OtterMind Krystian
Wydro
PL
An antique statue wearing futuristic glasses, surrounded by flowers and cosmic dust: an AI frame by Krystian Wydro
Knowledge base Guide: AI graphics

AI models for graphics
- matching the tool to the task

Which AI image generator to choose? Matching the model to the task, plus my test: 8 models, 40 images, 80 measurements of hex colours. Midjourney, Nano Banana, Seedream.

- AI Creative Director, OtterMind  ◆   ◆  16 min read

You have a campaign, product post or printable poster to make and you open yet another list of the 'best AI image generators'. Each recommends something else, because they all compare tools ignoring what you actually want to do.

Every AI image generator has its specific use. You match the tool to the task: finding a style, editing a ready image, automating or preparing a file for print. Below I describe six groups of tools, each for a different stage of work. At the end you will find my own test: the same prompts in 8 models and 80 measurements of colours given as hex codes (results).

AI image and video tools change practically every few weeks. New models, new features, new platforms.

That is why knowing the tools is not enough. A process takes you from an idea to a finished piece: first you know what you want to make, then you pick the model. That is how I arranged this overview: from the task to a specific tool. Tool list as of 29 September 2026, colour test: 30 October 2026.

The links to ComfyUI, Magnific, Figma Weave and DFirst in this article are affiliate links. When you use them to purchase a paid plan, I receive a commission. The price for you remains the same, and the commission helps me develop the knowledge base.

Tool for the task

Which AI image generator for which task?

You want to…Reach for
create a graphic from scratch and choose its styleMidjourney, Recraft, Ideogram, Adobe Firefly
change the lighting, background, clothing or detail in a ready photoNano Banana, GPT Image 2, Seedream, FLUX.2
have the same character or product in multiple framesNano Banana
make a graphic with text: a banner, infographic, product creativeNano Banana, GPT Image 2, Ideogram
keep a brand colour given as a hex codeNano Banana 2, Seedream 5 Pro and colour correction after generation
generate regularly on your own hardware, with zero generation feesFLUX, Qwen-Image, Z-Image, Stable Diffusion in ComfyUI
compare models and build the entire process in one placeMagnific, Krea, Flora, Figma Weave, DFirst
integrate generation into an application or automationfal.ai, Replicate, Gemini Enterprise Agent Platform
enlarge an image for print or restore a poor-quality photoMagnific Upscaler, Topaz, Krea Enhancer, SUPIR
01 - Commercial image generation models

Which AI image generator to choose when creating graphics from scratch?

Commercial generators are where you can start looking for the style and visual direction of a project. At this stage, Midjourney is the most important one for me. I use it when establishing what the image should look like.

Here you work with one model in its own interface: Midjourney, Ideogram, Recraft or Adobe Firefly. You enter a prompt and get an image. Each has a different aesthetic and different strengths.

Midjourney has set the direction for the entire image generation industry for a long time. Recraft is strong in vector graphics and product illustrations. Ideogram handles text in an image better than other models, something the whole category struggled with for a long time.

Midjourney or Nano Banana? If you are creating something from scratch and looking for a drawn, 3D or any creative style, Midjourney will be the best choice. If you need photorealism, models from Google, meaning Nano Banana, will work just as well.

From practice: brand style in Midjourney. A moodboard represents the style in which Midjourney will generate images. If you have a set of graphics for your brand, create a moodboard from them and generate in that style.

From practice: style described in words. I suggest describing the effect in your own words instead of using artists' names: colour palette, textures, composition, era, saturation, technique. You can then transfer such a description to any other model.

With a photo generator, start with your goal too. You can match different models to subsequent stages of work.

02 - Contextual models and multimodal editing

How to change a fragment of a photo in AI and keep the rest?

In contextual models, an existing image can be the starting point. You pass it to the model and describe the change: different lighting, background, clothing, prop, detail or composition. The remaining elements of the image are to be kept.

Contextual models, such as Nano Banana, FLUX Kontext or Seedream, allow you to change a shirt to a different one, insert a product into a lifestyle scene or recreate the same character in different shots. Previously, this required Photoshop, manual graphic design work or a photo shoot. Now all you need is a reference photo and a prompt.

From practice: one sentence that saves the frame. Sometimes you have to explicitly ask the model to keep elements from the reference. If I am changing something in a photo, I type: keep the entire structure of the photo, and replace the given character.

From practice: brand colour. A HEX code in a prompt gives a colour from the same family, but rarely the same one: in my test only 2 of 80 samples were indistinguishable from the code. It is much better to provide the model with a graphic containing that colour, and still check the finished file with a colour picker.

How to keep the same character or product in multiple AI graphics?

Nano Banana has changed the way we work with consistency. You can provide a photo of a product or a character, and then develop subsequent images while preserving their most important features.

With a product, you keep the object and change the surroundings. With a character, you keep the face and silhouette, and change the outfit, framing or the entire scene. Prepare a character sheet: a portrait from the front, profile, three-quarters and a full silhouette. You return to it with each subsequent shot, and the model holds the character's appearance more precisely.

The simpler the product shape, the better the result. Fine details, labels with small text or complex shapes require more iterations, and sometimes manual correction. Midjourney would recreate the product from scratch, resulting in slight variations. Contextual models have an advantage here and allow you to create additional formats for social media.

How to make an AI graphic with text?

For product creatives, I recommend Nano Banana: it maps the product best and handles text on the image well. To get the text right, keep it below 400 words in the entire graphic: more text means more errors and distortions. Nano Banana 2 writes very well, although occasional errors still happen.

GPT Image 2 and Seedream 5 also offer great control during generation and editing.

03 - Open-weight and locally run models

How to generate AI images on your own computer and is it worth it?

Open-weight models allow you to work on your own infrastructure and control the generation environment more precisely. You can combine them with additional models and build a process tailored to a specific task. These models come in handy for automation and working with your own data.

You run Z-Image, FLUX, Qwen or Stable Diffusion on your own hardware: generation costs zero. However, full control requires a powerful graphics card and patience during configuration, but in return, it offers possibilities unavailable in any cloud.

Is it really for free? Local work is free, although we still pay for the graphics card, other components and energy, and these have become more expensive lately. It makes the most sense when we generate regularly and use the same setup across multiple projects.

In ComfyUI we have much more control: we choose how many steps the model will perform, what prompt tools it uses and so on. It is an environment where you assemble your own workflow from blocks.

Stable Diffusion has historical significance here. For many creators, it was the starting point for working with image generation on their own computers, using their own models and extensive workflows. At one time, LoRAs were very popular for images, but models supporting references have reduced their popularity because they handle this much better.

In the case of Ideogram, the manufacturer provides the weights of version 4.0 for download and running on your own hardware. You can find access information on the official model page (as of 29 September 2026).

04 - Multimodel platforms and aggregators

Where to have multiple AI models in one place?

Multimodel platforms gather various models and tools for further image processing. Some allow you to connect subsequent stages using nodes. This way, you build an entire process, rather than performing each generation separately.

You test different models on the same project and choose the result that fits best, and when one model struggles, you simply use another. Some models still experience difficulties with hands and fingers. In my test with a tennis player, Nano Banana performed best.

From practice: advertising creatives on nodes. Magnific Spaces acts like a creative board: each block performs a specific operation and passes the result to the next one. In the Designer tool, you choose your own font, and you provide the colour from the brandbook to a tool that understands colours, rather than leaving it to the model. You get greater brand consistency compared to fully generated text. Such a workflow pays off with repetitive content: building it takes some time, and it has to pay that time back.

Magnific is convenient for campaign materials because it combines generation with a stock photo library and editing tools. Individual platforms differ in their workflow and additional features. Figma Weave provides access to generation and the ability to add models from fal.ai. Magnific has a large creative community and extensive teamwork capabilities. DFirst is a Polish company.

An aggregator can also simplify your subscription choice. You get access to many necessary models and tools in one place. It is worth paying attention to the entire process you want to build. If you want such a process in your team, I help design it through AI implementations.

05 - API and infrastructure

Generating images via API: when does it make sense?

API allows you to incorporate image generation into your own applications and production processes. Corporate implementations and automations rely on connecting models via API. Gemini Enterprise Agent Platform from Google (formerly Vertex AI) is an example of a solution for companies, while fal.ai provides an API for generative models.

In an API, you pay for model usage, usually per operation or generation. At a larger scale, it is worth controlling the number of calls and the cost of the entire process. From practice: I can run generation through fal.ai, but then I pay for credits. For my own needs, I run it through a local computer, controlled from a browser.

Prompt in JSON. Use it when you are building automation, working with a model via API or the model is prepared for such prompting, like FLUX.2. In everyday work in visual interfaces, JSON adds little value.

06 - Upscaling, restoration and enhancement

How to upscale an AI image so it is fit for print?

Upscaling tools increase resolution, restore photos and enhance images. Additional enlargement comes in handy when preparing material for print, working with large formats and restoring older photographs.

Sometimes a photo turns out blurry after enlargement, yet we need it for print. Or we have generated an image and want to make a flyer out of it. Raster graphics, unlike vector graphics, scale less effectively. That is when we need an upscale.

We reach for upscaling when the image is already selected and refined.

In this way, we prepare a file in the resolution required for publication or print. Topaz and Magnific enlarge the image, sharpen details and remove noise. This is often the last step before sending the material to production. From a blurry face with indistinct wrinkles and eyelashes, the Magnific upscaler extracts sharp details even at five times magnification.

07 - Model selection

How to match a model to a task and save hours of work?

Match the model to the task, and when something gets stuck, switch the tool instead of struggling with the product. In practice, 2-3 graphics models will suffice: Midjourney, Nano Banana and Seedream. The rest are situational add-ons.

The 30-45 minute rule. Sometimes a model stubbornly resists cooperating despite good prompts. In that case, try a related one. I have had many situations myself where a different model captured a jacket or a Scandinavian interior much better. If something takes you longer than 30-45 minutes and the result remains poor, jump to another model.

Before you publish:

  • Free tools usually lack a commercial licence: check the subscription terms.
  • The exact same model might keep your data private in one environment, while in another, it might use your graphics for algorithm development or marketing.
  • With products, check the text on the label, shape and legibility of the logo.

I teach how to select tools for subsequent stages of work in my AI graphics and video training courses. If you prefer to learn on your own, we go through the entire process from an idea to a finished creative in the Sprint AI course.

The next step is describing the image you want to achieve. I break the description into six fields: subject, background, details, technique, lighting and composition.

I cover video generators separately in Which AI video generator to choose?.

08 - Colour test · 30 Oct 2026

Which AI image generator reproduces a hex colour most faithfully?

Brand graphics have their own colours, and those have to match the brand book. So I checked what models do with a colour given in the prompt as a hex code.

I gave the same prompts to eight models in Runway. Five portraits, each with two colours given as codes: 10 colours and 80 samples. Then I measured in the files how much of those colours was left.

In short: Nano Banana 2 came closest to the codes (mean ΔE 9.0), with Seedream 5 Pro right behind (10.3). GPT Image 2.5 (17.3) and Grok Imagine 2 (19.9) were furthest away. Exact hits are rare, though: only 2 of 80 samples were indistinguishable from the code, and 49 were a different colour altogether.

How I measured

I ran the test on 30 October 2026. Models: Nano Banana Pro, Nano Banana 2.1, Nano Banana 2, Nano Banana 2 Lite, GPT Image 2.5, Grok Imagine 2, Seedream 5 Pro and Ideogram 4.5. Each got the same 5 prompts and one attempt per prompt. Portrait 3:4, 2K resolution where the model allows it (Nano Banana 2 Lite gives 896 × 1200 px). All people in the images were generated by the models.

On each of the 10 elements with a code I marked the lit part and took the median of its brighter pixels, without shadows and highlights. I measured the difference from the code in ΔE (CIEDE2000), the standard measure of colour difference:

  • below 2: indistinguishable,
  • 2-5: visible when the colours sit side by side,
  • 5-10: a clearly different shade,
  • above 10: a different colour.

I measure what you get in the file. Light in the scene, such as dusk or a tungsten lamp, also shifts the colour, but that file is what ends up in your project. One attempt per model is a signal, not a verdict, which is why I run the same prompts every month.

ModelMean ΔESamples with ΔE below 5Closest to the codeFurthest from the codeRunway credits per image
Nano Banana 29.03 of 10jacket #F2E8D5 (1.4)turtleneck #A4472A (23.0)11
Seedream 5 Pro10.33 of 10jacket #F2E8D5 (1.4)turtleneck #A4472A (26.2)9
Nano Banana 2 Lite11.32 of 10coat #4F5D6B (4.5)hair #B86B3A (18.9)4
Nano Banana Pro11.63 of 10sweater #D9772B (3.8)turtleneck #A4472A (25.3)20
Nano Banana 2.111.91 of 10sweater #D9772B (2.6)turtleneck #A4472A (24.2)6
Ideogram 4.514.00 of 10coat #6D1F2C (5.3)jacket #F2E8D5 (30.4)25
GPT Image 2.517.30 of 10background #4A2C1F (7.4)hair #B86B3A (27.2)10
Grok Imagine 219.91 of 10sweater #D9772B (4.7)jacket #F2E8D5 (51.5)8

Credits are the cost of one image in Runway. One prompt across 8 models cost 93 credits, the whole test 465. The most expensive, Ideogram 4.5 (25 credits per image), came sixth on colour. The middle of the table, Nano Banana 2 Lite, Pro and 2.1, is practically a tie: with one attempt, a 0.6 ΔE difference settles nothing.

What happened to the colours

No model hit the rust turtleneck #A4472A. Results ranged from 12.1 to 26.2 ΔE. Seven models made it much darker, and most pushed it towards red: brick and burgundy. Ideogram 4.5 came closest but drifted into brown. In the photo your eye sees rust; the sample from the file does not.

Darker and greyer. 64 of 80 samples came out darker than the code, and 68 less saturated. Every model turned the copper hair #B86B3A brown, with half the saturation or less.

Muted colours change hue. Ideogram 4.5 turned the sage skirt #9CAF88 turquoise, the slate coat #4F5D6B a greenish grey and the bottle-green jacket #2F4A3A emerald. Seedream 5 Pro turned the slate coat navy, and Nano Banana 2 Lite made the skirt khaki.

Every model hits saturated orange. The sweater #D9772B came out at ΔE 2.6 to 8.7 across all eight models.

Cream: two hits, two misses. Nano Banana 2 and Seedream 5 Pro rendered the cream jacket #F2E8D5 at ΔE 1.4, which is indistinguishable. GPT Image 2.5 gave beige (25.2), Ideogram 4.5 grey (30.4). For Grok (51.5) most of the difference comes from the dark scene: in context, the jacket looks cream.

My eye was wrong too. The Seedream coat looked red rather than wine to me. In the measurement it was the closest of all models to #6D1F2C (ΔE 3.5); the others gave a darker burgundy. That is why I check colour in the file, not by eye.

All measurements

ΔE for each of the 80 samples. In each row, the bold value is the one closest to the code.

Element and codeNB ProNB 2.1NB 2NB 2 LiteGPT Image 2.5Grok 2Seedream 5 ProIdeogram 4.5
coat (arcade) #6D1F2C7.78.211.97.911.617.03.55.3
scarf #D4A23A22.122.47.914.822.913.610.720.2
sweater #D9772B3.82.64.57.08.04.78.77.6
skirt #9CAF8812.613.39.113.615.715.86.513.9
hair #B86B3A16.312.19.418.927.230.114.215.3
velvet jacket #2F4A3A9.87.75.29.817.113.813.911.0
background (velvet) #4A2C1F4.312.615.216.37.416.64.811.0
bouclé jacket #F2E8D53.95.51.44.725.251.51.430.4
turtleneck #A4472A25.324.223.015.224.821.526.212.1
coat (coast) #4F5D6B10.110.42.64.513.014.713.212.8
Mean11.611.99.011.317.319.910.314.0
80 crops of fabric, background and hair from the test images in eight model columns, each with the code colour and the measured colour
Each sample is a crop from the image. Below it, the left square is the code from the prompt, the right one is the colour measured in the file, and the number is ΔE.

Beyond colour

The prompts also described pose, action, skin tone and framing. Here, too, the models did not always listen.

2 of 8 models rested the forehead on the glass. In the study portrait, the man was supposed to rest his forehead against the cold window. GPT Image 2.5 and Seedream 5 Pro did it; the others placed him next to the window.

Ideogram 4.5 lightened dark skin. In both portraits described with “deep brown skin”, it lightened the skin and changed the facial features.

A second head. In the portrait with the cup, Ideogram 4.5 drew a second, empty head with a bun above the woman. It is easy to miss in a thumbnail, not in print. It was also the most expensive: 25 credits per image.

Four Google models, almost the same face. Nano Banana Pro, 2.1, 2 and 2 Lite produced a very similar person in three of the five portraits. For a series with one character, that is an advantage. When you need different people, change the face description or the model.

Framing and pose. In the glasshouse I asked for a full figure with one knee drawn up. Seedream 5 Pro cropped the frame to half the figure and Ideogram 4.5 lost the pose. The Google models, GPT Image 2.5 and Grok Imagine 2 kept both.

What this means for your work

  • If the brand colour has to match, check it in the file with a colour picker, not by eye, and plan a colour correction after generation.
  • For images with brand book colours, start with Nano Banana 2 or Seedream 5 Pro.
  • Check muted colours most carefully: rust, sage, slate, copper. This is where models most often darken the colour or add their own hue.
  • View the image at full resolution. A second head or altered skin tone can hide in a thumbnail.
  • Price per image says nothing about quality: the most expensive model came sixth on colour.
09 - 40 images and 5 prompts

What does the same prompt look like in 8 models?

Below are all 40 images from the test: one from each model per prompt, unedited. Under each image you will find ΔE for both coded colours. The prompts are below the galleries, exactly as the models received them. You can copy them and test them in your own tool.

1. Arcade: wine #6D1F2C and mustard #D4A23A

Seedream 5 Pro came closest to the wine (ΔE 3.5) and Nano Banana 2 to the mustard (7.9). For six models, the scarf drifted from the code by more than 13 ΔE.

Masterclass award winning film still photography, overcast editorial street portrait with soft film grain and muted, slightly lifted blacks. A grey limestone arcade on a cold autumn morning, flat diffused daylight from a white overcast sky, damp stone floor with a faint grey sheen. A man in his late twenties with a square jaw, strong straight brows, dark brown almond eyes with amber rims, a broad nose and a thin scar line through the right eyebrow; deep brown skin with cool undertone, light texture on the forehead, a small razor bump under the left jaw; close-cropped black hair with a slightly uneven hairline at the temple. He wears a burgundy (#6D1F2C) boiled-wool double-breasted coat with the collar turned up and a mustard ochre (#D4A23A) ribbed-knit scarf wrapped twice, one end hanging loose. Wind catches the scarf end as he turns his head over his shoulder toward something off-frame, jaw relaxed, eyes narrowing slightly. A single dark wine (#6D1F2C) dahlia pinned to the coat lapel. Waist-up shot, camera at eye level slightly off-axis, shallow depth of field, arcade columns dissolving behind him. Color palette built on burgundy #6D1F2C and mustard #D4A23A against neutral stone grey. Prompt 1: Arcade

2. Glasshouse: pumpkin #D9772B and sage #9CAF88

All models hit the orange sweater (closest: Nano Banana 2.1, ΔE 2.6). Seedream 5 Pro rendered the sage skirt best (6.5) but cropped the frame. Ideogram 4.5 made the skirt turquoise and lost the pose.

Masterclass award winning film still photography, soft medium format fashion editorial with gentle grain and pastel-lifted shadows. An old glasshouse with sage-green (#9CAF88) painted iron frames and milky panes, overcast daylight diffusing through the glass from above, cool even light with a warm bounce from terracotta floor tiles. A woman in her late twenties with a heart-shaped face, wide-set dark brown eyes, softly arched brows of slightly uneven height, a rounded nose tip and naturally full lips; medium brown skin with golden undertone, light freckling across the nose, a faint scar on the chin; a thick waist-length black braid, frayed with loose flyaways near the crown. She wears an oversized pumpkin-orange (#D9772B) chunky mohair sweater and a pleated sage-green (#9CAF88) silk midi skirt. She sits on an old wooden potting bench, one knee drawn up, hands wrapped around it, looking straight into the lens with a steady, unhurried gaze, mouth at rest. Two potted sage plants and one small pumpkin (#D9772B) rest at the edge of the bench. Full-body shot from a low position at bench height, foreground leaves softly blurred at the frame edge. Color palette built on pumpkin #D9772B and sage #9CAF88 with soft neutral light. Prompt 2: Glasshouse

3. Study: copper #B86B3A and bottle green #2F4A3A

Nano Banana 2 rendered both colours best (hair 9.4, jacket 5.2). Only the men from GPT Image 2.5 and Seedream 5 Pro rest their forehead on the glass.

Masterclass award winning film still photography, low-key interior portrait with dense film grain and deep warm blacks. A dark forest-green (#2F4A3A) wood-panelled study at dusk, a single tall window on frame-left, cool fading daylight falling across the face, the rest of the room sinking into shadow, a small brass lamp glowing warm far in the background. A man in his early thirties with a long narrow face, pronounced cheekbones, grey-green heavy-lidded eyes, a slightly hooked nose and a thin upper lip; pale skin with pink undertone, dense freckles across nose and cheeks, faint redness around the nostrils; shoulder-length wavy copper-red (#B86B3A) hair, tangled at the ends, one strand falling over his eye, and a short trimmed copper beard. He wears a forest-green (#2F4A3A) cotton velvet jacket with a worn sheen on the cuffs over a dark knit. He rests his forehead against the cold window glass, breath leaving a small fog patch on the pane, eyes half closed. Close-up in profile from slightly behind his shoulder, shooting past the window frame, shallow depth of field, cool daylight on copper hair against deep green shadow. Color palette built on copper #B86B3A and forest green #2F4A3A with deep shadows. Prompt 3: Study

4. Velvet: chocolate #4A2C1F and cream #F2E8D5

Nano Banana 2 and Seedream 5 Pro hit the cream (ΔE 1.4); Nano Banana Pro rendered the background best (4.3). In the Ideogram 4.5 image, a second, empty head appears above the woman.

Masterclass award winning film still photography, warm intimate editorial portrait with fine grain and soft halation in the highlights. A room draped in chocolate-brown (#4A2C1F) crushed velvet, lit by a single warm tungsten source from high camera-right, a narrow triangle of light on the shadow cheek, everything beyond the subject falling into deep brown. A woman in her early forties with a wide oval face, high forehead, dark brown eyes with a slight hood on the left lid, a broad nose and full defined lips; deep brown skin with red undertone, a soft sheen on the cheekbone, a few raised dark spots near the temple; natural black coils with silver streaks pinned in a loose low bun, several coils escaping at the nape. She wears a cream (#F2E8D5) wool bouclé jacket over an ivory silk blouse with a high ruffled collar. She glances back over her right shoulder toward the camera, chin lowered, one corner of her mouth lifting slightly, holding a cream (#F2E8D5) porcelain cup close to her collarbone with a thin thread of steam rising. Medium close-up at eye level, Rembrandt lighting, shallow depth of field. Color palette built on chocolate #4A2C1F and cream #F2E8D5 in warm tungsten light. Prompt 4: Velvet

5. Coast: rust #A4472A and slate #4F5D6B

Nobody hit the rust turtleneck (closest: Ideogram 4.5, ΔE 12.1). Nano Banana 2 rendered the slate coat best (2.6); Seedream 5 Pro turned it navy.

Masterclass award winning film still photography, raw coastal editorial portrait with coarse film grain and desaturated cool shadows. A slate-rock shoreline under a low heavy overcast sky, flat cold daylight, wet slate-grey (#4F5D6B) stone with a dull silver sheen, a slate-blue sea fading into mist behind. A woman in her late twenties with an angular face, strong jaw, thick dark eyebrows, deep brown eyes, a narrow nose with a small bump and wide lips; light olive skin with neutral undertone, wind-reddened cheeks, a tiny healed scar above the upper lip; shoulder-length dark brown hair whipped across her face by the wind, a few strands caught at the corner of her mouth. She wears a rust-red (#A4472A) ribbed merino turtleneck and a heavy slate-grey (#4F5D6B) wool coat hanging open. She stands braced against the wind, holding a bunch of dried rust-red (#A4472A) hydrangea against her chest, looking slightly past the lens, eyes squinting against the cold. Medium shot from a low angle, horizon low in the frame, rust fabric as the single warm tone in a cold grey scene. Color palette built on rust #A4472A and slate #4F5D6B. Prompt 5: Coast
FAQ

AI image generator: frequently asked questions

What is the best AI image generator?

The best choice depends on the task. For style and creative images, reach for Midjourney; for photorealism and photo editing, choose Nano Banana; for vector graphics, use Recraft; and for text in an image, pick Nano Banana, GPT Image 2 or Ideogram.

Will an AI image generator reproduce a brand colour exactly from a hex code?

Rarely. In my test of 30 October 2026 (8 models, 80 samples), only 2 samples were indistinguishable from the code, and 64 came out darker. Nano Banana 2 came closest, followed by Seedream 5 Pro. Check the brand colour in the file and plan a correction after generation.

Midjourney or Nano Banana?

Midjourney, when you are creating from scratch and looking for a drawn, 3D or creative style. Nano Banana, when you need photorealism, are altering your own photo or making banners and infographics.

Is there a free AI image generator?

Open source models, such as Z-Image, FLUX, Qwen or Stable Diffusion, run locally, are free of generation costs: you pay for the hardware and energy. Free online tools rarely grant a commercial licence.

How to achieve a consistent character in AI?

Prepare a character sheet (front, profile, three-quarters, full silhouette) and work in a contextual model, e.g. Nano Banana: you preserve the face and silhouette, while changing the outfit and scene.

How many AI models for graphics do you need?

In practice, 2-3: Midjourney, Nano Banana and Seedream. The rest are situational add-ons.

How to enlarge an AI graphic for print?

At the end of the process, when the image is already selected and refined. An upscaler (Magnific, Topaz, Krea Enhancer, SUPIR) increases the resolution, sharpens details and removes noise.

Can I use AI graphics commercially?

It depends on the tool and plan. Before using a graphic in a campaign, check the subscription terms.