Redrob Image

담당

모델 선정(Z-Image Turbo 오픈 웨이트), 블라인드 선호 설계, Studio 베이크오프 운영을 담당. 호스팅은 Studio 셸과 공유.

목적

인터랙티브 베이크오프가 가능한 지연과, 강한 호스티드 베이스라인 대비 선호에서 계속 개발할 근거가 되는 Apache-2.0 오픈 웨이트 이미지 모델. 사용·상업적 이용·재배포 자유이고, 소비자용 GPU 한 장이 차선책이 아니라 기준이 될 만큼 가볍게 만듭니다.

결정

• 자연어 해부학을 위해 Stable Diffusion 폐기

  Stable Diffusion을 Klein 4B, Z-Image Turbo와 비교. SD는 깨끗한 해부학에 좁은 프롬프트 방언이 필요하고, 일반 언어에서는 실패. 그 방언은 수용 불가.

• 포토리얼리즘에도 불구하고 Klein 기각

  Klein 4B는 피부·빛에서 강했으나 그래픽·인쇄 인접 작업에서 타이포그래피를 깨뜨림. 포토리얼리즘만으로는 부족.

• Z-Image Turbo 오픈 웨이트로 출시

  공개 가중치는 Z-Image Turbo 머지(Diffusers transformer 또는 ComfyUI UNET 하나)이고, 한 장에 약 8 DiT 스텝을 씁니다. 베이스보다 피부·빛·텍스처를 현실적으로 맞춤. 코드가 아니라 웨이트 자체가 Apache-2.0: 사용·상업적 이용·재배포 자유.

• 따로 배울 런타임이 없음

  표준 Diffusers 파이프라인 API의 from_pretrained로 불러오고, ComfyUI에는 병합 UNET 하나로 올라갑니다. 경로에 우리가 만든 래퍼가 없고, 가입도 키 발급도 없습니다.

• 터보 샘플링의 대가는 네거티브 프롬프트

  8 스텝은 classifier-free guidance를 끄고서만 성립하고(guidance_scale=0, ComfyUI cfg 1), 가이던스를 끄면 네거티브 프롬프트는 아예 무시됩니다. 피하고 싶은 것도 포지티브 프롬프트에 써야 합니다. 문서로 덮을 문제가 아니라 실제 제약입니다.

• 강점을 명시

  사진·포트레이트·무드에 강하고, 한글을 비롯한 대부분의 비라틴 문자 등 읽히는 텍스트에 약하며, 타이포 비중이 큰 그래픽·인쇄 작업은 사진 결과보다 한참 아래입니다. 선호 스터디 패턴도 같은 방향. 텍스트가 필요한 표면은 frontier API 탈출 경로를 유지.

과제

• 호스티드 API는 반복 실험에 너무 느림

  장당 20초-3분이면 LoRA·체크포인트 스윕이 비현실적. 동일 하드웨어 조건에서 자체 서버 end-to-end 약 1.8초로 인터랙티브 베이크오프가 가능.

• 카테고리 셀은 검정력 부족

  카테고리당 6회 비교로는 스타일별 승자를 확정할 수 없음. 출시는 전체 n=90 선호 결과만 사용. 그래픽·인쇄 셀은 동일 방향으로 실패.

• 한계

  n=90, 내부 평가자, 단일 프롬프트 분포. 설계상 underpowered. 평가자는 시스템 구축 주체에 대해 블라인드가 아님.

성과

• 측정된 경로와 오픈 웨이트

  44-16-30은 계속 개발의 근거이지 동등성 주장이 아님. 지연은 인터랙티브 베이크오프를 지원. 단가는 측정 가동률 50%에서 API 정가의 약 1/40(소유 하드웨어 vs 공개 가격; 가동률 의존). 가중치는 Apache-2.0이며 상업적 이용 가능.

• 증류이지 폐쇄 스택이 아님

  속도는 Z-Image Turbo 베이스 증류에서 옴. 지속 자산은 평가 증거와 오픈 웨이트.

장당 비용

자체 서버 완전 부하 장당 비용 vs GPT Image 2 50원. 측정 가동률 50%에서 API 정가의 약 1/40. 소유 하드웨어 비용 vs 공개 가격; 동일 조건 비교 아님; 가동률 의존.

자체 서버 단가는 고정 GPU 용량이 지배. 가동률 가정 없이 숫자는 정의되지 않음.

결과물

• [Medium and Style] Surreal fine-art editorial photography with analog film aesthetics, heavy grain, soft halation, and poster-like color blocking. Vertical full-body composition, centered and nearly symmetrical, photographed from a slightly low eye-level angle with a normal-to-short-telephoto lens. Strongly stylized red-and-black palette with cyan-blue clothing, resembling an experimental album cover or vintage screen-printed photograph. [Subject(s) and Clothing] A solitary adult human figure stands centered in the frame. The person has a shaved or closely cropped head, but the face and scalp are completely overexposed into a smooth, featureless white silhouette. They wear an oversized pale cyan-blue pullover jacket with a high raised collar, broad dropped shoulders, voluminous sleeves, gathered elastic cuffs, and a loose bloused waist. A subtle curved seam or pocket-like fold appears on the upper right side of the chest. The lower body is covered by very dark, nearly black, loose straight-leg trousers. Both hands are visible as dark silhouettes emerging from the sleeves. No accessories or footwear details are clearly visible. [Pose and Expression] The figure faces directly toward the camera and stands upright with squared shoulders. Both arms hang naturally at the sides, with the hands slightly separated from the body. One leg appears subtly advanced, suggesting a slow forward step. The head is held straight. No facial expression is visible because the face is entirely erased by intense white overexposure. [Background and Environment] A barren, rocky outdoor landscape stretches across the lower half of the image. The foreground is almost completely black, containing irregular stones, low scrub, and sparse branching plants outlined by red-orange rim light. Larger dark rock formations sit behind and beside the figure, including a prominent jagged rock to the figure's lower right. The distant horizon is obscured by red haze, with several soft, rounded dark silhouettes resembling boulders or low vegetation. The upper two-thirds of the frame is filled by a flat, intensely saturated orange-red sky with subtle mottled texture and film grain. [Lighting, Atmosphere and Colors] Extreme frontal or backlit overexposure turns the head into pure ivory-white light and creates a glowing orange-red halo around the shoulders and outer silhouette. The jacket transitions from pale white-blue near the collar and shoulders into deeper cyan and muted blue toward the torso and sleeves. The lower body and ground remain almost entirely black, producing stark tonal separation. Dominant colors are vermilion, scarlet, burnt orange, cyan-blue, ivory white, and deep black. The atmosphere is apocalyptic, alien, ominous, and cinematic, with dense red haze, strong contrast, visible grain, slight color bleed, and soft glowing edges.
• film grain, faded color tones, low contrast, long shadows, artistic framing, almost cinematic feel, editorial photography, moody photography - analog black-and-white portrait of h4sr3t, her head tilted downward and to one side, eyes closed, lips painted a deep matte shade. She wears a fitted ribbed light beige turtleneck and a pinstriped high-waisted skirt, seated sideways on a stone rooftop ledge with her torso angled toward the camera and one shoulder dropped lower than the other. Her right hand is raised beside her temple, two fingers loosely holding a lit cigarette close to her fringe, thin curls of smoke drifting up and tangling into her hair, while her left arm reaches back and down, palm braced flat against the ledge behind her hip for support. Behind her sprawls a rooftop skyline of mansard buildings, and further back the dark, sharply defined silhouette of the Eiffel Tower cuts against the pale, flat sky. Soft overcast light wraps around her features while grain and subtle contrast deepen the folds of her sweater and the texture of her hair, lending the image a hushed, detached elegance, wholly analog and cinematic in tone.
• A high-contrast surrealist painting of a brilliant white desert under a cloudless, searing midday sun. The landscape is dominated by impossibly long, pitch-black shadows that possess their own life and form, constructing entire cities across the dunes. These shadow-cities feature towering skyscrapers, marketplaces teeming with shadowy figures, and swaying shadow trees that exist without physical trunks. Shadow birds fly through the empty air above the sand. In stark contrast, a few "real" objects - a lone wooden chair, a solitary stone staircase, and a single window frame - are scattered across the desert, yet they cast no shadows at all, appearing disconnected from the laws of light. The image emphasizes a sense of reality as negative space, where the absent world is overwhelmingly populated and the physical world is sparse. The aesthetic is minimalist and eerie, featuring sharp edges, deep blacks against blinding whites, and a sense of profound, haunting mystery.
• The central figure in the image is a classic, stern profile of a contemplative Sherlock Holmes, depicted from the waist up in a cloak and his famous hat, with a smoking pipe in hand. The image is built on the principle of double exposure, where the lower part of the detective's profile artistically includes a depiction of a Victorian-era London street at night, along which, in the haze and fog of the night, a darkening cab moves away from us. It's an atmospheric and somewhat eerie image, combining the smooth transition of the rising fog and haze of the London street with the smoke from Sherlock Holmes' pipe, depicted against the blurred background of his office. Literary dreams, the cinematic feel of the picture, and the powerful style of the best visualizations of the famous detective’s stories aim to create an incredibly stylish image.
• colossal red torii gate on a jagged black cliff, solitary warrior silhouetted centrally beneath the archway, stone lanterns flanking the path, massive black sphere dominating a vibrant crimson sky, dark rocky foreground terrain, minimalist two-tone palette of pure black and blood red, high-contrast graphic illustration
• an ukiyo-e woodblock print of traffic including a cybertruck on the golden gate bridge.
• Image that is super artsy and Japanese painting stylized. Image looks like it was painted on a piece of bamboo parchment. In the background there is a big red circle (resembling the sun) and black silhouette of Fiji mountains with some black silhouettes of sakura trees. In the foreground there are Japanese stylized wojaks dressed as samurai. They are looking at viewer pogging with their mouths open and they are pointing back at the mountain with outstretched arms. On the side of the image there are three japanese words written verically "最も洗練された", "最も印象深い" and "傑作"
• analog image, film grain, saturated color, high contrast long shadows, artistic framing, almost cinematic feel, This minimalist digital illustration captures a charming animal scene peeking over a solid ledge, shifting the mood to something cute, friendly, and filled with character. The black cat's silhouette remains a dark, soft-edged shadow with alert, pointed ears framing the center-left of the composition. However, its large, round eyes now glow with a vibrant gradient that incorporates deep violet and electric purple tones, narrowing slightly into a pair of thin, vertical black pupils that give the feline a slightly annoyed, grumpy expression. A tiny, luminous magenta nose sits beneath its gaze, adding a whimsical touch to its comical irritation. Sitting right beside the cat is an adorable, fluffy corgi silhouette that brings a burst of friendly warmth to the piece. The corgi's signature large, rounded ears perk up joyfully, and its head is tilted slightly with a soft, bright smile evident in its silhouette. A heavy, stippled noise texture coats both animals, giving their edges a fuzzy, airbrushed quality that feels like a cozy, vintage risograph print or a grainy pastel drawing, perfectly maintaining the signature retro art style while leaning into a playful dynamic. The background frames the duo with a smooth, horizontal gradient sky that evokes a serene, colorful sunset. The top half of the frame is a calm turquoise blue that blends seamlessly down into a radiant, warm coral pink, casting a soft aura behind the contrasting animal shapes. The lower foreground ledge mirrors these colorful tones, shaded in deep purple and violet hues that ground the artwork in a cohesive, neon-infused color palette, turning the scene into a delightful, story-filled interaction between a grumpy cat and a happy pup.
• A highly detailed, professional real-life photograph taken on a Sony A7R IV with a 50mm f/1.8 lens, showing a cozy, cluttered wooden desk. A tiny, adorable chibi dragon, hand-drawn directly onto the surface of a real paper notebook in a rough, smudged charcoal pencil sketch style, has physically come to life. The flat, 2D black-and-white charcoal dragon has crawled partially off the page, and from its open mouth bursts a stream of fire rendered in the same rough, expressive charcoal style - thick smudged lines, sketchy strokes, and messy charcoal dust swirling outward to form the flame. This charcoal-drawn flame billows toward a real, fluffy marshmallow on a toothpick held by a real human hand in the foreground. The warm, intense light from the flame casts dramatic, flickering orange highlights on the charcoal sketch lines and the textured paper, creating a stunning, cross-dimensional depth of field that blurs the background.
• Aerial top-down drone shot capturing the vibrant, dynamic interplay of nature's elements: a narrow, golden-orange sandy beach sharply bordered by dense, dark green coastal vegetation on the left and churning turquoise ocean water with frothy white waves crashing against the shore on the right; the scene is rendered in hyperrealistic photorealism with high detail, sharp focus, and vivid color saturation - emphasizing the stark contrast between the warm earth tones of the sand and the cool, textured blues and greens of the sea, with visible underwater coral or reef structures scattered throughout the shallow waters, under bright, direct daylight with minimal shadows, conveying an intense, immersive, and visually arresting natural landscape devoid of human presence.,
• [Medium and Style] Surreal mixed-media digital collage with a dark cinematic poster aesthetic, vertical square-cropped composition, grainy analog texture, scratched film overlay, high-contrast experimental lighting. The image combines photographic elements with glowing abstract light effects, creating a dreamlike, symbolic, melancholic visual style. [Subject(s) and Clothing] A large glowing human silhouette occupies the left-center of the image, rendered as an overexposed white figure with a vivid electric-blue aura. The silhouette appears faceless and featureless, shaped like a person walking leftward, with a rounded head, torso, arm, loose pants, and bright shoes. The body is blown out into pure white with soft cyan and pale yellow diffusion inside the glow. On the right side, a second human figure stands partly hidden behind a tall vertical blue textured panel or doorway. This figure wears a reddish-brown oversized jacket or coat, loose white pants, and light-colored sneakers. The person has medium-length hair and is shown in profile or partial side view, with the upper body leaning or positioned behind the panel. Their face is partially obscured by shadow and the vertical blue structure. [Pose and Expression] The glowing white figure appears to be walking toward the left, one leg extended forward and one arm hanging downward, but the form is blurred and abstract. The right-side figure stands still behind the vertical panel, with one leg visible and bent slightly, creating a paused, hidden, or transitional posture. The expression is not clearly visible due to shadow, distance, and partial concealment. [Background and Environment] The background is almost entirely black, with a rough, worn texture resembling scratched photographic film or a distressed dark wall. Thin pale scratch marks, dust specks, and vertical streaks are scattered across the image, especially along the right edge and upper area. A rectangular blue light field stretches behind the glowing figure, fading into soft cyan and white near the center. The tall vertical blue panel on the right divides the glowing silhouette from the partially hidden figure, functioning like a doorway, screen, or barrier. [Lighting, Atmosphere and Colors] Low-key lighting dominated by black shadows and a powerful white-blue glow. The glowing figure emits intense white light with saturated cobalt-blue edges, creating a halo that spills across the center of the image. The right figure is dimly lit, with muted reddish-brown, off-white, gray, and beige tones. The palette is mostly black, electric blue, white, cyan, reddish brown, and muted gray. The mood is introspective, lonely, mysterious, and symbolic, suggesting separation, transformation, or leaving behind a former self. [Text and Typography] Small italic lowercase text near the top center reads "goodbye to the person I no longer am". The text is bright cyan-blue, slightly glowing, and set in a clean sans-serif italic style.
• surreal composite photograph, a giant dark skinned man in an oversized charcoal double-breasted suit and deeply pleated wide-leg trousers striding as a towering figure over a miniature San Francisco cityscape, one worn brown suede derby shoe lifted mid-stride looming enormous in the foreground close to the camera, the tiny city spread below - recognizable San Francisco landmarks rendered in miniature including the Transamerica Pyramid, Salesforce Tower, Victorian painted ladies, and steep hilly streets with cable car tracks, tiny yellow cabs and cars dotting the roads below, black beanie, narrow dark sunglasses, woven black crossbody bag, black crow perched on shoulder wings slightly spread, figure looking coolly downward or straight ahead, bright vivid sunny blue sky, sharp photorealistic composite, wide angle lens with strong perspective distortion, cinematic scale and grandeur
• This is a digital magazine cover with a dark, surreal theme. The dominant color is a deep, fiery orange, creating a dramatic and intense background. At the center, a partially obscured female face is visible, with only her eyes and upper nose showing through a black, jagged, and textured pattern that resembles sharp, pointed teeth or spikes. The black pattern is positioned horizontally across the middle of the cover, partially covering the woman's face. Her eyes are wide and expressive, with a slightly anxious or intense look. The overall style is dark, provocative, and thought-provoking, with a strong emphasis on identity and perception.
• A striking, high-contrast graphic poster features the stark black silhouette of a cat's head peeking over a textured horizontal barrier. The composition is defined by its minimalist yet aggressive design, dominated by a raw, textured crimson red background. The cat's fur blends completely into the shadows, with only its pointed ears and piercing gaze defined against the backdrop. Its left eye glows with an intense, vivid yellow hue, while its right eye is cast in a darker, shadowed orange tone, creating a mesmerizing and slightly ominous focal point. The artwork heavily incorporates a gritty, low-fidelity analog texture with deep grain, fine scratches, and subtle print distortions. The red barrier in the foreground features heavy black ink distressing, mimicking a raw woodcut or screen-printed punk zine aesthetic. The clean, sharp geometry of the cat's silhouette clashing with the chaotic, stained textures of the background gives the overall piece a powerful, contemporary graphic design edge.
• A sophisticated concept art illustration of a contemporary architectural masterpiece that blends technical precision with atmospheric artistry. The image combines the clean lines of a blueprint with the expressive textures of a moody digital painting. It depicts a multi-tiered urban sanctuary characterized by sweeping curves and sharp, geometric planes. Fine charcoal lines and delicate ink sketches outline the structural framework, while soft watercolor washes of deep indigo and warm ochre define the shadows and highlights. The building features expansive glass facades that reflect a vibrant, hazy cityscape. Layers of transparency reveal the interior layout, showing floating walkways and lush indoor forests. The composition uses a vertical perspective, looking upward from a bustling plaza, making the structure feel monumental and towering. Golden light spills from the windows, illuminating the surrounding mist and creating a sense of tranquil sanctuary amidst a high-tech metropolis.

시스템

• Redrob Image

  1.8s

  이미지당

  오픈 웨이트 Z-Image Turbo 머지(Apache-2.0, 상업적 이용 가능). g7e.2xlarge(NVIDIA RTX PRO 6000 Blackwell Server Edition, 96GB)에서 모델 상주 이후, 큐 제외 이미지당 약 1.8초. GPT Image 2 블라인드 90쌍 44-16-30(동등성 미입증). 베이스 대비 피부·빛·텍스처가 낫고 사진·포트레이트·무드에 강함. 실측 가동률 기준 단가는 API 정가의 약 1/40(동일 조건 비교 아님).