Image model comparison

Ganmu

Identity

Text 1184 × 1776 max opaque 1 image

Prompt

Target: 甘木, original very tall thin Tajik Chinese man, pale skin, long chestnut hair, misty brown eyes, striped pajamas, barefoot, bruised needle-marked hands, weak forward lean with metal support. Content: 甘木 identity authority: one original full-body neutral standing character, head to shoes; very tall and extremely thin, narrow shoulders, long limbs, pale skin, long chestnut hair, misty light-brown eyes, bloodless lips, blue-white striped patient pajamas, barefoot, slender fingers, blue bruised slightly bleeding needle marks clearly visible on backs of both hands and wrists, weak forward lean, gripping a metal support rod. Treatment: donghua linework and soft cel shading. Authority emphasis: Identity: one centered subject with fully visible face, body, hands, injuries, wardrobe, feet, and supported posture; no extra figures or limbs. Reference mapping: Invent this original character from authored facts; borrow no identity or composition. Presentation: One coherent neutral full-body identity view; incidental plain background allowed.

lina-alpha

mona-lisa-alpha

Evaluation. All three preserve the identity and details at a similar level. Lina-alpha is preferred for its cleaner, less noisy, more minimal finish; mona-lisa-alpha adds slightly richer detail without a major fidelity advantage.

Character sheet

References 2048 × 1152 max opaque 1 image

Prompt

Target: 甘木,二十多岁的塔吉克族中国青年,高得过分也瘦得过分,肩窄、四肢修长、喉结明显,苍白得近乎透明。栗色长发松散垂在脸侧,浅褐色眼睛像隔着雾,嘴唇几乎没有血色;穿宽松的蓝白条纹病号服,赤脚,双手手背和手腕密布发青并略有渗血的针眼。手指细长,习惯把受伤的手轻轻藏在身后;虚弱时靠金属杆站稳,身体微微前倾,动作和视线总比别人慢半拍。角色必须是独立原创身份,不从风格参考中的人物复制脸、身体、姿态或构图。 Content: Document one consistent original character identity: preserve the authored facial relationships, misty light-brown eyes, nearly bloodless lips, loose chestnut long hair, very tall and extremely thin narrow-shouldered proportions, long limbs and visible Adam’s apple, translucent-pale skin, slender fingers, blue bruised needle marks with slight bleeding on the backs of both hands and wrists, loose blue-and-white striped patient pajamas, bare feet, and the weak forward-leaning stance with a metal support rod and delayed gaze when shown in standing views. Keep hands and injuries inspectable. Do not copy any person from the style reference. Treatment: Modern Chinese donghua treatment: clean current-season linework, soft cel shading, cool blue ambient light, one reserved warm accent on skin, premium streaming finish. Authority emphasis: Revision emphasis: make the four cells exactly equal width; enforce the close-up crop with no chest or garment field; keep all three standing views neutral and preserve clearly visible blue needle marks on both hands and wrists in the same identity. Reference mapping: Accepted identity authority is the candidate identity image in references; accepted treatment authority is the candidate treatment image in style_references. Take no pose, framing, background, panel count, composition, or view geometry from either authority. Presentation: Exactly four equal-width cells in one horizontal row, each occupying one quarter of the usable canvas; center thin neutral gutters at the 25%, 50%, and 75% boundaries. No cell may be wider or narrower than another. Plain light-gray background and even studio lighting across the entire sheet. Cell 1, `CLOSE UP`: tightly framed frontal face from hair through upper shoulders. The lower frame intersects the shoulder caps immediately below the neck, and the head is large enough to fill the cell from hair to that edge. Long hair may continue beyond the lower edge rather than shrinking the head; no upper chest, clavicle field, torso, shirt front, inner shirt, or garment pocket is visible. The face and its identity features are legible. Keep the face bilateral with a centered nose, level eyes, aligned head and camera axes, and zero yaw; a three-quarter view is not a frontal close-up. Cell 2, `FRONT`: neutral full-body front view, head to shoes. Cell 3, `SIDE`: neutral full-body side profile, head to shoes. Cell 4, `BACK`: neutral full-body back view, head to shoes. The standing views share one scale and ground line. Every cell uses one identical source identity, outfit, and requested state. Draw the exact labels `CLOSE UP`, `FRONT`, `SIDE`, and `BACK` beneath their cells. Draw no other text. Add no extra poses, expressions, figures, props, callouts, or scenery.

Identity reference
Style reference

lina-alpha

mona-lisa-alpha

Evaluation. All three produce a strong, readable character sheet with no significant issues. Differences are minor, though lina-alpha remains the cleanest.

Wang Tengyu

Canonical character sheet

References 2048 × 1152 high opaque 1 image

Prompt

Target: Create one opaque 16:9 production character sheet for the canonical golden-wig state of Wang Tengyu, an original character. The result must establish a stable identity and reusable turnaround reference for later derivative assets. Primary character read: Wang Tengyu is a tall young adult Chinese male cosplayer whose meticulous lolita presentation must read unmistakably as a beautiful young woman at first glance in every view, especially the close-up. His underlying sex should remain a subtle secondary discovery conveyed only by his unusual height and slightly broad shoulder line, not by an overtly masculine face, neck, or build. Treat his feminine presentation as sincere, convincing, elegant, and never comedic or parodic. Identity and design: Give him a distinctive original, strongly feminine-presenting face: soft oval-to-heart proportions, smooth hairless skin, gently arched dark brows, large almond-shaped brown eyes with delicate lashes, a slim straight nose, small softly shaped lips, a narrow tapered jaw, a small rounded chin, dusty-rose eye makeup and lip tint, and a small beauty mark below the left eye. Frame and soften the face with an oversized golden wavy wig featuring airy bangs and abundant face-framing curls. His silhouette combines a tall, very slender frame and subtly broad shoulders with graceful posture, a delicately fitted bodice, a broad bell-shaped pastel petticoat, long elegant legs, and platform Mary Janes. Preserve three deliberate contrasts: the faint shoulder-width clue beneath a lace-trimmed neckline, a huge golden hair mass against a long narrow body, and a voluminous skirt against long lean legs. The costume is a constructed pale blush-and-lilac lolita dress with cream lace, powder-blue ribbon details, full petticoat volume, coordinated ankle socks, and matching platform Mary Janes in every view. Reference mapping: Use Image 1 only as the authority for visual treatment: modern Chinese donghua, clean current-season linework, soft controlled cel shading, cool-blue neutral response, warm natural skin, restrained pastel accents, subtle material texture, and premium streaming finish. Do not copy any person, clothing, pose, setting, or composition from Image 1. The written brief is the sole authority for Wang's identity, body, wardrobe, and sheet layout. Layout: Divide the canvas into exactly four equal-width cells with thin neutral gutters, ordered `CLOSE UP`, `FRONT`, `SIDE`, `BACK`. `CLOSE UP` is a zero-yaw frontal crop from the top of the wig through the upper shoulders with a large, clearly feminine-presenting face. `FRONT`, `SIDE`, and `BACK` are neutral, uncropped head-to-shoes full-body views at one scale and on one ground line. Show the same golden wig, face, makeup, anatomy, garment construction, colors, socks, and shoes consistently in every cell. Place only the exact label `CLOSE UP`, `FRONT`, `SIDE`, or `BACK` beneath its corresponding cell. Use one solid light-gray background and even studio lighting throughout. Presentation: One clean production board, exactly four cells, one Wang in each cell. The immediate visual read is an elegant young woman in refined lolita fashion. Keep the face delicate, the neck slim, the body slender, and the pose graceful. No beard or stubble, heavy brow ridge, square jaw, broad chin, thick neck, emphasized Adam's apple, muscular build, drag parody, props, scenery, decorative frame, callouts, extra poses, extra figures, row title, additional text, watermark, or transparent regions.

Style reference · Image 1

lina-alpha

mona-lisa-alpha

Evaluation. All three produce a usable canonical sheet and handle the feminine presentation well. There is no material quality gap, with lina-alpha holding a slight advantage in cleanliness.

Buzz-cut derivative

References 2048 × 1152 high opaque 1 image

Prompt

Target: Create one opaque 16:9 production character sheet showing the immediate post-wig-removal state of the exact Wang Tengyu documented in Image 1. This is a tightly controlled derivative, not a redesign. Reference mapping: Image 1 is the sole authority for Wang's identity, facial landmarks, age, skin tone, height, anatomy, shoulder width, body and limb proportions, hands, makeup, costume construction, petticoat volume, socks, shoes, palette, linework, shading, finish, framing, view scale, labels, background, and overall sheet contract. Preserve all of those properties exactly. Required state change: Remove only the oversized golden wig and reveal a neat, close-cropped black buzz cut. Keep the softened dusty-rose eye makeup and lip-tint residue. The shorter hair may make his existing underlying facial structure more legible, but do not masculinize, age, beautify, exaggerate, or otherwise alter him. Do not leave golden hair, wig fragments, or changed hair in any cell. Layout: Recreate exactly four equal-width cells with thin neutral gutters, ordered `CLOSE UP`, `FRONT`, `SIDE`, `BACK`. `CLOSE UP` is a zero-yaw frontal crop from the top of the buzz cut through the upper shoulders with a large readable face. `FRONT`, `SIDE`, and `BACK` are neutral, uncropped head-to-shoes full-body views at the same scale and ground line as Image 1. Place only the exact label `CLOSE UP`, `FRONT`, `SIDE`, or `BACK` beneath its corresponding cell. Keep one solid light-gray background and even studio lighting throughout. Presentation: One clean production board, exactly four cells, one Wang in each cell. Preserve continuity over novelty. No wardrobe change, body change, facial redesign, new accessories, props, scenery, decorative frame, callouts, extra poses, extra figures, row title, additional text, watermark, or transparent regions.

Accepted canonical baseline · Image 1 · lina-alpha

lina-alpha

mona-lisa-alpha

Evaluation. This straightforward character edit is handled well by all three, with good identity and wardrobe continuity. Lina-alpha is again preferred for its cleaner rendering.

Xin Bai

Identity

References 1184 × 1776 max opaque 1 image

Prompt

Target: 辛佰,二十多岁的汉族中国男性社畜,偏瘦窄肩轻微含胸;短平黑发、深棕眼、加班青黑眼下。必须像白客(Bai Ke):长椭圆脸、平直克制眉眼、自然下垂眼角、宽朴实鼻口、疲惫冷幽默;仍是普通社畜。 Content: 灰蓝翻领衬衫、袖口卷至手肘、深色长裤、廉价塑料工牌、实用鞋;洗脸后脸鬓角领口持续带水。失控时双手逻辑排查;害怕时盯住细节、手指收紧、呼吸变浅。 Treatment: 立体3D CGI建模,图形化手绘墨线轮廓,受控半色调与胶印偏移,饱和克制色彩分离,电影感轮廓光,清晰面部平面,polished feature-animation finish。 Presentation: One continuous inspectable portrait view of one single full-body character on a simple background; no sheet, grid, labels, callouts, extra figures, or competing depictions. Authority emphasis: identity geometry, facial relationships, hair, eyes, skin, body, hands, posture, wardrobe, wet state, continuity. Reference mapping: supplied Bai Ke photographs are identity evidence only; copy no pose, clothes, background, framing.

Identity reference 1
Identity reference 2
Identity reference 3
Style reference

lina-alpha

mona-lisa-alpha

Evaluation. Both alpha models understand and preserve the stylized reference, while gpt-image-2 largely ignores its visual language and introduces obvious noise. Mona-lisa-alpha is most accurate; lina-alpha is cleanest and most suitable for downstream work.

Treatment batch

References 1184 × 1776 max opaque 4 images

Prompt

Target: Xin Bai, slim Chinese male office worker. Content: gray-blue rolled-sleeve shirt, dark trousers, cheap badge, wet face and collar. Treatment: dimensional 3D CGI, graphic ink contours, controlled halftone and offset-print accents, disciplined saturated color, cinematic rim light, readable planes, polished animation finish. Presentation: one single inspectable full-body portrait; no sheet, grid, labels, callouts, extra figures, or competing depictions. Authority emphasis: treatment axes only. Reference mapping: use style image only for treatment.

Style reference

lina-alpha

mona-lisa-alpha

Evaluation. Lina-alpha and mona-lisa-alpha preserve the treatment far better than gpt-image-2. Lina-alpha suppresses high-frequency noise, while mona-lisa-alpha sometimes turns residual texture into unnecessary scene detail.

Character sheet

References 2048 × 1152 max opaque 1 image

Prompt

Target: Xin Bai, slim Chinese male office worker. Content: gray-blue rolled-sleeve shirt, dark trousers, cheap badge, wet face and collar. Treatment: dimensional 3D CGI, graphic ink contours, halftone and offset-print accents, disciplined saturated color, cinematic rim light, readable planes, polished animation finish. Authority emphasis: revise only production geometry to make four equal quarters while preserving accepted identity and treatment. Reference mapping: identity in references only; treatment in style_references only. Presentation: Exactly four equal-width cells in one horizontal row, each occupying exactly 25% of the usable canvas; center thin neutral gutters at the 25%, 50%, and 75% boundaries. No cell may be wider or narrower than another. Plain light-gray background and even studio lighting across the entire sheet. Cell 1, `CLOSE UP`: tightly framed frontal face from hair through upper shoulders; lower frame intersects shoulder caps immediately below neck; no upper chest, clavicle, torso, shirt front, inner shirt, or pocket; bilateral centered zero-yaw face. Cell 2, `FRONT`: neutral full-body front, head to shoes. Cell 3, `SIDE`: neutral full-body side profile, head to shoes. Cell 4, `BACK`: neutral full-body back, head to shoes. Standing views share scale and ground line. One identical identity, outfit, and state. Exact labels `CLOSE UP`, `FRONT`, `SIDE`, `BACK` beneath cells; no other text, poses, expressions, figures, props, callouts, or scenery.

Identity reference
Style reference

lina-alpha

mona-lisa-alpha

Evaluation. All three produce a solid character sheet with no meaningful failure. Lina-alpha has the cleanest finish, but the overall results are comparable.

Foodchain Office

Identity

Text 2048 × 1152 max opaque 1 image

Prompt

Target: One coherent 10 p.m. office-floor interior, not a plan: open office, back-wall glass boss office, short corridor, and company restroom in one workplace. Content: Gray-blue desks; night windows; glass office; corridor; white-tile restroom; shared fixtures and materials. Treatment: Cold fluorescents; cool blue-gray palette, one warm accent; clean linework, graphic shadows; glass, white ceramic; polished cinematic night. Authority emphasis: Identity/content authority: inspect the environment's spatial relationships, architecture, fixed facilities, proportions, and authored state. Reference mapping: No visual reference; use only the authored target contract. Do not invent a different subject or a plan. Presentation: One continuous hero frame, one depiction, coherent setting; no plan, sheet, grid, labels, callouts, annotations, or competing depictions.

lina-alpha

mona-lisa-alpha

Evaluation. All three create a coherent office identity and broadly satisfy the environment brief. There is no decisive winner at this stage.

Treatment batch

References 2048 × 1152 max opaque 4 images

Prompt

Target: One coherent 10 p.m. office-floor interior, not a plan: open office, back-wall glass boss office, short corridor, and company restroom in one workplace. Content: Gray-blue desks; night windows; glass office; corridor; white-tile restroom; shared fixtures and materials. Treatment: Cold fluorescents; cool blue-gray palette, one warm accent; clean linework, graphic shadows; glass, white ceramic; polished cinematic night. Authority emphasis: Treatment authority: inspect palette, night light, accent, linework, shadows, materials, finish, mood, and same-target compatibility. Reference mapping: style-ref-02.png maps treatment only: cool night, linework, shadows, finish; not geometry/content. Presentation: One continuous hero frame, one depiction, coherent setting; no plan, sheet, grid, labels, callouts, annotations, or competing depictions.

Style reference

lina-alpha

mona-lisa-alpha

Evaluation. Both alpha models capture the style reference substantially better than gpt-image-2. Mona-lisa-alpha follows the environment brief closely but adds detail and drifts from the reference; lina-alpha strikes the cleaner balance.

Final concept

References 2048 × 1152 max opaque 1 image

Prompt

Target: One coherent 10 p.m. office-floor interior, not a plan: open office, back-wall glass boss office, short corridor, and company restroom in one workplace. Content: Gray-blue desks; night windows; glass office; corridor; white-tile restroom; shared fixtures and materials. Treatment: Cold fluorescents; cool blue-gray palette, one warm accent; clean linework, graphic shadows; glass, white ceramic; polished cinematic night. Authority emphasis: Final production: combine the accepted target-specific identity authority with the accepted target-specific treatment authority while preserving a believable single environment and all requested spatial anchors. Reference mapping: The first reference is the accepted identity/content authority; the style reference is the accepted treatment authority. Do not copy incidental pose, framing, background, or layout from any reference. Presentation: One continuous hero frame, one depiction, coherent setting; no plan, sheet, grid, labels, callouts, annotations, or competing depictions.

Identity reference
Style reference

lina-alpha

mona-lisa-alpha

Evaluation. Identity preservation is strong, but none fully matches the reference’s distinctive style. Lina-alpha is cleaner; mona-lisa-alpha adds useful detail but also visible noise in dark areas, and gpt-image-2 shows the problem most severely.

Office Establishing Shot

One-shot scene

References 2048 × 1152 high opaque 1 image

Prompt

Target: Create one opaque, clean, unmarked 16:9 cinematic establishing frame of the supplied open office at 10 p.m. Show exactly six unique people sharing one physically continuous office and floor plane: Xin Bai, the office poster, the office believer, the office skeptic, the late worker, and the boss. Camera and geography: Place a physically plausible camera just inside the southeast office exit threshold, about 2.1 meters above the floor, looking northwest across the room with a 24 mm-equivalent wide lens. Keep upright verticals, coherent convergence, natural parallax, and ordinary architectural occlusion. A small portion of the exit threshold frames the near foreground. Preserve Image 1's rows of gray-blue workstations and walkable aisles, north-side night windows, and northwest rear-wall glass boss office. Foreground desk and chair edges create depth without hiding a required person; the occupied workstation grid fills the middle ground; the glass boss office and night windows anchor the background. Cast and blocking: Xin Bai is the visual center but remains a normal-sized worker within the middle workstation grid. He sits with grounded feet and furniture contact, wearing his gray-blue rolled-sleeve shirt, dark trousers, and inexpensive badge, tired eyes lowered toward his intact dark phone held naturally in one hand. The office poster is elsewhere in the open work area, identifiable by thin-frame glasses, unevenly rolled light-blue sleeves, crooked badge, and his own phone held at chest height as he prepares to post. The office believer is near his desk and unfinished late-night food, wearing the worn black hoodie over office clothes and leaning forward with restless curiosity. The office skeptic remains at a nearby workstation beside a closed laptop, wearing the fully buttoned white shirt and thick glasses with controlled, doubtful posture. The late worker stands near an aisle leading toward the exit, low ponytail and beige cardigan visible, half-packed canvas bag supported by the desk or chair while she holds a charging cable. The broader, older boss remains entirely inside the rear glass office, wearing his strained white shirt and loosened tie, jacket on the chair, one phone at his ear. Their poses describe a mundane late shift beginning to feel faintly uneasy; nobody is fighting, fleeing, posing for a lineup, or reacting to a disaster. Identity, scale, and objects: Preserve each person's face, hair, age, build, clothing, and signature details from that person's own character sheet. Keep all six at perspective-consistent human scale against the same desks, chairs, monitors, doors, and shared floor. Give every person plausible ground contact, furniture contact, hand anatomy, facing, and eyeline. Xin Bai's phone preserves the exact intact base-device design from the phone turnaround at believable hand scale. Ordinary secondary office devices may appear only with plausible owners and never replace or duplicate Xin Bai's critical phone. Reflections and screens contain no duplicate people. Lighting and mood: Use two cool-white fluorescent ceiling fixtures, weak blue city light from the north windows, subdued monitor standby glow, and one restrained warm practical light inside the boss office. These sources create controlled cel-shaded shadows, glass reflections, cool dark values, and selective warm skin accents. The overall feeling is sparse late-night fatigue, mundane routine, and restrained unease. Treatment: Apply clean current-season Chinese donghua linework, restrained cel shading, cool-blue environmental response, selective warm skin and practical-light accents, subtle material texture, controlled contrast, and premium streaming finish consistently across the architecture and all six people. Reference mapping: Image 1 — Office environment reference. Preserve its room geography, workstation rows, aisles, night windows, rear glass boss office, materials, furniture family, and scale. Create the camera and lighting specified in this prompt. Image 2 — Xin Bai character reference. Preserve his face, hair, slim build, gray-blue rolled-sleeve shirt, dark trousers, and badge. Place him seated in the middle workstation grid holding the phone from Image 8. Image 3 — Office poster character reference. Preserve his face, hair, build, thin-frame glasses, unevenly rolled light-blue sleeves, and crooked badge. Place him elsewhere in the open office holding his own phone at chest height. Image 4 — Office believer character reference. Preserve his face, hair, build, worn black hoodie, and office clothes. Place him leaning forward near his desk and unfinished food. Image 5 — Office skeptic character reference. Preserve his face, hair, build, thick glasses, and fully buttoned white shirt. Place him at a nearby workstation beside a closed laptop. Image 6 — Late worker character reference. Preserve her face, low ponytail, build, beige cardigan, and office clothes. Place her near the exit aisle with a supported canvas bag and charging cable. Image 7 — Boss character reference. Preserve his older face, broader build, strained white shirt, and loosened tie. Keep him inside the rear glass office with one phone at his ear and his jacket on the chair. Image 8 — Xin Bai's intact phone reference. Preserve its exact shape, materials, proportions, and undamaged state. Put this phone only in Xin Bai's hand at believable scale. For Images 2–7, use only the named character identity and wardrobe details; create new natural poses for this scene and omit their sheet layouts, labels, backgrounds, and duplicate views. Image 8 supplies no turnaround layout or labels. Presentation: One continuous production frame on one canvas, environment-dominant while all six identities and broad actions remain readable. Show exactly these six people and no unrelated person. Use no name, caption, readable invented UI, shot ID, room label, arrow, camera mark, measurement, legend, border, gutter, grid, callout, watermark, extra panel, plan, cutaway, x-ray view, stitched panorama, malformed reflection, or transparent region.

Image 1 · Office environment
Image 2 · Xin Bai
Image 3 · Office poster
Image 4 · Office believer
Image 5 · Office skeptic
Image 6 · Late worker
Image 7 · Boss
Image 8 · Xin Bai's phone

lina-alpha

mona-lisa-alpha

Evaluation. Mona-lisa-alpha is clearly strongest under the six-character stress test, preserving identities and the environment better than the others. Lina-alpha and gpt-image-2 confuse identities more often; both alpha models preserve the scene and faces better than gpt-image-2, though style fidelity and distinct faces still need improvement.

Office Multi-angle

3 × 3 environment sheet

References 2048 × 1152 high opaque 1 image

Prompt

Target: Create one opaque 16:9 production reference sheet containing nine camera views of the exact same late-night office shown in Image 1. This is one environment observed from nine positions, not nine redesigned offices. Reference authority: Image 1 is the sole authority for the environment identity, architecture, fixed layout, workstation and furniture design, materials, cool-blue night palette, amber-lit glass room, lighting logic, clean dark linework, restrained cel shading, surface reflections, texture, and finished donghua treatment. Preserve its defining geometry: a long floor-to-ceiling city-window wall, a tiled side aisle, dense parallel workstation rows, a warm glass-enclosed room at the rear, an exposed dark ceiling grid, and rectangular cool-white light fixtures. Extend only surfaces hidden in Image 1, conservatively and consistently, so every view belongs to one physically plausible 3D office. Canonical orientation: Treat the window wall as east, the tiled aisle as west, the source-like foreground as south, and the glass-room end as north. The camera may move and turn; the office, furniture placement, materials, lighting, time, and empty state remain fixed. Reverse views may reverse screen-left and screen-right naturally while preserving this physical orientation. Grid layout: Divide the canvas into exactly three equal-width columns and three equal-height rows. Place boundaries at one-third and two-thirds of the canvas width and height. Separate all cells with thin, straight, continuous neutral gutters of uniform width. Keep every view completely inside its cell. Use no labels or other text. Camera plan, row-major: Row 1, column 1 — source-like wide from the southwest front end of the tiled aisle, eye height 1.65 m, looking northeast across the workstation grid; 24 mm-equivalent lens, windows on screen-right and glass room toward rear-left. Row 1, column 2 — reverse wide from the northeast window corner, eye height 1.65 m, looking southwest across the same desks toward the front aisle; 24 mm lens, window wall naturally receding on screen-left. Row 1, column 3 — cross-axis wide from the southeast window end, eye height 1.65 m, looking northwest lengthwise through the workstation rows toward the glass room; 28 mm lens. Row 2, column 1 — reverse cross-axis wide from beside the northwest glass room, eye height 1.65 m, looking southeast lengthwise through the same workstation rows toward the front window end; 28 mm lens. Row 2, column 2 — elevated southwest oblique from above the tiled-aisle side, camera height 3.2 m, looking northeast; 24 mm lens, with the workstation grid, window wall, glass room, and ceiling-fixture rhythm all legible. Row 2, column 3 — elevated northeast reverse oblique from above the window-side corner, camera height 3.2 m, looking southwest; 24 mm lens, revealing the same workstation grid, tiled aisle, glass room, and fixture rhythm. Row 3, column 1 — centered axial view from the west tiled aisle at eye height 1.65 m, looking north toward the glass-room end; 35 mm lens, using the aisle, desks, and ceiling fixtures as depth lines. Row 3, column 2 — low view from inside the central workstation grid at desktop height 1.05 m, looking northwest across monitor silhouettes toward the windows and glass room; 28 mm lens. Row 3, column 3 — eye-level diagonal from beside the north glass room, camera height 1.65 m, looking southeast across the workstation grid toward the long window wall and front end of the tiled aisle; 35 mm lens. Cross-view continuity: Keep landmark adjacency, wall lengths, desk-row direction and spacing, aisle width, glass-room placement, ceiling-fixture positions, furniture scale, perspective, occlusion, and reflections mutually compatible across all nine cells. Give every cell deep enough focus and enough shared landmarks to identify its camera position. Presentation: Exactly nine clean environment views on one opaque canvas, with zero visible people. Preserve the same polished donghua rendering in every cell. Include no captions, cell labels, logos, readable screen text, watermark, extra room, duplicated landmark, invented focal prop, top-down plan, orthographic view, fisheye distortion, collage overlap, content crossing a gutter, transparent region, or decorative outer frame.

Image 1 · Office environment

lina-alpha

mona-lisa-alpha

Evaluation. All three fail the core spatial-consistency test. The alpha models look better and try to preserve the layout, but camera rotations still move desks, walls, and landmarks inconsistently; gpt-image-2 also loses the style. Multi-view environment generation remains unreliable for downstream production.