All posts
Image Generation API

FIBO 2.0 Prompts: 12 Prompts to Get the Same Shot Twice

Twelve FIBO 2.0 prompts you can paste on each::labs, in plain English and Bria's structured JSON, for packshots, packaging, portraits, interiors, characters and edits. Plus the one-field, fixed-seed method we tested.

By Eftal Yurtseven · Nov 16, 2026 · 9 min read

FIBO 2.0 Prompts: 12 Prompts to Get the Same Shot Twice

A good image prompt that you can't repeat isn't much use. You get a lovely packshot on Monday, the client asks for the same frame in golden hour on Tuesday, and the model gives you a different mug on a different table. Most FIBO 2.0 prompts online ignore that problem: one long paragraph, run once, hope for the best.

FIBO 2.0 Image Generation rewards a different habit. Bria trained the model on structured captions, so it reads a scene as named parts: lighting, composition, palette, objects. Write your prompt in those parts, fix the seed, and you can change one thing while the rest of the picture stays put. Below are twelve copy-ready prompts, including edits for FIBO 2.0 Image Editing, and the three test runs we did on each::labs.

For the model overview, read our FIBO 2.0 spotlight. This post is only about what to type.

How FIBO 2.0 Prompts Work on each::labs

Bria's model card describes FIBO 2.0 as an 8B model that does text-to-image, instruction edits, masked edits and multi-reference edits in one set of weights. Bria's pipeline expands a short prompt into a structured one in its Visual GenAI Language (VGL), then renders it. The card's pitch is blunt: tweak a single attribute without breaking the scene, and iterate without prompt drift.

On each::labs the model comes through two doors. Generation takes a prompt of up to 4,000 characters, nine aspect ratios from 1:1 to 16:9, 1MP PNG or JPEG output, an optional seed, and either one reference image or the photoreal preset (not both). The editing endpoint takes one to four images, a plain instruction or a structured_instruction JSON string, an optional mask and its own seed.

The generation endpoint has no separate JSON field, but you can paste a VGL object as the prompt text. We did exactly that in our tests, and the images followed the fields closely. So the prompts below come as plain English for exploring and as JSON for the shot you'll need again.

Name every part and nothing gets lost.
Name every part and nothing gets lost.

The VGL Fields Worth Learning First

Bria's VGL documentation lists more fields than one image needs. Learn eight.

short_description is the one-line anchor: what the picture is. background_setting is the place. style_medium picks the medium (photography, digital illustration, 3D render, watercolor, anime and a few more). context says what the image is for, such as a catalogue.

lighting is an object with conditions, direction and shadows. aesthetics holds composition, mood_atmosphere and a color_palette written as RGB triplets, like [196, 110, 74]. photographic_characteristics covers depth_of_field, focus, camera_angle and lens_focal_length. objects is an array of items with a description, shape_and_color, texture, orientation and, for people, age, expression, clothing and action. Typography goes in a separate text_render array with text, color and font.

The rule that makes all of this useful: one decision, one field. If the light is described in lighting and again in mood_atmosphere, you have two places to break when you change it.

Two frames, one scene. The seed keeps the promise.
Two frames, one scene. The seed keeps the promise.

Same Shot Twice: Change One Field, Keep the Seed

We ran prompt one on FIBO 2.0 Image Generation with seed 4242 at 4:5. It's the template for the rest of the post.

{
  "short_description": "A matte terracotta ceramic coffee mug on a pale oak table, steam rising",
  "background_setting": "Minimal kitchen counter, soft out-of-focus plaster wall",
  "style_medium": "photography",
  "context": "Product shot for a homeware catalogue",
  "lighting": {
    "conditions": "soft, diffused morning daylight",
    "direction": "from a window on the left",
    "shadows": "soft, short shadows falling to the right"
  },
  "aesthetics": {
    "composition": "centered, mug fills the lower third, generous negative space above",
    "mood_atmosphere": "calm, quiet",
    "color_palette": [[196, 110, 74], [222, 205, 180], [245, 240, 232]]
  },
  "photographic_characteristics": {
    "depth_of_field": "shallow",
    "focus": "sharp on the mug rim",
    "camera_angle": "eye level, slightly above",
    "lens_focal_length": "85mm"
  },
  "objects": [{
    "description": "ceramic coffee mug",
    "shape_and_color": "rounded cylinder, terracotta orange",
    "texture": "matte",
    "relative_distance": "foreground",
    "orientation": "handle facing right"
  }]
}

The result matched the brief almost line by line: terracotta mug, handle right, black coffee with a thin wisp of steam, oak table, pale plaster wall, lots of empty space above. Then we swapped only the lighting object for this, same seed, everything else untouched:

"lighting": {
  "conditions": "warm golden-hour sunlight",
  "direction": "low, from the left",
  "shadows": "long, crisp shadows stretching to the right"
}

The second frame is clearly the same shot in different light. Same mug, same handle side, same framing, with a long hard shadow running right. It wasn't pixel-identical, though. The mug got a touch rounder, the table edge moved up, and the wall picked up darker texture with a hint of kitchen at the edge. With a fixed seed and one changed field, composition and subject hold while the details around them breathe. For a true lock on the product, take the approved frame to the editor instead (our third test, below).

FIBO 2.0 Prompts for Product Shots and Packaging

Product work is where structure pays, because clients ask for variants. Prompt two explores a skincare packshot:

A frosted glass serum bottle with a brushed gold dropper cap standing on a wet slate slab, a few water droplets on the glass, soft overhead light, dark charcoal background, centered, macro product photography.

Prompt three is the same idea as a structured brief you can reuse across a range. Change shape_and_color per product and the set stays consistent:

{
  "short_description": "A frosted glass serum bottle with a gold dropper cap on a wet slate slab",
  "background_setting": "Seamless dark charcoal studio backdrop",
  "style_medium": "photography",
  "context": "E-commerce hero image for a skincare range",
  "lighting": {"conditions": "soft softbox light", "direction": "from above, slightly behind", "shadows": "tight contact shadow"},
  "aesthetics": {"composition": "centered, bottle fills two thirds of the height", "mood_atmosphere": "clean, cool", "color_palette": [[40, 42, 46], [214, 186, 120], [232, 236, 238]]},
  "photographic_characteristics": {"depth_of_field": "medium", "lens_focal_length": "100mm macro"}
}

Prompt four adds type. Text lives in its own text_render array, so you can change the copy without touching the box. Run it at 3:4:

{
  "short_description": "A kraft paper tea box with a minimalist printed label on a linen tablecloth",
  "background_setting": "Bright Scandinavian kitchen, blurred",
  "style_medium": "photography",
  "context": "Packaging mockup for a tea brand",
  "lighting": {"conditions": "bright, even daylight", "direction": "from the right", "shadows": "soft"},
  "aesthetics": {"composition": "box angled three quarters, lower center", "mood_atmosphere": "fresh, honest", "color_palette": [[184, 150, 108], [250, 248, 242], [74, 98, 72]]},
  "text_render": [
    {"text": "MORNING LEAF", "color": "[74, 98, 72]", "font": "bold geometric sans serif", "appearance_details": "printed on the front label, centered"}
  ]
}

Keep label copy short and check the spelling in every frame.

Light first. The room follows.
Light first. The room follows.

Editorial and Interior Prompts

Editorial is the opposite problem: you want mood, not a spec sheet. Prompt five is plain English and deliberately loose:

Editorial portrait of a woman in her sixties with silver cropped hair, wearing a rust wool coat, standing in a windy coastal car park at dusk, looking past the camera with a half smile, overcast light, muted film colors, 50mm.

Prompt six fills in the person fields. Keep this version once the art director signs off, so you can later change clothing or expression alone:

{
  "short_description": "Editorial portrait of an older woman on a windy coast at dusk",
  "background_setting": "Empty coastal car park, grey sea behind",
  "style_medium": "photography",
  "lighting": {"conditions": "overcast dusk", "direction": "diffuse, from the sea", "shadows": "very soft"},
  "aesthetics": {"composition": "subject on the left third", "mood_atmosphere": "wistful", "color_palette": [[150, 78, 48], [128, 136, 140], [210, 206, 198]]},
  "photographic_characteristics": {"depth_of_field": "shallow", "focus": "eyes", "lens_focal_length": "50mm"},
  "objects": [{
    "description": "woman with silver cropped hair",
    "age": 63,
    "expression": "half smile, looking past the camera",
    "clothing": "rust wool coat, grey scarf",
    "action": "standing still, wind in her hair"
  }]
}

Prompt seven is an interior. Give the room one light source and one hero object, or it gets furnished like a showroom:

Wide interior photo of a small Japandi living room, a low walnut sofa as the only large piece, a paper floor lamp switched on, late afternoon light through sheer curtains on the left, pale clay walls, wide angle 24mm, calm and uncluttered.

Run it at 16:9. Once you like the room, move it into JSON and treat lighting as your time-of-day dial. Morning, noon and blue hour from one seed beats three unrelated rooms.

Keep a Character Consistent Across Shots

Character consistency is a two-step job with FIBO 2.0: generate the character once, then do everything else as an edit. Prompt eight builds the character sheet. Set a seed you'll keep:

{
  "short_description": "Full-body character of a young courier on a cargo bike",
  "background_setting": "Plain light grey background",
  "style_medium": "digital illustration",
  "artistic_style": "clean flat shading, bold outlines",
  "lighting": {"conditions": "even light", "direction": "front", "shadows": "minimal"},
  "aesthetics": {"composition": "full body, side view", "mood_atmosphere": "upbeat", "color_palette": [[230, 92, 52], [40, 60, 92], [244, 240, 230]]},
  "objects": [{
    "description": "courier with a short black bob and round glasses",
    "clothing": "orange rain jacket, navy cargo trousers, yellow helmet",
    "action": "riding a green cargo bike"
  }]
}

Then hand that image to FIBO 2.0 Image Editing with a new scene. Bria's card says references are read in order as image 1, image 2 and so on. Prompt nine puts the courier into a street you supply as the second image:

Place the courier from <image_1> riding through the rainy street in <image_2>, seen from the front. Keep the face, glasses, jacket, helmet and bike exactly as in <image_1>. Match the street's evening light.

Up to four images fit in one request: character, location, a prop and a light reference.

Change one line. Keep the plate.
Change one line. Keep the plate.

FIBO 2.0 Edit Prompts That Change One Thing

Edit prompts fail when they describe the whole new picture. Say what changes, then say what stays. Our third test sent the morning mug to the editor, seed 4242, with prompt ten:

Change the mug glaze to a deep sage green. Keep the mug shape, the coffee, the steam, the table, the wall and the lighting exactly as they are.

It did precisely that. Same mug silhouette, same handle, same wisp of steam, same wall cracks. The only drift we spotted was a slightly warmer cast on the oak. That's our recommended colorway workflow: generate once, approve, then edit every variant from the approved file.

Prompt eleven is a background swap. Paint a black-and-white mask the same size as the photo: on each::labs black areas are preserved and white areas are edited, and a mask only works with exactly one input image. Paint the backdrop white:

Replace the background with a sunlit terracotta tiled terrace with a blurred olive tree. Match the shadow direction to the light already on the product.

Prompt twelve is a relight:

Relight the scene as a cool blue-hour exterior, with a warm practical light glowing from the lamp on the right. Keep every object, its position and its color unchanged.

When an edit like this has to run across a hundred catalogue images, switch from instruction to structured_instruction. The schema describes it as the field for reproducible programmatic edits, passed as a JSON string. Pair it with a fixed seed. Our Flux 3 Edit prompts apply the same principle on another model, and our mask editing guide compares mask conventions.

Habits That Make FIBO 2.0 Prompts Repeatable

Write the plain version first. Explore with short prompts and no seed, because you want variety while you're looking. When something works, translate it into VGL fields and pin the seed.

Change one field per run. If you change lighting and composition together and the result is worse, you have no idea which one did it.

Keep palettes in numbers. RGB triplets in color_palette stop "warm beige" from meaning five different beiges across a campaign.

Choose the aspect ratio before you approve anything. On the editor the output follows the first input's ratio unless you set one, and recomposing later breaks the layout you signed off.

Finally, store the trio with the file: prompt, seed, aspect ratio. Run the same chain every week and it belongs in an each::labs flow, not a notes app. And when a brief needs a different model, the text-to-image catalogue and the full Bria lineup sit behind the same API.

Frequently Asked Questions

Do FIBO 2.0 prompts have to be written in JSON?

Plain English works. The generation endpoint on each::labs takes a text prompt of up to 4,000 characters, and Bria's own pipeline expands short prompts into structured ones. JSON pays off once you need to repeat a shot or change one attribute.

How do I get the same image twice with FIBO 2.0?

Pin three things: the prompt, the seed and the aspect ratio. The schema describes the seed as deterministic, so that trio is your recipe. Change one field with the seed held and the subject and framing stay while small details shift. For an exact lock on the subject, edit the approved image instead of regenerating.

How many reference images can FIBO 2.0 editing use?

Up to four on each::labs. Refer to them in order as image 1, image 2 and so on in your instruction. Masks are the exception: a mask is only valid when you send exactly one image.

Start with what you need.

Scale when you’re ready.