Pool A · Method A · Forward Redescribe

📅 2026-04-11 🐊 Crocodile pool · 5 frames 💰 $0.737 🔄 6 регенов всего ✅ multi-ref [prev + frame_01] = best approach
← Все отчёты
5/5
Кадров принято
11
Total generations
$0.737
Total cost
$0.147
Per accepted frame
~35m
Total time
★★★★★
Continuity (final)

Continuity check — финальные 5 кадров

Последняя версия каждого кадра. Все scene anchors (дом, забор, газон, кусты, форма крокодила, угол камеры) сохранены через всю последовательность благодаря multi-ref подходу: каждый кадр генерится с двумя референсами — предыдущий accepted кадр + frame_01 (canonical anchor).

frame 1
1. Outline
frame 2
2. Excavation
frame 3
3. Concrete shell
frame 4
4. White plaster
frame 5
5. Filled + LEDs

Что такое Forward Redescribe + Multi-ref

Стратегия: frame 1 генерится как T2I (без референсов). Frames 2-5 — I2I с двумя референсами:

1️⃣ Primary: предыдущий accepted кадр (контекст текущего состояния стройки)
2️⃣ Anchor: frame_01.png (canonical scene + crocodile shape constraint)

Каждый промпт полностью переписывает всю сцену с нуля (дом, забор, газон, угол камеры). Источник: AI Maskman + Grow with Dani + Salmaan Mohamed (3/15 гайдов сходятся на full-redescribe). Multi-ref с anchor — наша адаптация после первого review.

Из официальной доки Gemini: модель блендит до 14 reference images как unified context — нет «primary», все равны. Передача frame_01 как второго референса работает как shape-constraint, не давая форме крокодила «плыть» во время excavation phase.

Модель: gemini-3.1-flash-image-preview (Nano Banana 3.1) — $0.067 per image.

Frame-by-frame breakdown с галереей всех версий

frame 1

01 Empty + crocodile outline

Канонический кадр. T2I (без референсов) — устанавливает scene anchors для всех последующих кадров.
T2I (0 refs) 1511 chars ✅ accepted v1 $0.067

Aerial 70° drone shot, белый двухэтажный дом сверху, кедровый забор с двух сторон, газон с диагональными mowing stripes, силуэт крокодила в белой краске чётко виден. Все якоря сцены установлены чисто. 0 регенов.

Полный prompt (1511 chars)
Static overhead aerial drone shot, fixed camera position pointing straight down at a 75-degree tilt, no zoom, no pan, no tilt change, no rotation. A modern American suburban backyard photographed from approximately 15 meters above. The yard is enclosed by a tall horizontal-plank cedar wood privacy fence running along the back edge and the right side of the frame, painted a natural warm cedar tone. A two-story white clapboard suburban house with a dark gray shingled roof and white-trimmed windows is partially visible at the top edge of the frame, behind the back fence. The ground is a perfectly maintained emerald green Bermuda grass lawn, freshly mowed in clean diagonal stripes. In the center of the lawn, a large outline of a giant crocodile-shaped swimming pool has been carefully sprayed onto the grass using bright white marking paint. The crocodile shape is clearly recognizable: a long pointed snout aimed toward the house at the top of the frame, two front legs and two back legs extending outward, a thick body, and a long tapering tail pointing toward the bottom of the frame. Six small wooden surveyor stakes with orange flagging tape mark the corners and joints of the outline. Bright sunny late-morning lighting from the upper right, casting soft sharp shadows. Photorealistic, hyper-realistic, shot on a DJI Mavic 3 Pro drone, sharp focus throughout, natural color grading, 9:16 vertical orientation, no people visible, no captions, no text overlays, no watermark, no CGI, no cartoon style.
frame 2

02 Excavation in progress

Yellow CAT excavator + Bobcat skid-steer mid-dig. ~50% формы выкопано.
I2I (1 ref: 01.png) 1821 chars ✅ accepted v1 $0.067

Все anchors сохранены. CAT excavator в теле, Bobcat у головы, mounds свежей земли, tire tracks. 0 регенов.

Полный prompt (1821 chars)
Static overhead aerial drone shot, fixed camera position pointing down at a 70-degree tilt, no zoom, no pan, no rotation. The exact same modern American suburban backyard as before, photographed from approximately 15 meters above. The yard is enclosed by the same tall horizontal-plank cedar wood privacy fence running along the back edge and the right side of the frame, painted natural warm cedar tone. The same two-story white clapboard suburban house with dark gray shingled roof and white-trimmed windows is visible at the top of the frame behind the fence. The same emerald green Bermuda grass lawn with diagonal mowing stripes surrounds the work area. The same row of low landscaping shrubs runs along the right edge inside the fence. Inside the previously sprayed white outline of the giant crocodile-shaped swimming pool, excavation work is now actively underway: a yellow CAT 320 mini excavator with its boom arm extended is positioned in the middle of the body section, mid-swing as it digs out the brown earth. A small Bobcat skid-steer loader sits near the snout area. Approximately half of the crocodile shape has been dug down to about one meter depth, exposing raw brown soil and dark wet clay at the bottom of the pit. Large mounds of fresh dark soil are piled around the perimeter of the excavation, partially covering the white outline. Two parallel tracks of dirt cross the lawn from the gate area where the equipment was driven in. A faint cloud of dust hangs in the air over the dig site. Same bright sunny late-morning lighting from the upper right, casting the same soft sharp shadows. Photorealistic, hyper-realistic, shot on a DJI Mavic 3 Pro drone, sharp focus throughout, natural color grading, 9:16 vertical orientation, no people visible, no captions, no text overlays, no watermark, no CGI.
frame 3

03 Concrete shell + rebar

Полностью выкопанная форма, gunite shell, видимая grid-структура. Excavator припаркован у забора. Финальная версия — multi-ref.
multi-ref: 02 + 01 1948 chars ✅ accepted v2 (multi-ref) $0.134 (2 calls)

✅ Sequence из регенов — multi-ref решил drift

  • v1 (single-ref): ⚠️ форма крокодила слегка «плыла» — голова и лапы менее чёткие чем в frame 1, surveyor stakes исчезли. Принято с минорным замечанием.
  • v2 (multi-ref [02, 01]): ✅ форма крокодила восстановлена, surveyor stakes снизу вернулись (модель подхватила из frame 1!). Это новый baseline.
Полный prompt (1948 chars)
Static overhead aerial drone shot, fixed camera position pointing down at a 70-degree tilt, no zoom, no pan, no rotation. The exact same modern American suburban backyard as the previous frame, photographed from approximately 15 meters above. The yard is enclosed by the same tall horizontal-plank cedar wood privacy fence running along the back edge and the right side of the frame. The same two-story white clapboard suburban house with dark gray shingled roof and white-trimmed windows is visible at the top of the frame behind the fence. The same emerald green Bermuda grass lawn with diagonal mowing stripes surrounds the work area, although now scuffed and patchy near the dig site. The same row of low landscaping shrubs runs along the right edge inside the fence. In the center of the lawn, the giant crocodile-shaped pool is now fully excavated to a uniform depth of about two meters. The walls and floor of the crocodile-shaped pit are now lined with sprayed gray concrete (gunite shotcrete), forming the rough but structurally complete shell of the pool. A grid of vertical and horizontal steel rebar bars is clearly visible reinforcing the walls, especially around the curves of the head, the legs, and the tail. Wooden formwork planks line the upper rim of the pool, holding the gunite in place during curing. Stacks of unused rebar bundles and a small pile of cement bags are placed on the grass near the snout. The yellow CAT excavator and Bobcat skid-steer have been moved off to the side near the right fence, parked. The mounds of dirt around the perimeter have been mostly cleared away, leaving only thin remnants of soil along the edge. Bright sunny midday lighting from nearly directly above, casting short shadows. Photorealistic, hyper-realistic, shot on a DJI Mavic 3 Pro drone, sharp focus throughout, natural color grading, 9:16 vertical orientation, no people visible, no captions, no text overlays, no watermark, no CGI.
frame 4

04 White plaster (dry, finishing stage)

Dry pool basin с white plaster, видна глубина с тенью, aluminum ladder, plaster bags + tools на газоне.
multi-ref: 03 + 01 2409 chars ✅ accepted v4 (multi-ref) $0.268 (4 calls) 3 регена всего

⚠️ История регенов — критический урок про негативные промпты

  • v1 (1841 chars, описание «aquamarine blue tiles»): ❌ модель налила воды — конфузила «aquamarine blue tiles» с водой. (потерян на диске)
  • v2 (2293 chars, добавил CAPS «NO WATER, NO BLUE, NOT FILLED» × 5): ❌ снова налила воды. Негативные промпты на Nano Banana НЕ работают. (потерян на диске)
  • v3 (2409 chars, убрал ВСЕ упоминания «blue/tile/water», описал позитивно «white plaster + aluminum ladder + dust + shadow», single-ref): ✅ работает но форма чуть менее чёткая.
  • v4 (тот же prompt 2409 chars, multi-ref [03 + 01]): ✅ лучшая версия — форма ещё острее, plaster bags/tools видны, ladder в теле. Multi-ref снял проблему dry-state с первой попытки потому что frame 1 тоже не имеет воды.

🔑 Урок: описывай цель, а не «чего нет» (модель видит «BLUE»/«WATER» и использует их). + Multi-ref с frame_01 anchor усиливает эффект если canonical кадр уже не имеет конфликтующих элементов.

Полный prompt v3/v4 (2409 chars)
Static overhead aerial drone shot, fixed camera position pointing down at a 70-degree tilt, no zoom, no pan, no rotation. The exact same modern American suburban backyard as the previous frames, photographed from approximately 15 meters above. The yard is enclosed by the same tall horizontal-plank cedar wood privacy fence running along the back edge and the right side of the frame. The same two-story white clapboard suburban house with dark gray shingled roof and white-trimmed windows is visible at the top of the frame behind the fence. The same emerald green Bermuda grass lawn surrounds the work area, freshly resodded and pristine where the construction equipment used to be. The same row of low landscaping shrubs runs along the right edge inside the fence. In the center of the lawn, the giant crocodile-shaped excavated pit is in the white plaster finishing stage. The interior walls and floor of the crocodile-shaped concrete pit have just been resurfaced with smooth white pool plaster, a uniform off-white cream color, slightly textured. The plaster surface is dusty and chalky, freshly applied, completely dry. A bright aluminum extension ladder leans against the inside wall at the body section, providing access to the bottom of the pit. The pit is two meters deep — you can see the matte white plaster floor at the bottom in shadow, the shadow cast by the high midday sun pooling along one inside wall. A white-handled long-bristle floor squeegee and a yellow plastic bucket sit on the dry plaster floor near the tail end of the basin. A continuous border of light cream-colored concrete coping stones has been installed around the entire rim of the pit, neatly outlining the crocodile shape with a clean stone edge approximately 30 cm wide. Several stacks of plaster bags and a white plastic mixing tub sit on the grass just outside the snout area. The lawn around the pit is perfectly restored — no dirt, no equipment, no machinery, no debris. The wooden surveyor stakes are gone. Bright sunny midday lighting from nearly directly above, casting short crisp shadows. The empty pit interior is half in sunlight, half in deep shadow, clearly showing the pit is hollow and dry. Photorealistic, hyper-realistic, shot on a DJI Mavic 3 Pro drone, sharp focus throughout, natural color grading, 9:16 vertical orientation, no people visible, no captions, no text overlays, no watermark, no CGI.
frame 5

05 Filled + twilight + LEDs (HERO SHOT)

Twilight night, окна дома горят, garden lights, бассейн доминирует с электрическим cyan glow.
multi-ref: 04 + 01 2577 chars ✅ accepted v3 (multi-ref) $0.201 (3 calls) 2 регена всего

⚠️ История регенов — lighting transition

  • v1 (2341 chars, twilight упомянут в середине промпта, после описания сцены): ❌ модель сохранила midday lighting из ref-кадра 04. Бассейн заполнен водой, но небо дневное. (потерян на диске)
  • v2 (2577 chars, «EVENING TWILIGHT BLUE-HOUR SCENE» вынесено в самую первую строку, single-ref): ✅ перебило baseline lighting. Twilight сцена с LED glow.
  • v3 (тот же prompt, multi-ref [04 + 01]): ✅ лучшая версия — форма крокодила более solid, композиция драматичнее (пул заполняет больше кадра), 4 окна дома горят (vs 2 в v2), garden lights чётче. Multi-ref не сломал twilight потому что 04 уже dry-state не противоречит ночи.

🔑 Урок: при смене lighting/времени суток новое состояние нужно описывать в самом начале промпта. + Multi-ref с frame_01 НЕ ломает lighting transition если предыдущий кадр уже задаёт нужный контекст.

Полный prompt v2/v3 (2577 chars)
EVENING TWILIGHT BLUE-HOUR SCENE. The sky is dark cobalt blue overhead, gradient down to a warm burnt-orange and dusty pink near the horizon behind the house. Long deep blue evening shadows stretch diagonally across the entire lawn from the back fence and the house. The ambient light is dim and cool. It is the moment just after sunset. Static overhead aerial drone shot, fixed camera position pointing down at a 70-degree tilt, no zoom, no pan, no rotation. The exact same modern American suburban backyard as the previous frames, photographed from approximately 15 meters above. The yard is enclosed by the same tall horizontal-plank cedar wood privacy fence running along the back edge and the right side of the frame. The same two-story white clapboard suburban house with dark gray shingled roof is visible at the top of the frame behind the fence — and now several windows of the house are glowing brightly with warm yellow interior light: the kitchen window on the ground floor and one upstairs bedroom window are clearly lit, casting warm yellow rectangles into the dim evening yard. The same Bermuda grass lawn surrounds the work area, now appearing dark teal-green in the dim twilight, with the diagonal mowing stripes still faintly visible. The same row of low landscaping shrubs along the right edge inside the fence is now lit by a string of small warm-white garden lights placed at the base of each shrub, each light glowing as a small bright point in the dim scene. In the center of the dark lawn, the giant crocodile-shaped swimming pool is the dominant brightest object in the frame. The pool is brim-full with crystal-clear water, and a powerful network of underwater LED lights set into the walls and floor casts a vivid bright electric-cyan and aquamarine glow upward through the water. The entire crocodile shape glows brilliantly in the dark scene, with the LED light spilling out over the cream-colored coping stones and onto the surrounding dark lawn in a wide blue halo. The water surface is glassy still, faintly mirroring the deep cobalt twilight sky. The bright glow of the pool stands in vivid contrast to the surrounding dim evening lawn. Photorealistic, hyper-realistic, shot on a DJI Mavic 3 Pro drone with long exposure for evening lighting, cinematic blue-hour color grading, sharp focus throughout, 9:16 vertical orientation, no people visible, no captions, no text overlays, no watermark, no CGI. The dominant color palette of this frame is cool blue and cyan, with warm yellow accents only from the house windows and tiny garden lights.

📚 Финальные уроки — как минимизировать регены

Из 11 generations на 5 кадров: 3 ушли с первой попытки (frames 1, 2 + некоторые multi-ref повторы), 8 потребовали регенов (включая single→multi-ref upgrade). Multi-ref подход полностью устранил регены на финальной итерации.

Правило 1 — Multi-ref [previous + frame_01] по дефолту

refs=[previous_frame.png]
refs=[previous_frame.png, frame_01.png] — frame_01 как canonical anchor

Frame 1 (canonical scene с outline формы) работает как shape-constraint, не давая форме «плыть» во время excavation/concrete phases. Bonus: если frame_01 не имеет конфликтующих элементов (воды, ночи), он косвенно помогает с состоянием.

Правило 2 — Никогда не используй негативные промпты для целевого состояния

"empty pool, NO water, no liquid, not filled"
"dry concrete basin with aluminum ladder leaning inside, dust on the floor, deep shadow cast inside the empty pit"

Модель видит токены «water», «filled» в тексте и использует их как присутствующие концепты, даже после «NO». Описывай что должно быть нарисовано.

Правило 3 — При смене lighting/state — выноси новое состояние в самое начало промпта

"...same backyard...<200 слов сцены>... and the lighting is twilight"
"EVENING TWILIGHT BLUE-HOUR SCENE. The sky is cobalt blue... ...<потом сцена>"

Правило 4 — Используй конкретные объекты-маркеры состояния

Правило 5 — Сохраняй ВСЕ версии регенов с суффиксами

Naming convention: 03_v1_singleref.png, 03_v2_multiref.png, финальная копируется в 03.png. Тогда галерея в отчёте показывает полную историю. В этой итерации первые регены frames 4 и 5 потеряны — они помечены как «потерян» в галерее.

Правило 6 — Smooth state-смены работают с первой попытки

Frames 1, 2 прошли с первой попытки потому что переходы между ними не конфликтовали. Регены нужны только когда новое состояние радикально отличается от ref:

📊 Прогноз для следующих storyboard'ов с multi-ref + правилами 2-4

Total estimate для оставшихся 5 storyboard'ов: ~$2.00. Cumulative: ~$2.74.