from your answer to the next row. I haven't timed this way of working yet, so those first times are the number we try to beat. \nI'm taking \"from scratch\" to mean a new character and a new way of working. The current bodies, VRoid base and renderer stay on the shelf, used only if they help. If you meant those too, tell me, because it changes the first hour. \nBackground: the retro \nShort version \nThe bodies worked. Nothing since has had that feeling. Your last words for a whole thing were it looks perfect (10-12). Since then praise has been for parts. \nThe real complaint all month is parts that don't belong together. You said it best: there has to be cohesion in it (10-31). I turned that into a lash rule. The word was bigger than the lash. \nI never told you the big limit. I don't know if this route reaches what you compare it to, and I said things that sounded like I did. Studios get that look from a person's eye making many hand decisions, per angle. \nWe've never seen the current bodies dressed, moving, or in a game frame. They're 14 bare T-poses, judged on black tiles. \nWhat I'd do: restart on one whole character in one game frame, and have you choose between rows early. Free. One week answers the ceiling question better than more fixes. \n1. What happened 9 active days: 10-07 to 10-13, 10-30, 10-31 13,979 lines of tool code 40 lessons written, 1,440 lines 32 open todos, 19 of them from the first week 0 dressed characters on the current bodies 0 game frames composed, 0 lines of game \nWhere it stands \nCurrent cast, master c4f1015 : neutral, joy, angry, shout. The hair and colour vary a lot. The faces and open mouths look much alike. \n35°, sheet from 10-30, before the mouth work. Yin and Yang rigs by DillonGoo, study only. My read, and my eye isn't reliable: theirs has a big designed shadow down the side of the face, few large dark hair shapes, and small eyes that sit in the face. Ours has near-white skin with almost no shadow shape, many bright strands, and big eyes half under the hair. The first critique on 10-07 said the far side of the face is a flat lit plane . It still is. my read, untested \nCan this route get there? My honest estimate guess \nIt can reach a coherent, likeable cast that clearly belongs to one game, and one character at a time can go above that. I doubt code alone gets all 14 to Yin and Yang's face quality, because theirs rests on someone adjusting points by eye on a mesh built for it. Unicorn Overlord isn't 3D at all: it's painted parts. Nobody has measured this for a current model, so I can't give you a number. I should have told you this on 10-08, instead of the look can be reached . \nWhat has been hard, and why \nArea What you said Why it's hard \nFaces at an angle 20 degrees it looks cursed (10-07), too sharp (10-12), something still looks off (10-30) An anime face isn't a real solid. Studios say it can't fully work from every angle and fix it view by view. 1 We can only nudge a face whose mesh VRoid decided. \nExpressions eyes are weird on the youth (10-30), we can't apply lashes like this like it's generic (10-31) VRoid's eyes, lashes and mouths are paint on fixed shapes. Blending samples breaks what the paint assumed, so 10-31 was a stack of workarounds, each for a number that differs per face (lesson 40). \nBases I think the base model constrains us (10-08), I'm still made wondering about bases (10-10) The base sets the ceiling, and its limits only show once it's reshaped and dressed. See the list below. That fight ended well. \nHair too sharp (10-07), seven pointed bangs , we really need to study how to do the hair thing (10-08), a slice down the middle (10-30) Our own notes called it the weakest part on 10-08, and it still is. A VRoid style is made for one head, and warping it moves its gaps. Our kit makes a fine lock, but nobody has designed a hairstyle you love with it. The study is still open (#735), and so is the slice. \nClothes, motion pieces look floating (10-08), our pipeline isn't polished and iterations take long (10-09) We did these first, on bases later thrown out, then stopped for good reason. Nothing is known about how the current bodies hold up. \nSeeing it hard to explain (10-10), I can't explain it (10-30) You can see it and can't always name it. I can name things and see badly. My lash check passed lids you'd reject at a glance. In the one test I found, models judged visual results well below people (older models; mine is unmeasured). 25 \nEight starting points in six days 10-07 Hand-placed pixels: they don't look anime styled to me 10-07 Code-built head: 20 degrees it looks cursed 10-08 VRoid: the face is great now , then the base model constrains us 10-09 MakeHuman: bodies vary, not anime styled 10-09 VRoid head on MakeHuman body: the head surgery we do is kinda offputting 10-09 Other free bases tested (Quaternius, Blender Studio): none better 10-09 MakeHuman muscle carried onto VRoid: didn't sculpt well 10-10 VRoid shaped by a MakeHuman cage: this looks amazing now , it looks perfect (10-12) \nWhat kept going wrong the same way \nA check appears after you catch the fault. Elaine's arm through her chest (10-08), body holes (10-09), lashes (10-31). Each got a checker afterwards. \nI check in a view that hides the fault. Arms down hid her waist (10-08). The first rig sheet used the old pixel renderer (10-30). Mouths tuned on two bodies broke on the other twelve (10-31). \nA fix breaks what you approved. Saber's waistband (10-08), the heavy woman's thighs (10-12), the combined row's lashes (10-31). \nNumbers stand in for taste. Ratios from paintings stretched the faces (10-12). The heroic man passed every measure and still read as a stretched strong man. Cohesion became a 15% lash limit. \nLessons don't carry. On 10-07 I wrote authored shading beats computed shading , with a forced shadow on the far side of the face. By 10-30 I'd set the face shading so the front stays lit and only the far jaw shades. \nParked means you have to remember. we still have repairs to do for some models, right? (10-31). Mature faces have been waiting since 10-12 (#770). \nThe nature of your feedback \nWhat you respond to \nA whole character beating your doubt: I'm sorry for doubting you. that saber was actually pretty good (10-08). This is the high point of the record. \nSeeing the climb: I loved how we got better while going forward (10-08). \nA long fight ending: it looks perfect, we made headway (10-12). \nA reference beside ours. One look at Unicorn Overlord's arms and you had it: there's no muscle bulge (10-09). They were right about an hour later. \nWhat you reject \nAnything stuck on or one-size: surgery , scar , generic , seven pointed bangs , doesn't fit her body . It's the same complaint each time: this doesn't belong to this person. \nWhat you keep asking for \nWhere we're going: where to? (10-08), what next? (10-12), where are we next headed (10-30). \nHow far this goes: see what our ceiling is (10-07), full confidence (10-08), I wonder how else (10-07). \nFewer rounds: guaranteed after a turn (10-10). \nSimpler: I'm lost (10-30), I just need the sheets (10-31). \nWhat's hard to put into words \nYou reach for images ( cursed , scar ), draw an orange line on a screenshot (10-08), and ask for comparisons so I can imagine it easier (10-30). Your sharpest verdict all month was the blind pick between three rows (10-30). \nIn those moments I mostly named a cause and built: I think I can name the \"something off\" (10-30). I found no case of me making options just to help you find the word. \nWhere my approach made it harder \nToo much, too soon. Costumes, cloth, a duel and enemies by day 2, on a base we later dropped. You called the retreat (10-09). \nNo target frame. Nobody settled what the game shows: where did we get the idea that it will be rendered so small (10-08), until we actually compose game scenes (10-30). Everything is judged on T-posed tiles at four angles. \nYou steered, I brought lists. Nearly every pivot was yours. My best contribution, the cage, came after I wonder what's a way out . I held no map. \nI oversold. On 10-30, when you said it wasn't what you expected, I said the shading layers are what close the gap and promised you can see the gap closing . You never said it closed. Your reference also moved from Unicorn Overlord to Yin and Yang, and I never asked about that. \nResearch after failure. The chin shadow took three builds and a scar before you said we might need to research (10-30). When study came first (sleeves, idle, Saber), results jumped. \nTools grew faster than art. The lessons are nearly all about meshes and caches. Almost nothing records your taste. \nToo much to read, and you as logistics. Long replies with ledger numbers, plus running a command for me, PC schedules, reopening pages. \nThe hair report (10-31) answered your question and changed nothing on screen. \nWhat I'd keep: the bodies and their fit checks, VRoid as the base, the painted renderer and its cast chin shadow, regression checks, the Saber armour kit, the one-lash rule, and blind rows. \nWhat this rests on: your typed messages in the two main sessions, the handoff, the portrait guide, commits, the ledger, sheets, and a fresh reader with no stake in the work, whose read I checked and largely agree with. I did not read comments left on review pages. My guess at what you want, for you to correct: a character you'd love, whole, doing something, like Saber, and someone who says plainly how far this goes. \n2. How professionals do it \nNumbers point to sources, read 2026-11-08. guess marks my inference. Gaps: I found no HoYoverse primary source, and no studio publishes hours per character. The Guilty Gear GDC PDF, vroid.com, Polycount and ArtStation refused automated reads. I couldn't confirm the idea in my own lesson 03 that Guilty Gear stores geometry per angle. \nFaces \nWhat they do \nMake the three-quarter face work first. Straight-on and pure profile are rare, and if 3/4 works, the angles between tend to. 1 \nFour things decide it: the cheek line, brow and eye depth, how far the mouth sits back, the jaw line. The cheek line sets sharp or soft and is called a lifeline for pretty characters. 1 \nCommon failure: squash the head front to back to slim the 3/4, which gives a festival mask . 1 \nBreakage can't be avoided , so fix it with bones or shapes. 1 Gakumas has one shape that pushes the nose dot back so it vanishes in profile. 9 Guilty Gear keeps separate battle and close-up heads, 3 and in finishers the rig moves eyes, nose and mouth so the face reads from that one camera. 4 \nExpressions don't transfer. Uma Musume authors named presets per character, because parts sit differently. 6 \nThe drawing is a guide, and copying it exactly isn't the goal. 1 Arc trains new modellers for a year. 13 \nFor us \nToo sharp probably lives in the cheek line. guess \nThe four moves are small, so code can make them per character at 3/4. guess \nOne shared expression list is the wrong shape. Every character needs a tuned set, which is why the open mouths look alike. \nIf the game shows both battle and dialogue, plan on two faces. guess \nShading a face so it reads as drawn; outlines \nWhat they do \nQuality sits in authored control data, not polygons. Uma Musume: under 20k polygons a character. Its shadow mask, a painted map of where shadow may fall, has the biggest effect on quality and takes most of the texture work. Shape-based shading can't look anime. 5 \nWhere there is no vertex, you can't change the shadow shape. A vertex is a mesh point. 1 \nGuilty Gear paints a value on each mesh point: how readily it goes into shadow, and how thick its line is. 2 3 \nNormals are the direction a surface claims to face, which decides light or shadow. Faces get them edited by hand; hair and clothes get them copied from simple shapes. 1 3 Outlines keep a separate set. 2 5 \nGuilty Gear adjusts each character's light per motion, and per frame in cutscenes. 3 Blue Protocol keeps light to shade at 7:3, because 5:5 went flat in backlight, and cuts the light's angle on skin by half to stop stray face shadows. 7 \nFace shadow maps: paint masks at about 9 light angles and merge them. It's a community method, mirrored from one side. 10 \nOutlines: a dark inverted copy of the mesh gives per-point control, while screen-wide lines are hard to tune. Not showing lines where unintended matters as much as showing them. 2 7 \nFor us \nThis explains the scar : the shadow's edge followed VRoid's mesh points. guess \nAll 26 VRoid faces share one texture layout (lesson 29), so one designed shadow mask could serve them all, tuned per character. This is the best cheap next step for faces. guess \nMoving light per shot is a cheap cheat once we have shots. guess \nHair \nWhat they do \nBundles of locks with thickness. A study of 37,000 VRoid styles found 25 to 130 per style. 14 \nStart from the drawing's outline and shadow positions. Keep only what's needed, and delete locks that get in the way from other angles. 11 Too many fine strands make posing hard. 11 \nBuild the inside first, decide the bones then, and add the outer hair last. 17 \nHair can cost as much as the body: Gakumas 23k to 26k polygons, body 25k to 26k. 16 \nMotion: root keyed, tip simulated, simulation off for dramatic moments. 6 A hobbyist's settings for long cel hair: high stiffness, no gravity, to hold the drawn shape. 18 \nFor us \nOur lock kit is the same family as VRoid's, so building hair in code is sound. guess \nWhat's missing is design: a silhouette first, then few big dark shapes before any strands. guess \nA dark inner layer is still the fix for the slice. \nCorrection: my 10-31 report said to lay back, then sides, then front. No source here confirms that order. \nClothing \nWhat they do \nFolds are placed by hand for common poses, exaggerated, with shapes taken from animation rather than life. 11 Revisions moved to pose shapes an animator dials in. 12 \nSkirts follow the legs first and sway second. Getting the skin weights right (how much each bone moves each mesh point) beats physics settings. 18 22 Colliders on legs alone fail, and a keep-inside cage works better. 19 VRM now has inside colliders. 20 \nNobody reaches zero clipping. Uma Musume adds special motions for big sleeves. 6 Blue Protocol hides clashing parts. 7 Dark Souls capes intersect pretty much everywhere and nobody really notices . 21 \nReal simulation is for big items like capes. 21 \nFabric pattern shows in light and mostly vanishes in shadow. 24 \nArmour: only forum evidence found. Why simulated folds look wrong: no source says so outright. \nFor us \nOur folds are sine ripples that don't change with pose. A few named fold shapes tied to joint angles would suit code. guess \nOur stiff skirt cone is close to the opposite weighting. guess \nGrade clipping by what the game camera sees, not by millimetres. guess \nKeep our cloth sim for capes only. guess \nVRoid, studios, and where code falls down \nWhat's known \nVRoid: hair guides don't survive export. Files hold merged, baked meshes. No scripting API found, and our driver is blocked by Windows until you allow it (#1105). \nBetter base: none found that ships in a closed game. MB-Lab and CharMorph require AGPL, which can block closed-source games; Ready Player Me ended in January; Character Creator is paid and realistic. 32 \nStudios: modeller, texture artist, rigger, technical artist. A freelance listing prices an anime model at about 195,000 to 295,000 yen over 3 to 4 weeks. 31 My estimate for a hero with rig: 3 to 8 person-weeks. guess \nAgents judging looks: in BlenderGym the best model's error was about 10 times a person's on shape edits and 5 times on geometry. Judging which result was better, the best model agreed with people at 0.66; people agree with each other at 0.79. These are older models. 25 \nAgents in Blender: good at scene work, poor at organic forms, proportion and weights, per mostly vendor reports. 29 \nCode-built garments: about 64 to 72% simulate cleanly, with no fine straps and nothing on armour or anime. 27 Models do far better on a small spec language: 94% vs 65% in one test. 26 \nAI-to-3D anime: fused meshes or research-only weights, all trained on VRoid models. 28 \nFor us \nKeep VRoid. Keep our spec language small. guess \nHave code produce shadow masks, normals and line weights, not more geometry. guess \nYou stay the eye, but earlier. One CG studio keeps fixing the block model until it reads as the character. 30 \nA licence catch: our sample exports say Redistribution_Prohibited , OnlyAuthor , commercial Disallow inside the file. I checked two. It's probably the export dialog's default, since VRoid Hub lists samples A to C as open. Pixiv's A to Z help page has never been read; it blocks me. \n3. Tools \nCost: $0. Nothing here needs buying, and I'd buy nothing yet. You said on 10-30 you're open to paid extensions. No paid hair tool I found has a documented way to script it, so they only help if someone styles by hand. Versions read 11-08, except where noted. \nFree What it does Suits us? Licence, how alive Say \nBlender 5.2.3 hair nodes , Curve to Tube Braid, Curl, Frizz, Clump; tube meshes ready for textures Yes. Scriptable, already installed, no install needed GPL. LTS to Jul 2028 try \nVRoid Studio Face and hair editor. Its sliders keep a face coherent, and it's the only source of mature faces By hand, or through our driver once you allow it Free; game use allowed per 10-09 notes. v2.14, Jun 2026 use \nthree-vrm The standard VRM renderer Cross-check our models outside our own renderer MIT. 3.5.5, Jul 2026 use \ntrimesh , glTF-Validator Ray and mesh checks; export file check Yes, for clipping checks and an export gate MIT; Apache-2.0 use \nVRM Add-on for Blender VRM in and out of Blender, by script Only if the Blender trial earns its place. Installing it needs your go MIT or GPL-3. 4.10.4, 4 Nov 2026 later \nBlender 5.2 Cloth and Hair Dynamics nodes Node-based physics One skirt trial, only if a rigged skirt fails. Marked experimental GPL. In our copy later \nVRM 1.0 capsule colliders , spring limits Fewer collision spheres, possibly less hair through bodies. We're on VRM 0.0 A half-day test. Low urgency Open spec. Limits landed Oct 2026, not ratified later \nOur own shadow-map bake A clean face shadow that moves with the light No free Python baker found. QuickSDFTool is Unreal-only. Write it in numpy if a mask isn't enough Ours later \nGarmentCode Sewing patterns from code Ideas only. One garment averages 30k triangles, more than a whole Uma Musume character MIT. 2.0.1, 2024 ideas \nRejected, free: official Blender MCP (running Blender by script is simpler); mcp-for-blender (four security flaws reported by a third party); UniRig (we have a rig); Abnormal (hand tool); CATS (on life support); Goo Engine (doesn't reach three.js); lilToon, Poiyomi, Unity Toon Shader (Unity only, read them for technique); MB-Lab and CharMorph (AGPL); StdGEN and CharacterGen (see above). \nPaid, all on hold Cost Why hold \nStylized Hair PRO $22, MIT (read 10-31) The first buy if you want hair styled by hand. Anime locks, braids, curls. No scripting. A separate build per Blender version. \nHair Tool 4 , Anime Hair Maker $52; $25 to $27 (read 10-31) Hand tools. Stylized Hair PRO comes first. \nMarvelous Designer $39 a month or $280 a year Hand tool for simulated clothes. I found no confirmed way to script it. \nLightning Boy, Anime Shading Plus, Character Creator 5 $30+; $30; $299 Blender 3.4 only per its listing; Unity; realistic and hand-driven. \n4. How to move forward \nRestart on one whole character in one game frame. Everything else waits. Cost: $0, about 10 minutes of your time on most days. \nWhat I'd change \nOne whole character first: dressed, moving, in the frame the player sees. Range comes after. Saber is what won you over, and it was a whole person. \nBig shapes before detail. You approve hair and outfit as flat shapes, plus the light and dark pattern, before any fine work. \nYou choose between rows. Two to four options beside your references, blind when it matters. You point, and I work out why. \nPer-character work over global fixes: a tuned expression set and a designed shadow mask per character, and hair designed from a silhouette. \nStudy first, test the same day. One look at how it's done before the first build of anything new, and every piece of research ends in a test on our model. \nI hold the map. One short page: the target, this week, what's parked. And each week one plain line: how far we are, and whether I still think this route gets there. \nTwo rounds on one complaint with no visible change, and I stop and say so. \nWhat I'd ask you, one at a time \nQuestion What I'd assume if you have no strong view \nWhat do you wish ours looked like now? Five images. Yin and Yang for faces, Unicorn Overlord for costume. Your reference moved, and I never asked. \nWhat does the player actually see? Camera and size, in battle and in dialogue. A side-view battle frame, plus one dialogue close-up. \nWho's the first character, and what are they doing? An original swordswoman on the slender body, in the Saber armour kit, doing one attack. Saber was the high point, and the kit exists. \nWhat's this for: a game you'll ship, or finding how far a model can go? Both, with the second driving. It changes what \"done\" means. \nIs model-only part of the point, or would a person's eye on key decisions be welcome, if free? Model-only for week one. No money implied. \nCould you open pixiv's AvatarSample A to Z help page by hand? It blocks me, and we need it before anything ships. Not urgent, but nobody has read it. \nSmall ones still waiting: the old man's fangs (I'd remove them), and allowing the VRoid driver through Windows (#1105), which any mature face needs. \nA first week \nDay What I do What you see \n1 The questions above, one at a time, played back to you. No code. One page: references, target frame, the character. About 30 minutes of your time. \n2 Put the current body in the frame on plain ground. Test the light and dark pattern: skin tone, a shadow down the side of the face, darker and simpler hair. Three rows beside your references. Pick one. \n3 Hair and outfit as big flat shapes, three or four options. The dark layer under the hair goes in here, which closes the slice. Options in the frame. Pick or mark up. \n4 Her face only: the 3/4 moves, a shadow mask, six expressions tuned for her. Blind rows at 3/4 and in the frame. \n5 Dress her. Folds placed for the pose, skirt following the legs, hair and cloth motion. One idle, one attack. Clipping graded from the game camera. A short clip and a turning view. \n6-7 Your verdict, then my plain note: how close, what's between, what more work buys. Then we decide: next character, more depth, or change route. One page, one decision. \nHow we'd know it worked: you'd show her to someone, or say something like 10-08 again. If you don't, the route gets questioned out loud instead of patched. \nWhat I'd stop doing \nFixing 14 bodies before one looks right. \nJudging on T-poses on black. \nWriting new checkers for looks. Checks stay for things with a right answer: proportions, holes, clearance, regressions. \nPatching VRoid's painted eyes and mouths face by face without asking whether that's the right fight. \nReports that end with nothing changed on screen. \nLedger numbers and long replies during review. A sheet and under 100 words. \nSaying something closes the gap. I'll show it, and you say. \nCarrying 32 open todos. Close or merge the ten or so about retired characters and old dressed work, and keep the rest on the one page. \nSources \nAll read 2026-11-08 through a summarising fetch. I checked the main quotes from 1, 5 and 9, and the figures from 25, directly. Local evidence: docs/portrait-guide.md , docs/vroid.md , the handoff, session transcripts, and the VRM files themselves. \nArc System Works, modelling tips: docswell.com \nArc System Works, toon lines: docswell.com \n4Gamer, Guilty Gear Xrd part 1 (pages 1-2): 4gamer.net \n4Gamer, Guilty Gear Xrd part 2 (page 1): 4gamer.net \nGame Watch, Uma Musume models: game.watch.impress.co.jp \nGame Watch, Uma Musume conversation scenes: game.watch.impress.co.jp \nGame Watch, Blue Protocol, CEDEC 2020: game.watch.impress.co.jp \nGame Watch, Gakumas, CEDEC 2024: game.watch.impress.co.jp \nAnime Shading Plus, face shadow map workflow: erichu33.github.io \nAutodesk cel-look modelling tutorial: part 1 , 2 , 3 \nCGWORLD, Revisions: cgworld.jp \nCGWORLD, Arc System Works feature (partial): cgworld.jp \nCHARM, a study of VRoid hair: arxiv.org \nInside Games, Gakumas budgets: inside-games.jp \n80.lv, long curly anime hair: 80.lv \nZenn, a hobbyist's memo on hair and skirts: zenn.dev \nAvatar makers on skirts: cgbox.jp , signyamo.blog , note.com \nVRM extended colliders: github.com \n80.lv, cloth simulation for games (2020): 80.lv \nMagicaCloth, preventing penetration: magicasoft.jp \nCGWORLD, Soten no Ken: cgworld.jp \nBlenderGym (GPT-4o and Claude 3.5 era; photometric loss): arxiv.org \nDesign2GarmentCode: arxiv.org \nGarmentCodeData: arxiv.org \nAI-to-3D: StdGEN , CharacterGen , PAniC-3D \nAgents in Blender, practitioner and vendor reports: mindstudio.ai , 3daistudio.com \nCGWORLD, Snack World (search snippet, likely page): cgworld.jp \nPrices: Lancers listing (search snippet only, two price tiers) \nBases: MB-Lab licence , CharMorph licence , Ready Player Me shutdown , Reallusion "}{"chunk_id":"5d1738","wall_time_seconds":0.00000437,"exit_code":0,"original_token_count":1293,"output":"   517 ~/context/lowmarch.md\n  1442 docs/portrait-guide.md\n   271 docs/vroid.md\n  2230 total\n# Map of the repo\n\nWhat's current, what's stale, where the big files are. Checked 2026-11-08 at `c4f1015`. Nothing was moved or deleted.\n\"Live\" means the armory build or the first-session kit reaches it. Reach was read from imports, including the\n`__import__` calls in `outfit.py`, not from running every path.\n\n## Docs\n\n| File | Status |\n|---|---|\n| `handbook.md` | **Start here** |\n| `map.md`, `todos-2026-11-08.md` | Current; suggestions only |\n| `portrait-guide.md` | Detail log, 40 lessons. True when written, some since contradicted; marks are in the handbook |\n| `vroid.md` | Current. Says 54 expression shapes at the top; the files carry 57 (2.14) and 56 (older). Licence: see #1443 |\n| `anthro.md` | Current for bodies |\n| `proportions.md` | Bodies current; face ratios are reported, not held, since lesson 30 |\n| `makehuman.md` | Current: references for the cage |\n| `study.md` | Current; one clause is guessed (handbook, Things that don't bend) |\n| `quaternius.md` | Stale: a spike that lost (lesson 21) |\n\n## Tools\n\n**First-session kit (new):** `tools/session/` (`frame.py`, `rows.py`, `smoke.sh`, `smoke_spec.json`), `review/frame/index.html`.\n\n**Render scripts (copied in 11-08 from `~/Projects/lowmarch-bench/kit`, which still has the originals):** `tools/shots/`:\n`arm_shoot.py` (ran 11-08), `variant_build.py`, `all14.sh`, `all14_sheet.py`, `expr_build_tiles.py`, `expr_page.html`. Only\n`all14.sh` changed, to find its own folder.\n\n| Status | Files |\n|---|---|\n| Live: the armory | `cage`, `heads`, `fit`, `proportions`, `anthro`, `makehuman`, `collide`, `vroid_portrait`, `vrm`, `export_glb`, `poses`, `rig`, `fitting`; `review/vroid-02/build.py`, `sheets.py`, `index.html`; `art/cast/cast.py`, `expressions.txt` |\n| Live checks | `lash_check`, `lash_shots`, `checks`, `regress`, `review/vroid-02/snapshot.py` (keeps the baseline `regress` reads), `review/vroid-02/duel_check.py` (build imports `chars_for` from it), `hair_check` (standalone, nothing imports it) |\n| Kept for later: costumes and motion on old bases, reached from the build | `outfit`, `plates`, `gear`, `mobs`, `hair`, `clips`, `duel`, `terrain`, `saber_study` (benchmark, not for the game). Not unused: the armory's inspection shorts come from `mobs.underwear` and the hair texture from `hair` |\n| Spike that lost, still imported by the build | `mass`, `quaternius` |\n| Pixel era, abandoned; still imported | `head3d`, `pix`, `vecpix`, `study` |\n| Blocked | `tools/winvroid` (V.dll, #1105) |\n| Looks dead: nothing imports it | `features.py`, `compare.py` (study side-by-sides) |\n| Looks dead: old one-off tests, nothing imports them | `review/vroid-02/`: `battle_test`, `closeup_test`, `costume_test`, `fit_test`, `pose_test`, `mouths` |\n| Stale pages | `review/rowan-01/`, `art/portraits/rowan*`, `art/templates/`, `review/vroid-02/base_spike_notes.html`, `mh_workshop_notes.html` |\n\n`vroid_portrait.py` is named for the abandoned pixel renderer, but the build still uses its `Portrait` class. Don't edit it\nlightly: it's in cage's cache key, and so are the other cache-key files listed in the handbook.\n\n**Scripts that live outside the repo:**\n\n- `~/Projects/lowmarch-bench/kit/`: the benchmark's kit, including `front_sheet.py`, `build_expr.sh`, `blind_sheet.py`, `terra_review.sh`.\n- `~/.local/share/lowmarch/trials/*/scripts` and loose scripts there, such as `patch_lens_flattice.py` (a perspective lens for the painted page). Copy one in if it's needed again.\n\n## Big files\n\nThe repo's tracked files are small: `.git` is 5 MB. Everything below is outside git.\n\n| Where | Size | What |\n|---|---|---|\n| `~/.local/share/lowmarch/vroid/lowmarch_bodies` | 14 GB | Built bodies and heads; holds cage's cache (`fits.json`). The bulk of everything |\n| `~/.local/share/lowmarch/vroid/samples_2_14` | 466 MB | The 26 VRoid samples |\n| `~/.local/share/lowmarch/makehuman` | 1.7 GB | MakeHuman data |\n| `~/.local/share/lowmarch/trials` | 1.9 GB | Sheets and shots; `combined-2026-10-31` is 1.8 GB |\n| `~/.local/share/lowmarch/tools` | 1.5 GB | Portable Blender 4.5, tar included. Unused since the spike |\n| `~/.local/share/lowmarch/winvroid` | 803 MB | VRoid driver data |\n| `~/.local/share/lowmarch/refs3d`, `study`, `quaternius`, `blender-base-meshes`, `refs`, `anthro` | 1.2 GB together | Study and spike data. References stay out of the repo |\n| `~/.local/share/lowmarch/sessions` | new | First-session folders from `rows.py` |\n| `~/Projects/lowmarch-bench/work` | 4.9 GB | Benchmark builds; `e` is 3 GB. Transcripts and judging keep the benchmark's record |\n| `review/out/` (git-ignored) | 1.1 GB | `versions` 410 MB, `vroid-02` 249, `base-spike` 206, `armory` 115, `mh-workshop` 89, `frozen` 14 |\n| `~/.local/opt/blender-5.2.3-linux-x64` | | Blender 5.2.3, unused by the pipeline |\n\nIf space is ever needed, the safest candidates to discuss are the Blender 4.5 tar, `review/out/versions` and\n`review/out/base-spike`. Nothing has been removed. Ask first, and check `~/Projects/lowmarch-bench/README.md`: the\nbenchmark's raw material is being kept for a blog post (#1141).\n## astra\n"}