A prompt that packs seven life stages into narrow vertical strips with torn deckled edges, four different color temperatures, and a 2-pixel softness instruction creates predictable failures: per-panel scale drift, edge texture loss, and white-balance averaging toward beige.
A photorealistic cinematic composite divided into six vertical strips with torn deckled paper edges separating each panel, illustrating the complete arc of female life from conception through elder years. Leftmost panel: grainy monochrome fetal ultrasound with ethereal blue-white glow and visible amniotic fluid texture. Second panel: sleeping newborn swaddled in cream linen, soft north-facing window light at 5600K, visible skin translucency. Third panel: curious toddler in rust-colored corduroy overalls taking first steps, vintage sepia photographs floating in shallow depth of field around her. Fourth panel: school-age girl with pink canvas backpack and charcoal grey pinafore dress, brass pocket watch gears spinning behind her with motion blur marking time's passage. Fifth panel: young woman in mustard merino sweater and indigo raw denim, confident weight-shifted posture, golden hour light at 3200K streaming through venetian blinds creating stripe shadows. Sixth panel: professional woman in tailored black sleeveless sheath dress, arms crossed, determined expression, city skyline reflection in window behind her at twilight blue hour 7500K. Final panel: elegant senior woman with silver hair in cream cashmere turtleneck, seated by sunlit window at 6:47pm warm 2700K light, peaceful gaze toward horizon, visible hand veins and skin texture. Scattered antique photographs with curled edges throughout all panels, subtle 35mm film grain, deliberate warm-to-cool color temperature progression left to right, volumetric god rays penetrating dust particles, shot on Arri Alexa 65 with Cooke S7i 75mm and 100mm lenses at T2.8, shallow depth of field transitions between strips with 2-pixel edge softness --ar 1:1 --style raw --s 750 --q 2
Seven vertical strips with torn deckled edges inside a 1:1 square leave each panel about 14% of the frame width. The prompt asks each of those narrow columns to hold a full figure, a distinct light temperature, and a specific material — cream linen, rust corduroy, charcoal pinafore, mustard merino, indigo raw denim, black sheath dress, cream cashmere — plus floating sepia photographs, brass watch gears, and venetian-blind shadows. That density is where the structure breaks.
The exact model and version that generated the reference are unknown. The tail flags --ar 1:1 --style raw --s 750 --q 2 are Midjourney-specific. They set aspect ratio, raw style, stylization and quality on that platform and do not transfer as universal commands. An image-inspired prompt is an approximation; it cannot guarantee an identical reconstruction elsewhere.
Scale drift when every strip gets its own subject
When a prompt lists a newborn, a toddler, a school-age girl, a young woman and a senior woman each in their own vertical column, the model tends to frame each figure to fill its local column height. The infant swaddled in cream linen expands to fill the strip. The woman in the black sleeveless sheath does the same. Without a shared constraint, the toddler can read larger than the adult.
This happens because attention resolves per-panel framing before it resolves cross-panel proportion. A possible untested approach is to declare a single proportion rule before the panel descriptions, for example that all figures share a 1:6 head-to-frame ratio with eye-line near 55% of panel height. That anchors size to the figure itself rather than the column. If the toddler still dominates, reduce the wording that emphasizes her action — “taking first steps” often inflates scale — or remove one panel to give each figure more horizontal pixels.
Torn edges versus depth of field
Deckled paper needs a hard, fibrous boundary with a very narrow transition. The prompt asks for “shallow depth of field transitions between strips with 2-pixel edge softness.” At this strip width, even a few pixels of softness consume the visible fiber texture and the tears dissolve into a gradient. Worse, the model often reads that phrase as a global depth-of-field instruction and softens the subjects along with the borders.
Keep the two ideas separate in the text. Describe the paper first — torn deckled edges, sharp inter-panel boundary, narrow 2-pixel feather only at the paper tear — and then state that no depth-of-field blur should be applied to the panel content itself. The floating sepia photographs with curled edges and the brass pocket watch gears already carry their own shallow depth of field; they do not need an extra global blur. If the edges still bleed, shorten the softness phrase or replace it with “hard deckled tear, no gradient between panels.”
Why the color progression averages to beige
The prompt specifies 5600K north-facing window light, 3200K golden hour through venetian blinds, 7500K twilight blue hour with city reflection, and 2700K warm light at 6:47pm, plus a deliberate warm-to-cool progression left to right. Diffusion models carry a strong prior for a single white balance per image. Listing four to seven temperatures in one canvas pushes the model to average them toward neutral.
The twilight 7500K in the sixth panel is the first to disappear. It is not extreme enough to survive averaging. An untested suggestion is to anchor only three points — for example 5600K on the left, 3200K in the middle, 2700K on the right — and let the model interpolate, rather than naming every panel’s temperature. If you need that blue hour to hold, push it higher than 8000K or describe it visually instead of numerically: “deep blue twilight, cool window reflection, no warm correction.” If the whole composite still washes out, cut the temperature list and describe the light by material effect — stripe shadows from blinds, soft window wrap on skin translucency, warm rim on silver hair — rather than Kelvin numbers.
Unifying the collage and managing expectations
Order helps. Put shared constraints first: “single photograph of a torn-paper collage, consistent 35mm film grain across the full canvas, consistent exposure, volumetric god rays and dust particles across all panels.” Then list per-panel variations. Early tokens weight more heavily when instructions conflict.
The Arri Alexa 65 with Cooke S7i 75mm and 100mm at T2.8 in the prompt does not mean the model traces light through that glass. It is a style cue that tends to produce a photorealistic cinematic look with shallow depth of field, smooth falloff and gentle bokeh around the floating photographs and background gears. Describe the look you want if you move to another generator that does not recognize those names.
This is a simulation of a physical collage, not a photograph of one. At seven panels in a square, expect fine detail to soften: individual corduroy wales, hand veins and skin texture on the senior woman, and the fiber structure of the deckled edges all compete for very few pixels per strip. The suggestions above about shared head-to-frame ratios, isolated edge clauses, and three-point temperature anchors are untested structural ideas based on how models distribute attention, not guaranteed fixes. If eye-lines still diverge or the blue hour still neutralizes, the most reliable change is to use fewer panels — five instead of seven — which gives each strip more width and reduces both scale drift and edge contamination faster than adding more text.
Comments
Post a Comment