Ija cufattee suuraa kaasuu

AI moodeelota yeroo ammaa keessaa tokko suuraa akka siif kaasu gaafadhu; keessaa wantoota baay’ee adda ta’an lama keessaa tokko ta’a. Familiar image generators sentence irraa photograph uuman pixels kallattii paint godhu, yeroo hojjetanis canvas argu. Garuu gosa drawing lammaffaa tasgabbaa’aa tokko jira: moodeela paint hin godhu. Instructions barreessa: as circle kaasi, achi sarara kaasi, shape kana blue guuti. Instructions kun koodii dha — Scalable Vector Graphics, yookaan SVG, icons/logos web hedduu duuba jiru — fi appeal isaa dhugaa dha: bu’aa dhaabbataa pixel grid miti, shapes rescale/recolour/edit forever gochuu dandeessu blur malee.

Blue technical blueprint SVG document, vector paths, Bezier handles, editable nodes, grid lines fi small specification panels waliin.
Jalqabaa SVG blueprint artwork: image ofii isaa editable vector file, dhaabbataa pixels irraa osoo hin taane paths, grid sararoota, nodes fi handles irraa ijaarame.Laura Nesso / The Clean Paper

Rakkoon: yeroo dhihoo dura moodeela drawing koodii barreessu odeeffannoo garee dhoksee ture. Instructions sequence guutuu tokko pass keessatti baasa — circle, sarara, fill — rendering godhee maal akka uume hin argu. Ija cufattee fuula namaa sketch gochuu yaadi: ija tokko gara bakka sirrii kaa’uu dandeessa, garuu lammaffaan cheek irratti bu’e yookaan hair nose irra taa’e hin argitu. Moodeelota kun baay’ee koodii syntactically valid garuu visually messy uumu sababa kana.

Gareen Guotao Liang led godhe waan jechuun ni danda’ama obvious godhe: moodeela “ija bane”. Mala isaanii Render-in-the-Loop half-finished drawing tarkaanfii tokkoon tokkoon booda render godhee picture sana modeltti deebisa, next stroke dura. Interesting kutaa mul’ataa feedback gargaara qofa miti — gargaaraa akka ta’u gochuuf maal gochuu qaban dha.

Akkam do you score a drawing?

Barruu moodeela tokko “beats” kan biraa yeroo jedhu, maal irratti, akkam safarame jechuun fair dha. Generated picture judge gochuun ulfaataa: “laptop kaasi” deebii tokko hin qabu. Damee automatic scores muraasa proxy person looking at bu’aa ta’an irratti hirkata.

FID garee generated images waliigalaan istaatistikii flavour dhugaa ones waliin wal bira qaba; gadi-aanaa caalaatti gaarii, garee describe malee picture tokko miti. CLIP score adda AI image ajaja jechoota waliin match moo gaafata — faayidaa qabu, garuu judge sana qofa hanga sharp. DINO, SSIM, LPIPS reconstructed image galma waliin pikseloota jalqabaa irraa learned amaloota sadarkaalee irratti compare.

Kanneen keessaa tokko illee truth miti. Proxies dha, increments yeroo baay’ee xiqqaa. Moodeela 127.6 vs 128.8 score yoo argate intended direction dhugaa garaagarummaa dha, garuu nudge dha, landslide miti; lakkoofsa loosely eye judgment hordofu irratti nudge. mata-duree “moodeela 20× daataa irratti leenjifame caale” yeroo dubbistu yaadachuun gaarii.

Barreessitoonni maal godhan

Rakkoon odeeffannoo garee dhoksee drawing dha. Jiru moodeelota SVG writing pure text hojii akka ta’etti ilaalu: koodii hanga amma jiru irraa next chunk predict, render hin godhan. Ammayyaa multimodal moodeelota “eyes” — vision encoder — qaban garuu idle dhiisu. Render-in-the-Loop job step-by-step mul’ataa adeemsa godha. Drawing koodii fragment tokko booda partial SVG image ta’ee render, moodeela picture akka ta’etti receive, next fragment canvas yeroo sanaa ilaalaa filata.

Argannoo jalqabaa cautionary dha, barreessitoonni ifaa gabaasa: loop kana off-the-shelf moodeela irratti bolt gochuun hin hojjetu. Intermediate renders cimaa waliigalaa moodeelota addaa leenjii malee kennan, qulqullina hin fooyya’in — gadi bu’e keessatti board. Moodeela eyes hojii kanaaf akkam fayyadamu hin baranne suddenly hin beeku.

Kanaaf hojii baay’een teaching keessa. Leenjii daataa rebuild godhanii drawing tokko xiqqaa, visually hiika qabu tarkaanfiiwwan keessatti qoodu — walxaxaa shapes salphaa pieces taasisu, stage tokkoon tokkoon waan haaraa argu akka jiru. Achiis eight-billion-parameter banaa moodeela Qwen3-VL irratti built sequences step-by-step kana irratti fine-tune godhan. Mul’ataa Self-Feedback jedhu. Drawing yeroo irratti adeemsa hojii lammaffaa Render-and-Verify dabalu: stroke haaraa accept dura render godhee picture dhugumaan jijjiire moo last tokko repeat qofa sakatta’a; strokes homaa hin dabalne discard; nothing helps yeroo moodeela stop jedhamee prompted. kuusaa daataa comparatively xiqqaa — fakkeenyonni ~850,000, rival tokko daataa keessaa half gadi, kan biraa fraction xiqqaa.

Maal argatan

  • Akkas leenjifame moodeela odeeffannoo garee dhoksee counterpart caalaa gaariitti kaasa — kufaatii haalawwan keessatti ifaa: odeeffannoo garee dhoksee moodeela eye cheek irratti, yookaan requested bar chart skip godhee waliigalaa monitor uuma.
  • Istaandaardii madaallii wabii MMSVGBench irratti mala cimaa rivals waliin competitive, measures tokko tokko irratti slightly ahead — OmniSVG daataa dachaa lamaa ol, InternSVG gara dachaa 20 daataa irratti leenjifame dabalatee.
  • Margins xiqqaa. Icon set irratti ijoo image-quality score 127.6 vs isa gaarii rival 128.8; prompt-matching 0.293 vs 0.291 — dhugaa/wal-simu, garuu slender.
  • Pieces lamaan gatii qabu: addaa leenjii yookaan drawing-time verification off godhi, scores measurably drop. Verification moodeela thing wal-fakkaataa irra deddeebi’ee redraw keessatti stuck ta’u ittisa.
  • Barreessitoonni emphasize efficiency: leaders caalaa far xiqqaa leenjii daataa irraa raawwii kana ga’uu.

Kun maal hin mirkaneessu

  • Moodeela “akka argu” gochuun free win miti. Barruu ofii mul’ataa feedback retraining malee caalaatti hamaa godha. Gain retraining irraa, eyes qofa miti.
  • Guddaa decisive lead hin establish. Scores hedduun neck-and-neck; “20× daataa moodeela caale” true, garuu proxy safartuuwwan nudges, madaallii wabii tokko irratti.
  • Waliigalaa artistic ability hin demonstrate. 8B research moodeela icons/salphaa illustrations xiqqaa dhaabbataa 224×224 resolution irratti, waliigalaa designer miti.
  • GPT-5 fakkaatu waliigalaa moodeela waliin walbira qabuu apples-to-apples miti: moodeela sun narrow code-drawing hojii kanaaf built/tuned miti.
  • Free miti. Every tarkaanfii render/re-read generation odeeffannoo garee dhoksee one-pass caalaa suuta godha.

Ragaan hammam cimaa dha

  • Jajjabaa to’atame walbira qabuu terms ofii irratti. Ablations ifaa: leenjii remove yookaan verification remove, lakkoofsota fall; wantoota bu’uuraa lamaan hojii claimed godhu.
  • Mataa isaa surprise irratti honest. Naive mul’ataa feedback hurts argannoo buried miti; intuition “caalaa galtee always caalaatti gaarii” sirreessa.
  • Leaderboard himannaa irratti haphii. Larger-data rivals wins xiqqaa, tokko madaallii wabii rival barreessitoonni keessaa tokko built; proxy safartuuwwan margins xiqqaa suggestive, settled miti.
  • Safartuu/wild keessatti untested. Icons/salphaa illustrations gadi-aanaa resolution qofa; walxaxaa/high-res/haala addunyaa dhugaa qophii fuuldura hojii.

Maaliif barbaachisaa dha

Core yaada jechuun ni danda’ama embarrassingly salphaa, drawing ala illee bal’ata. Program koodii barreessuun something uumu — web page, chart, diagram, 3D mul’ata — guutuu thing odeeffannoo garee dhoksee barreessuu yookaan yeroo deemutti render godhee course sirrii gochuu danda’a. Namaa page yeroo hunda ilaala; moodeelota keessatti dhiyoo move dha.

Barruu worth dubbisuu taasisu asterisk isaa. Moodeela “argu” karaa kennuun “ilaaluu” barsiisuu waliin tokko miti. Eyes train ta’uu qabu; achiis loop payoff godha — payoff dhugaa garuu safarame: draughtsman steady caalaa, artist gosa haaraa miti. Meeshaalee description irraa editable graphics apps icons/illustrations uuman build godhanf lesson hojiirra oolu: win cheaper, smarter leenjii irraa, caalaa daataa qofa miti.

Cuunfaa gabaabaa

Vector graphics uumu afaan moodeelota traditionally “odeeffannoo garee dhoksee” — drawing commands hunda barreessanii bu’aa render hin argan. Guotao Liang gareen Render-in-the-Loop propose: half-finished drawing tarkaanfii hunda render fi modeltti feed, next stroke canvas ilaalaa godha. Central honest argannoo: jiru moodeela irratti simply kana dabaluu caalaatti hamaa godha; gain moodeela mul’ataa feedback fayyadamuuf retrain booda, drawing-time sakatta’a useless repeated strokes discard godhu waliin. Retrained 8B moodeela istaandaardii madaallii wabii irratti rivals hanga 20× caalaa daataa irratti leenjifame match/slightly beat godha — genuine bu’aa, garuu xiqqaa margins proxy scores irratti, icons/salphaa low-res illustrations keessatti. ergaa ijoo barreessitoonni emphasize efficiency: hojii ofii sirriitti ilaaluun sheer daataa safartuu bakka tokko bu’uu danda’a.

Sakatta’a ifaa

barruun maal agarsiisa: Vector-graphics moodeela mataa isaa half-finished drawing step-by-step render/look gochuuf retrain gochuun odeeffannoo garee dhoksee drawing caalaa caalaatti gaarii/guutuu bu’aawwan uuma, istaandaardii madaallii wabii tokko irratti larger-data rivals waliin competitive/slightly ahead, xiqqaa daataa irraa.

Waan amansiisaa ta’e garuu hin mirkanoofne: “Seeing your hojii” broadly daataa scaling caalaa recipe; wal-fakkaataa yaada walxaxaa/high-res/haala addunyaa dhugaa graphics irratti help; xiqqaa madaallii wabii margins humans notice garaagarummaa ta’uu.

Waan hin agarsiifne: Mul’ataa feedback alone helps (retraining malee hurts); guddaa decisive lead; general-purpose drawing; fair head-to-head waliigalaa moodeelota GPT-5 fakkaatan waliin.

Daangaawwan ijoo: Madaallii wabii tokko gariin rival barreessitoonni built; xiqqaa margins proxy safartuuwwan; 8B moodeela, icons/salphaa illustrations 224×224; render-every-step loop generation slower.

Dubbisaan waliigalaa amanamummaa hammam qabaachuu qaba? Closing loop — drawing yeroo render/re-read — genuinely helps booda leenjii irratti ol-aanaa. Kun moodeela rivals decisively beats irratti low-to-moderate; “beats 20× daataa” dhugaa garuu modest single-benchmark bu’aa. Yaada tokko irratti ol-aanaa: koodii drawing keessatti yeroo deemtu ilaaluu odeeffannoo garee dhoksee drawing caala.

Maddoota

Irratti hundaa'e: Render-in-the-Loop: Vector Graphics Generation via Visual Self-Feedback — Guotao Liang, Zhangcheng Wang, Juncheng Hu, Haitao Zhou, Ziteng Xue, Jing Zhang, Dong Xu, and Qian Yu, Preprint (arXiv:2604.20730).

Yaada gulaalaa

Barruun kun AI'n kan barreeffame yoo ta'u, garee gulaalaatiin ilaalameera. Hojii walqabate sanaaf ibsa ifaa fi of-eeggannoo qabuudha; hojii sana dubbisuu hin bakka bu'u. Filannoo, hiika fi jecha xumuraatiif itti gaafatamummaan gulaalaa bira jira.