Skip to content

Global-AI-Group-Inc/content-generation-skills / skills/mira-image-prompting

mira-image-prompting

Verified

How to write stills for Mira AI's image models: GPT Image 2, 2.5 and 2.6, Nano Banana Pro and 2.1, Seedream 4, 4K and 5 Pro, Muse Image, Ideogram 4.5, Flux. Which model holds text, which keeps a product identical across a series, how the craft banks for light, optics, composition, grade and materials are used, and the words that make an image worse. Use it before any generate_image call on Mira, or when the user says "image prompt", "промпт для картинки", "нарисуй", "сгенерируй картинку", poster, packshot, banner, cover. NOT for video (mira-video-prompting).

FileKindSize
SKILL.mdskill md4.2 KB
references/composition.mdmarkdown2.3 KB
references/craft.mdmarkdown3.0 KB
references/grade.mdmarkdown2.3 KB
references/light.mdmarkdown3.7 KB
references/materiality.mdmarkdown8.0 KB
references/models.mdmarkdown7.0 KB
references/optics.mdmarkdown2.4 KB

references/models.md

View on GitHub
# Image models on Mira

Fourteen models, one pipeline. Every model costs three credits per image except Muse Image at one,
Seedream 5.5 at two, Seedream 5 Pro and Nano Banana 2.1 at four, Ideogram 4.5 at six and the premium tier at eight: Nano Banana Pro and GPT
Image 2.6 and 2.5 Sunburst.
Pick the pricier ones only when their strengths are needed.

## gpt-image-2

The strongest at holding a detailed commercial brief with many requirements at once: exact copy,
brand colours, camera angle, lighting, what must not change. Photorealistic products, materials
and skin. Preserves product geometry, packaging and logos. The right default for e-commerce
packshots and structured ad layouts. In this setup it returns 2:3 for every tall ratio, so a 9:16
or 4:5 that must be exact goes elsewhere.

## gpt-image-2.6-flare

The newest GPT image family at the fast three-credit tier, served straight from OpenAI. It plans
the frame before it renders, so it holds a dense brief, complex layouts, small type and the
relationships between objects better than 2.5. Everything 2.5 Flare does below holds here too:
exact ratio on a free-form canvas, legible text in any script including Cyrillic, faithful
multi-photo edits. The first pick for a text-bearing post, story or banner.

## gpt-image-2.6-sunburst

Premium. The newest family at its most detailed: planned composition, rich light, the tightest
control over edits, and it keeps one character, place or product identical across a series - Mira
draws its own characters, bloggers, locations and props on it. Honours the ratio exactly. Slower
than Flare, so not for drafts.

## gpt-image-2.5-flare

The previous GPT image generation at the fast three-credit tier, served straight from OpenAI. The same
brief-holding, product fidelity and multi-photo edits as gpt-image-2, plus legible in-image text in
any script including Cyrillic, and it renders the exact requested ratio on a free-form canvas: a
9:16 story really is 9:16. The first pick for a text-bearing post, story or banner, and for a
faithful edit of the user's own photo when the premium tier is not needed. Carries lettering from
a reference verbatim.

## gpt-image-2.5-sunburst

Premium, previous generation. The same family at its highest quality setting: sharper detail, richer light and shadow,
and the tightest control over edits, changing only what you name and keeping the rest of the
reference. Production-ready campaign creative, polished product imagery, an edit that must stay
faithful. Honours the ratio exactly. Slower than Flare, so not for drafts.

## nano-banana-pro

Legible in-image text in any language and the only model to trust with Cyrillic. Headlines,
prices, offers, packaging copy, posters, infographics, diagrams. One design that must exist in
several languages. Many reference photos blended while several people stay recognisable. Honours
tall ratios. Premium price.

## nano-banana-2.1

Google's newer, faster Nano Banana, four credits - half the premium price. Sharp, legible text
in any language including Cyrillic, infographics, diagrams and layouts with several text blocks;
put every word in double quotes with its place and look, and write prose rather than tags. The
only model with the panoramic ratios 21:9, 4:1 and 8:1 (wide banners, site headers, strips) and
1:4 and 1:8 (tall scrolls, skyscraper banners); describe a panorama along its long side - what
sits left, centre and right - so it reads as one continuous scene. Holds every ratio exactly.
Takes up to four reference photos. For the hardest multi-person blends and maximum polish,
nano-banana-pro is still the pick.

## seedream-4

Keeps one product or one character consistent across several reference photos. The model for a
series: the same SKU in six scenes, the same person in a campaign. Honours tall ratios. Renders
at 1728x3072 on 9:16, three times the pixels of gpt-image-2, so skin, fabric and food close-ups
belong here.

## seedream-4k

The same model at 4K, up to 4096 px. Only for print, large format or a hero visual where fine
detail is inspected. For a social post it is waste.

## seedream-5-pro

ByteDance's flagship, four credits. Photoreal at true 2K on the exact requested ratio: skin,
fabric, glass and metal read as photographed rather than rendered, and light stays motivated
across the frame. Keeps one product or one person identical across several references and
follows a long, structured brief. Premium product and lifestyle stills, close-ups where texture
is inspected. Keep on-image text to a short line in quotes. Served straight from ModelArk; if
ByteDance is saturated the same model is delivered through Pika.

## seedream-5.5

ByteDance's newest Seedream, two credits. Photographic realism at 2K: skin with real texture,
fabric, glass and metal, at half the price of Seedream 5 Pro. The most careful editor in the
lineup: on an attached photo it changes only what the brief names and leaves every other region
as it was, so one image can be refined step by step, each result fed back as the next reference.
Option `background: transparent` returns a PNG with a real alpha channel, for stickers, cut-outs
and design assets. Ratios are 1:1, 3:4, 2:3, 9:16, 4:3, 3:2, 16:9 and 21:9; it has no 4:5 or
5:4, and no 4K. Keep on-image text to a short line in quotes. Served through Pika.

## muse-image

Meta's model, one credit, the cheapest in the lineup. It reasons over the brief in several
passes and checks itself against live web and image search, so real places, landmarks,
products and current designs come out recognisable: name them plainly instead of describing
them from memory. Holds the requested ratio and sizes the canvas itself. Good for a scene that
must show something real and specific, and for cheap drafts and variations. Keep on-image text
short.

## ideogram-4.5

Ideogram's typography and graphic-design model, six credits. Posters, covers, packaging, signage,
logos and editorial layouts where the lettering is the design. Open with the medium and the
layout zones, give every word that must appear in double quotes with its placement, size, case,
weight and type character, then describe the picture under the type. The brief is rendered
exactly as written - nothing rewrites it - so say everything explicitly. Holds the exact ratio.
With photos attached it edits the FIRST one (add a headline, swap a label, restyle the
background) and takes the rest as references. Latin and short Cyrillic lines render cleanly; for
long Cyrillic copy take gpt-image-2.6-flare.

## flux

The fastest. Quick drafts, mood exploration, variations the user will iterate on. Its in-image
text is the weakest of the lineup; never pick it for copy. Returns 2:3 for tall ratios.

## The rule that outranks the rest

Aspect ratio overrides everything above. When the ratio must be exact, pick gpt-image-2.6-flare,
gpt-image-2.6-sunburst, the 2.5 pair, nano-banana-pro, nano-banana-2.1, muse-image, ideogram-4.5 or a seedream model.
The panoramic 21:9, 4:1, 8:1, 1:4 and 1:8 exist only on nano-banana-2.1; any other model renders
them 1:1. Leave the model empty only if nothing
fits; it then falls back to gpt-image-2.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 323kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 323kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 323kUpdated Claude CodeCodexVerifiedInstall