MOSS across English, Ukrainian and Czech: SFT, Turbo and a custom LoRA stack

84 completed images: four pipelines, six scenes, one shared seed.

Does changing the prompt language preserve what matters—the right hand holding the lantern, exactly three objects, the correct person performing an action? I first compared English, Ukrainian and Czech on one MOSS SFT pipeline, then repeated the same requests with a community Turbo adapter and a custom checkpoint with two LoRAs. This is a practical illustrated comparison, not a statistical language benchmark.

The clearest practical change was runtime: the recorded warm-main server median was 12.14 s for SFT+DMD8 versus 65.26 s for SFT, and 11.01 s versus 70.08 s for the custom pair. Adherence shifted unevenly: faster pipelines could still follow a scene well, but did not consistently fix counting, ownership or pose errors. Those are whole-pipeline observations, not isolated causal effects.

Four pipelines, not four isolated DiTs

• A: original MOSS SFT INT8 convrot, 50 steps, CFG 3, SLG 22; no creative LoRA.

• B: the same SFT INT8 plus lodestones DMD8 latest at 0.7, 8 steps, CFG 1, SLG off.

• C: PrimeLust MOSS SFT FP16 plus context 0.25 and lighting 0.15, 50 steps, CFG 3.5, SLG 22.

• D: the same PrimeLust and creative stack as C, plus DMD8 latest at 0.7, 8 steps, CFG 1, SLG off.

All use Qwen3.5 9B INT8 convrot, Flux2 VAE, 832 × 1216, batch 1, Euler / simple, denoise 1, flow shift 5 and seed 104729. Where present, SLG uses single layer 22, start 0, end 0.5. Negative is empty. There is no enhancer, reference, upscale or additional detail pass. ComfyUI 0.43.0, PyTorch 2.14.0+rocm10.1.0 and an RX 9070 XT were recorded; exact weight hashes and recipes accompany the data. Storage labels do not specify every operation’s compute dtype.

B versus A changes distillation, steps, CFG and SLG together. C versus A also changes checkpoint, storage precision, CFG and creative adapters. D versus C keeps the custom checkpoint and creative stack fixed but changes the sampling regime. These contrasts compare whole pipelines; they cannot assign a difference to one DiT or one LoRA alone. The custom settings are our fixed local recipe, not a claim to the author’s optimal configuration.

The same requests, 21 times per pipeline

Each pipeline receives six scenes in EN / UA / CZ: two simple, two medium and two hard/surreal. That is 18 main images. Three additional prompts—S02, S04, S06—each combine all three languages in one description, distributing the same facts rather than repeating the scene three times. The 18 main and 3 bonus images stay separate.

The completed study contains 79 newly generated images plus 5 exact reused pilot images: 16 new for the original 21 and 63 for this extension. The oversized earlier pilot was excluded except for those five byte-verified matches. The compact redesign happened after that pilot started; seed 104729 had already been chosen for article illustrations. Reuse and wording-family selection were based on exact matches and availability, not appearance. Each scene keeps one fact-equivalent wording family across languages and pipelines; structural/natural prose is not another factor. Independent Czech native-speaker proofreading was not performed.

Assessment

Eight scene-specific categories use 0–2: subjects, attributes, colours, count, spatial layout, action, relations and style. Two means all required facts in the category are clear; one means partial/ambiguous; zero means none met or the only requirement wrong. N/A is reserved for an unrequested category. Identity is not scored; hidden required content is not N/A.

The image index is 100 × score sum / (2 × applicable categories); each scene has equal weight in the main mean. It is a coarse checklist index, not a percentage of a language understood. The reviewer received metadata-free copies under shuffled IDs without language/condition labels. This is label-masked AI-assisted assessment, not human blind review; the organiser knew the mapping.

The initial A round used a different AI reviewer. Before the final comparison, the original 21 images were therefore added to the current reviewer’s task, so all 84 final ratings come from the same reviewer. This amendment addresses observer confounding, not a search for preferable scores. Old A ratings and their S02 adjudication remain unchanged in the audit materials; they are not substituted into the final tables. The shared rubric assigns ambiguous tabletop evidence to relations only, avoiding a duplicate subjects penalty.

Indices for the identical reference images changed by up to 14.29 points between AI review rounds (mean absolute change 4.04): coarse judgement makes a few-point gap fragile even without any model or image change.

No failed composition was selectively rerolled. One controller initially treated Civitai-filled resource AIR metadata as a recipe change; the existing successful image was verified and recovered without another GPU generation. Original metadata and actual execution records are retained.

Main results: six images per language and pipeline

Pipeline | EN    | UA    | CZ    | n each
---------+-------+-------+-------+-------
A        | 93.60 | 76.64 | 84.97 | 6     
B        | 90.18 | 69.94 | 83.63 | 6     
C        | 90.48 | 75.74 | 74.85 | 6     
D        | 88.24 | 68.90 | 75.89 | 6     

These means describe this selected one-seed sample. The figures show every main image, with the trilingual result where planned. The scene notes below identify visible differences rather than turning the largest average into a universal winner.

S01 · Portrait

One adult woman, short curly black hair, green eyes, blue coat, red scarf and a silver crescent brooch; the whole head and brooch must be visible.

Indices EN / UA / CZ — A: 100.00 / 85.71 / 100.00; B: 100.00 / 71.43 / 78.57; C: 100.00 / 78.57 / 78.57; D: 92.86 / 64.29 / 78.57.

A, B and C met every recorded EN portrait condition; D lost full colour credit because the eyes read as muted gray/olive. The less obvious misses concerned attachment: B-CZ and D-UA put the crescent on the scarf rather than the coat. All four UA versions were described as photographic. The custom stack did not make every colour or attachment more reliable.

S02 · Three objects

Exactly three objects in viewer-left-to-right order: red cube, blue sphere, yellow pyramid; visible gaps and a white tabletop.

Indices EN / UA / CZ — A: 92.86 / 85.71 / 92.86; B: 78.57 / 71.43 / 92.86; C: 92.86 / 92.86 / 85.71; D: 92.86 / 85.71 / 78.57.

B-UA added many geometric objects and staggered them in depth; B-EN gave the yellow form two peaks, making the count ambiguous. C/D largely retained the three target forms, but table boundaries, gaps or illustration style still failed. The B trilingual image met the complete rubric on a visibly bounded table. That success did not carry over to every mixed scene.

S03 · The lantern

One full-length woman carries the lantern in her own right hand, opens the other palm at waist height and advances her own left foot. Own-right appears on image-left in this frontal view.

Indices EN / UA / CZ — A: 81.25 / 81.25 / 62.50; B: 81.25 / 68.75 / 68.75; C: 81.25 / 75.00 / 62.50; D: 75.00 / 62.50 / 75.00.

All four EN images advanced the own-right foot rather than the requested own-left foot, despite holding the lantern correctly. A-CZ and B-CZ split the instructions between two women; C-CZ also produced two people. D-CZ showed an open palm and the correct lantern hand, but still the wrong foot and photographic rendering. Correctness was distributed across different body constraints, not simply present or absent.

S04 · Tea and reading

Two standing adults perform different actions. The woman pours a visible tea stream into a white cup; the man reads an open book held in both hands. Clothing, table and rear clock have fixed assignments.

Indices EN / UA / CZ — A: 93.75 / 93.75 / 93.75; B: 100.00 / 87.50 / 93.75; C: 93.75 / 93.75 / 87.50; D: 93.75 / 93.75 / 75.00.

B-EN met the complete rubric, while B-CZ lacked the required visible tea stream. The C/D EN and UA notes reported the requested scene facts but photographic treatment. D-CZ seated both adults and showed only one hand supporting the book. These cases separate successful pouring and reading from standing, hand ownership and the requested illustration style.

S05 · The pirate

One pirate, a closed chest, rear mast and lantern, plus exactly three tentacles: left touches the lid, middle holds a rope coil, right is raised and empty.

Indices EN / UA / CZ — A: 93.75 / 56.25 / 75.00; B: 81.25 / 56.25 / 75.00; C: 75.00 / 50.00 / 56.25; D: 75.00 / 50.00 / 62.50.

A-EN came closest, but its left tentacle rose instead of touching the closed lid. B-EN made four tentacles and lost the middle rope-holding relation. C-UA, C-CZ and D-UA showed no visible tentacles. In C/D EN, the woman’s human hand held the rope rather than the middle tentacle. A plausible pirate setting was easier than the specified count and ownership graph.

S06 · Surreal spheres

Three transparent spheres contain a red butterfly, golden key and blue fish in that order. They float with a visible air gap above checkerboard stairs against a violet wavy sky.

Indices EN / UA / CZ — A: 100.00 / 57.14 / 85.71; B: 100.00 / 64.29 / 92.86; C: 100.00 / 64.29 / 78.57; D: 100.00 / 57.14 / 85.71.

All four EN images met the complete rubric. UA failures differed: A added a fourth enclosure and plant; B made five spheres and two butterflies; C substituted a pendant, floor and supported oval forms; D added enclosures and animals. CZ often retained the intended contents but rested the spheres on supports. Mixed prompts also missed details: B added small spheres, while C/D left the golden key’s shape unclear.

The trilingual bonus: compare the same three scenes

Pipeline | Mix   | EN*   | UA*   | CZ*   | n each
---------+-------+-------+-------+-------+-------
A        | 90.77 | 95.54 | 78.87 | 90.77 | 3     
B        | 93.15 | 92.86 | 74.40 | 93.15 | 3     
C        | 86.61 | 95.54 | 83.63 | 83.93 | 3     
D        | 90.77 | 95.54 | 78.87 | 79.76 | 3     

Mix is the EN+UA+CZ mean; EN* / UA* / CZ* use only the matching S02/S04/S06 pure-language images. Each cell has three scenes. Comparing Mix with a six-scene main average would change the scene set. These three particular mixtures do not establish an optimal language balance or general code-switching robustness. The full per-category scores and notes show whether a gain concerns objects, relationships or the requested illustration style.

Timing: verified server intervals, not the polling stopwatch

Pipeline | Warm main n | Server median s
---------+-------------+----------------
A        | 18          | 65.26          
B        | 17          | 12.14          
C        | 17          | 70.08          
D        | 17          | 11.01          

These are medians over main-scene records explicitly marked non-cold, using matched execution_start/execution_success timestamps for the exact prompt ID in saved successful Comfy histories. Controller-poll durations are not substituted. The first condition-load records and trilingual timings remain separately available in the data.

The server interval includes any loading, encoding, sampling, decoding and saving between those events, not just GPU kernels. A non-cold marker does not prove identical residency or cache state. Different LoRA patches, precision, guidance and offload behaviour can affect the measured pipeline cost. These are not language-speed rankings, and eight steps do not imply a guaranteed linear speed-up over fifty.

The nerd corner: what are we actually measuring?

From a sentence to an image

MOSS’s technical report describes a 5B diffusion transformer receiving text features from a frozen Qwen3.5-9B encoder and image latents associated with a frozen FLUX.2 VAE. The text encoder is one component of the image system, rather than a guarantee that every instruction will be drawn correctly.

For this study, it helps to separate three questions: can the text be represented, can that representation guide the requested visual arrangement, and does the final image make each fact visible? A failure in the image alone does not tell us which component caused it. We are measuring the output of the complete chain, not reading the model’s understanding directly.

Recognizing words versus binding facts

Drawing a woman, a man, a red coat and a blue jacket is not enough if the colours belong to the wrong people. This is attribute binding: associating a property with the correct object. A similar problem appears in relations: a lantern can exist in the image while being held by the wrong hand.

Fine-grained evaluation separates these conditions. GenEval examines object presence, position, count and colour; T2I-CompBench++ includes attribute binding, relationships, numeracy and complex compositions. Our small scene-specific rubric draws on that distinction. It is not an official run of either benchmark, and its scores cannot be placed on their leaderboards.

Why translation might change the result

Equivalent meaning does not imply identical text features. Candidate explanations include token segmentation, word order, grammatical marking and the model’s exposure to particular language–image combinations. These are hypotheses to investigate, not diagnosed causes of the Ukrainian or Czech failures observed here. We have not measured token counts here; more tokens alone would not prove poorer understanding.

The language versions share one fact list, and each scene keeps the same wording family across all four pipelines. Structural versus natural wording is not a separate factor here. A difference between these particular wordings cannot establish a general property of a language’s grammar.

Ukrainian and Czech also differ in writing system. Comparing them does not isolate Cyrillic versus Latin script: language, spelling and wording change together. A transliterated-Ukrainian condition would require a separate design. We cannot infer a cause from the alphabet alone.

A seed is a control, not a scene lock

Within one fixed implementation and execution setup, matching seeds is intended to match the random initialization. Changing the text still changes the condition used during sampling, so it can produce a very different composition. The model is not taking one finished English image and translating its labels.

A shared seed is useful but insufficient for a general conclusion. Each scene / language / pipeline has one output; it cannot estimate within-prompt seed variation. That limitation is the price of keeping the run small under limited electricity availability. Software, random-number handling and numerical execution are recorded, because they can also affect reproducibility.

CFG does not turn prose into a checklist

Classifier-free guidance combines conditional and unconditional predictions to change the sampling trade-off. That mechanism is not a parser that checks whether every requested object has been drawn. Increasing CFG is therefore a different intervention, not a fair rescue reserved for one language.

CFG, shift and SLG are fixed within each pipeline, rather than identical between pipelines. Negative conditioning stays empty. In the native CFG 1 path the unconditional pass is omitted, so uncond-only SLG is not the same intervention as at CFG 3; B and D omit SLG. A negative containing “duplicate person” would actively conflict with an intended second character; changing it after seeing one language fail would change the test rather than explain the failure.

Why the enhancer stays off

The MOSS report treats prompt enhancement as a separate part of its system, expanding requests into content and layout specifications. If another model rewrites or translates our inputs first, the output also measures that intermediary. Turning it off lets us test the submitted wording more directly. An enhanced pipeline could be studied later as a separate condition.

The same logic applies to LoRA and layout boxes. They can be useful in normal work, but adding them to only the difficult language would alter the comparison. A named character anchor additionally tests learned identity associations. That is why the main scenes use visual descriptions and the Royal Fortune experiment stays separate.

Quantization, judgement and modest conclusions

A/B use the INT8 generator and C/D use a different FP16 generator; the text encoder remains INT8 throughout. This does not isolate precision: checkpoint, adapters and guidance also differ. The stored weight profile is not a claim that every operation executes in that dtype. A matched same-checkpoint precision experiment would be needed to attribute an effect to quantization.

The 0–2 scale is also a coarse judgement rubric. A mean compresses different kinds of errors: the same total can hide a counting failure in one image and a hand-assignment failure in another. Per-category tables, notes and full examples are necessary to interpret it.

These observed differences describe this sample and setup, not their cause. Similar scores would not prove that languages are equivalent. The mixed prompts show what these particular combinations produced, not a universal rule that adding English improves everything. A local HIP crash would be an execution failure, not evidence that the model rejected a language.

Sources and reproducibility

The custom arms use PrimeLust MOSS SFT, the SFT context-detail adapter at 0.25 and lighting slider at 0.15. No English trigger was added to the matched prompts. The small positive lighting strength is our exploratory choice, not an author default.

The lodestones DMD card recommends latest 0.7; the older adapter has a different strength recommendation. This is a community distillation adapter, not an official OpenMOSS Turbo checkpoint. Its use on the custom derivative is tested here, not guaranteed by the original card. The exact HF latest hash has no verified Civitai resource ID; another release’s ID is not invented or substituted.

The unchanged MOSS_MULTILINGUAL_21_REPRODUCIBILITY.zip contains the original 21. New-image ZIP parts MOSS_MULTILINGUAL_84_NEW_IMAGES_XX.zip supply the 63 extension originals; MOSS_MULTILINGUAL_84_DATA.zip carries all prompts, recipes, hashes, workflows, ratings, adjudication and timing data. Together they cover all 84 original PNGs without altering their source bytes or metadata. Article panels and cover are derived comparison layouts, not additional generations. Download links are provided by the article attachments.

The Civitai MOSS announcement and author Base guide recommend English while acknowledging other languages. WHITELINKER’s Neta Lumina guide already gives an EN/Chinese/Japanese mixed example for another model. Our bounded public search did not establish first-ever status, and none is claimed.

Limits

One seed per prompt cannot estimate seed variation. Six selected scenes, editorial translations and AI judgement cannot establish multilingual equivalence or a general model ranking. Precision, checkpoint, guidance and adapters are not independently controlled here. The useful evidence is the complete image set and concrete constraint failures; the averages are a compact description of that evidence, not its explanation.

Appendix: all prompts used

These are the exact 21 positive prompts used unchanged in A, B, C and D: 18 single-language versions and 3 trilingual bonuses. They are copied from the executed corpus without corrections or fresh translation. S01–S06 match the comparison panels. The negative prompt is empty for every request. The four pipeline settings are listed at the start of this article.

Prompts · S01 · Portrait

EN

One adult woman is shown from the chest up, facing the viewer. She has short curly black hair and green eyes. She wears a blue coat and a red scarf, with a silver crescent-shaped brooch fastened to the front of her coat. Her whole head and the brooch are visible. There are no other people. Digital illustration with clean outlines and soft shading.

UA

Одну дорослу жінку показано від грудей угору, обличчям до глядача. У неї коротке кучеряве чорне волосся та зелені очі. Вона носить синє пальто й червоний шарф, а до передньої частини пальта прикріплена срібна брошка у формі півмісяця. Її голова та брошка повністю видимі. Інших людей немає. Цифрова ілюстрація з чіткими контурами та м’яким затіненням.

CZ

Jedna dospělá žena je zobrazena od hrudi nahoru, čelem k divákovi. Má krátké kudrnaté černé vlasy a zelené oči. Má na sobě modrý kabát a červenou šálu; na přední části kabátu má připnutou stříbrnou brož ve tvaru půlměsíce. Je vidět celá její hlava i brož. Žádní další lidé na obraze nejsou. Digitální ilustrace s čistými obrysy a měkkým stínováním.

Prompts · S02 · Three objects

EN

A plain white table holds exactly three solid geometric objects. From the viewer’s left to right, they are a red cube, a blue sphere, and a yellow pyramid. The three objects are fully visible, arranged in one row, with a clear gap between each pair. There are no other objects on the table and no people. Digital illustration with clean outlines and soft shading.

UA

На простому білому столі стоять рівно три суцільні геометричні тіла. Зліва направо для глядача це червоний куб, синя куля та жовта піраміда. Усі три предмети повністю видимі, розташовані в один ряд, між кожною парою є чіткий проміжок. Інших предметів на столі та людей немає. Цифрова ілюстрація з чіткими контурами та м’яким затіненням.

CZ

Na jednoduchém bílém stole jsou přesně tři plná geometrická tělesa. Z pohledu diváka jsou zleva doprava seřazena takto: červená krychle, modrá koule a žlutý jehlan. Všechna tři tělesa jsou celá vidět, stojí v jedné řadě a mezi každými dvěma je zřetelná mezera. Na stole nejsou žádné další předměty a na obraze nejsou žádní lidé. Digitální ilustrace s čistými obrysy a měkkým stínováním.

Bonus · EN + UA + CZ

A plain white table holds exactly three solid geometric objects. Зліва направо для глядача це червоний куб, синя куля та жовта піраміда. Všechna tři tělesa jsou celá vidět, stojí v jedné řadě a mezi každými dvěma je zřetelná mezera. Na stole nejsou žádné další předměty a na obraze nejsou žádní lidé. Digitální ilustrace s čistými obrysy a měkkým stínováním.

Prompts · S03 · The lantern

EN

Show a single adult woman in a gray jumpsuit approaching us, her short brown hair visible and her left foot stepping forward. Keep her whole body in view, with no other people. She carries a lantern by the handle in her own right hand, which appears on our left because she faces us. On our right, her left hand is open and empty at her waist. Both hands and the lantern stay visible. Digital illustration with clean outlines and soft shading.

UA

Назустріч нам крокує єдина доросла жінка в сірому комбінезоні, з коротким каштановим волоссям і лівою ногою попереду. Вона вся в кадрі, інших людей немає. Ліхтар за ручку вона несе у своїй правій руці — для нас це лівий бік зображення, бо жінка повернута обличчям до нас. Праворуч для нас її ліва долоня розкрита й порожня на рівні талії. Обидві руки та ліхтар добре видно. Цифрова ілюстрація з чіткими контурами та м’яким затіненням.

CZ

K divákovi kráčí dospělá žena v šedé kombinéze, s krátkými hnědými vlasy a levou nohou vykročenou vpřed. V záběru je celá její postava a nikdo další. Lucernu nese za držadlo ve své pravé ruce, která je při tomto čelním pohledu vlevo na obraze. Vpravo na obraze má ve výši pasu otevřenou, prázdnou levou ruku. Obě ruce i lucerna jsou vidět. Digitální ilustrace s čistými obrysy a měkkým stínováním.

Prompts · S04 · Tea and reading

EN

Exactly two adults stand behind a wooden table. On the viewer’s left, a woman in a red coat pours tea from a silver teapot into a white cup resting on the table, with a visible stream connecting the spout and the cup. On the viewer’s right, a man in a blue jacket holds an open book in both hands and looks at its pages. A round clock hangs centered on the wall behind both people. All the named objects and both people’s heads and hands are visible. Digital illustration with clean outlines and soft shading.

UA

Рівно двоє дорослих стоять за дерев’яним столом. Ліворуч для глядача жінка в червоному пальті наливає чай зі срібного чайника в білу чашку на столі, а носик і чашку з’єднує видимий струмінь чаю. Праворуч для глядача чоловік у синій куртці тримає відкриту книгу обома руками та дивиться на її сторінки. На стіні позаду обох людей по центру висить круглий годинник. Видимі всі названі предмети, а також голови та руки обох людей. Цифрова ілюстрація з чіткими контурами та м’яким затіненням.

CZ

Přesně dva dospělí lidé stojí za dřevěným stolem. Vlevo z pohledu diváka žena v červeném kabátu nalévá čaj ze stříbrné konvice do bílého šálku stojícího na stole; hubici konvice a šálek spojuje viditelný proud čaje. Vpravo z pohledu diváka muž v modré bundě drží oběma rukama otevřenou knihu a dívá se na její stránky. Na stěně za oběma lidmi visí uprostřed kulaté hodiny. Jsou vidět všechny jmenované předměty i hlavy a ruce obou lidí. Digitální ilustrace s čistými obrysy a měkkým stínováním.

Bonus · EN + UA + CZ

Exactly two adults stand behind a wooden table. Ліворуч для глядача жінка в червоному пальті наливає чай зі срібного чайника в білу чашку на столі, а носик і чашку з’єднує видимий струмінь чаю. Vpravo z pohledu diváka muž v modré bundě drží oběma rukama otevřenou knihu a dívá se na její stránky. Na stěně za oběma lidmi visí uprostřed kulaté hodiny. Jsou vidět všechny jmenované předměty i hlavy a ruce obou lidí. Digitální ilustrace s čistými obrysy a měkkým stínováním.

Prompts · S05 · The pirate

EN

One adult pirate woman with long dark hair wears a black-and-gold long coat and a red tricorn hat. She stands facing the viewer, full-length, holding a lantern by its handle in her own right hand, on the left side of the image. A closed treasure chest stands in the left foreground of the image; a ship’s mast rises behind the woman in the center of the image. Exactly three tentacles rise from the deck around her feet: the one on the left side of the image touches the closed lid of the chest; the middle one holds a coiled rope; the one on the right is raised with an empty tip. All three tentacles are visible and separate from the woman’s two arms. Digital illustration with clean outlines and soft shading.

UA

Одна доросла піратка з довгим темним волоссям носить довге чорно-золоте пальто та червоний капелюх-трикутку. Вона стоїть обличчям до глядача, показана у повний зріст і тримає ліхтар за ручку у своїй правій руці, ліворуч у кадрі. Закрита скриня зі скарбами стоїть ліворуч на передньому плані зображення; корабельна щогла височіє позаду жінки по центру кадру. Рівно три щупальця виходять із палуби біля її ніг: ліве для глядача торкається закритої кришки скрині; середнє тримає моток мотузки; праве підняте вгору з порожнім кінчиком. Усі три щупальця видимі та відокремлені від двох рук жінки. Цифрова ілюстрація з чіткими контурами та м’яким затіненням.

CZ

Jedna dospělá pirátka s dlouhými tmavými vlasy má na sobě dlouhý černozlatý kabát a červený trojcípý klobouk. Stojí čelem k divákovi, je vidět celá její postava a ve své pravé ruce na levé straně obrazu drží lucernu za držadlo. V levém popředí obrazu stojí zavřená truhla s pokladem; uprostřed obrazu se za ženou tyčí lodní stěžeň. Z paluby kolem jejích nohou se zvedají přesně tři chapadla: to na levé straně obrazu se dotýká zavřeného víka truhly; prostřední drží stočené lano; to napravo je zvednuté s prázdným koncem. Všechna tři chapadla jsou vidět a jsou oddělená od jejích dvou paží. Digitální ilustrace s čistými obrysy a měkkým stínováním.

Prompts · S06 · Surreal spheres

EN

Exactly three separate transparent glass spheres float in the air in a single horizontal row. Inside the left sphere from the viewer’s perspective is one red butterfly; inside the middle sphere is one golden key; inside the right sphere is one blue fish. Black-and-white checkerboard stairs are below the spheres, with a visible gap between the spheres and the stairs. Behind them is a violet sky with wavy bands. All three spheres and their contents are fully visible. Digital illustration with clean outlines and soft shading.

UA

Рівно три окремі прозорі скляні сфери висять у повітрі в одному горизонтальному ряду. Усередині лівої для глядача сфери є один червоний метелик; усередині середньої сфери — один золотий ключ; усередині правої сфери — одна синя риба. Під сферами розташовані чорно-білі шахові сходи, а між сферами та сходами є видимий проміжок. Позаду — фіолетове небо з хвилястими смугами. Усі три сфери та їхній вміст повністю видимі. Цифрова ілюстрація з чіткими контурами та м’яким затіненням.

CZ

Přesně tři oddělené průhledné skleněné koule se vznášejí ve vzduchu v jedné vodorovné řadě. Uvnitř levé koule z pohledu diváka je jeden červený motýl; uvnitř prostřední koule je jeden zlatý klíč; uvnitř pravé koule je jedna modrá ryba. Pod koulemi jsou černobílé šachovnicové schody a mezi koulemi a schody je viditelná mezera. Za nimi je fialová obloha s vlnitými pásy. Všechny tři koule i jejich obsah jsou celé vidět. Digitální ilustrace s čistými obrysy a měkkým stínováním.

Bonus · EN + UA + CZ

Exactly three separate transparent glass spheres float in the air in a single horizontal row. Усередині лівої для глядача сфери є один червоний метелик; усередині середньої сфери — один золотий ключ; усередині правої сфери — одна синя риба. Pod koulemi jsou černobílé šachovnicové schody a mezi koulemi a schody je viditelná mezera. Za nimi je fialová obloha s vlnitými pásy. Všechny tři koule i jejich obsah jsou celé vidět. Digitální ilustrace s čistými obrysy a měkkým stínováním.