Nano Banana 2 vs ChatGPT Images 2.5 vs Midjourney V8.2.
Thirty identical briefs, a fair scoring sheet and a decision workflow. This is a reproducible comparison plan, not a fabricated claim that we ran a benchmark.
What this comparison can—and cannot—tell you
A comparison is only as credible as its inputs. The proposed title for this slot said “30 prompts, same brief” and implied a grid of outputs. There is no underlying run in this repository, no saved prompts, no model settings, no output set and no scoring rubric. A responsible guide cannot turn that missing evidence into a winner. Instead, the 30 briefs below are a ready-to-run protocol. If you run it, save the exact prompt, account tier, model or version, settings, timestamp, output and score. Then you can publish results that another person can inspect.
The product facts are also deliberately narrow. OpenAI's current help page says ChatGPT Images creates and edits images, supports selected edits and aspect ratios, and is available on all tiers. It also says the official DALL·E GPT has been retired. Google's documentation identifies Nano Banana 2 as the stable gemini-3.1-flash-image model, with image and video inputs, image generation, search grounding and several aspect-ratio and size options. Midjourney's version documentation identifies V8.2 as the current default and describes its edit model. Those facts explain capabilities; they do not prove comparative quality.
The practical decision table
| Need | Start with | Why |
|---|---|---|
| Conversational photo edits | ChatGPT Images | OpenAI documents upload, selection and conversational edits. |
| API image workflow | Nano Banana 2 | Google documents a stable model ID and developer inputs. |
| Aesthetic exploration | Midjourney V8.2 | The vendor describes aesthetics, personalization and an edit model. |
| Exact text | Run a pass/fail trial | All tools need a full-size spelling check; never infer reliability from a marketing claim. |
| Reference preservation | Run a preservation trial | The decisive question is what changes, not how attractive one sample looks. |
This table is a starting decision, not a ranking. Test the exact task, input type and delivery channel you use. A web app and an API may expose different controls. A free tier may have different limits from a paid plan. A feature documented for a model may not be present in every interface.
The 30-brief protocol
Run every brief with the same canvas size where the product allows it. Do not rewrite a prompt to rescue one tool unless the protocol explicitly has a second pass. Store first attempts separately from revisions. Score the first attempt on instruction coverage, then score a revision on editability. This prevents a conversational tool from appearing worse simply because you stopped before its refinement loop, while still making clear that the first result is part of the user experience.
Brief 1
Text in an announcement
Create a clean square announcement for a community workshop. Leave a blank panel for text and do not render letters. Compare how easily each tool leaves a usable layout.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 2
Text that must be rendered
Create a simple poster with the exact headline: COMMUNITY CLEAN-UP, Saturday 10 AM. Use high contrast and a plain background. Treat spelling and punctuation as a pass/fail check, not an aesthetic opinion.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 3
Single-person portrait
Create a natural editorial portrait of a fictional person in soft window light. Preserve realistic skin texture, one clear light direction and a 4:5 crop.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 4
Reference portrait edit
Using the supplied reference, change only the jacket colour to forest green. Keep face, hair, pose, lighting and background unchanged. Record every changed element.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 5
Couple reference
Using two supplied reference images, create a relaxed outdoor portrait while keeping each person's facial features separate and recognisable. Note whether identity drift needs correction.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 6
Product hero
Create a studio product photograph of a reusable water bottle on a pale stone surface, with a clean reflection and negative space for copy. No brand or invented label.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 7
Packaging mockup
Create a blank paper pouch mockup with a flat front panel, three-quarter view and soft commercial lighting. Leave the panel empty for typesetting later.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 8
Food still life
Create a top-down still life of a bowl of jollof rice, plantain and salad on a warm ceramic table setting. Keep the composition believable and avoid text.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 9
Editorial illustration
Create a flat editorial illustration of a person sorting receipts at a desk, limited teal and orange palette, clean shapes and generous space for a headline.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 10
Watercolour scene
Create a loose watercolour of a rainy Lagos street at dusk, visible paper texture, restrained colour bleeding, no readable signs and no stereotypes.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 11
Diagram without labels
Create a clean three-stage diagram showing input, transformation and review as labelled empty boxes with arrows. Do not generate words; add them later in a design tool.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 12
Diagram with labels
Create a three-stage diagram with the exact labels INPUT, TRANSFORM and REVIEW, aligned in that order with arrows. Check every label at full size.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 13
Social portrait crop
Create a 9:16 vertical portrait with the subject in the lower two-thirds and uncluttered space above for later copy. Keep the background quiet and the eyes sharp.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 14
Website hero
Create a 16:9 website hero image for a budgeting app: a notebook, calculator and phone on a desk, with the left third intentionally calm for interface text.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 15
Thumbnail concept
Create a bold video thumbnail composition with one expressive fictional face, three large colour blocks and no text. Make the focal point readable at small size.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 16
Background removal
Use the uploaded object image and remove the background, preserving the object's edges, holes, transparent parts and natural shadow. Return a transparent-looking result with no halo.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 17
Object replacement
In the supplied kitchen photo, replace only the blue mug with a red ceramic mug. Preserve every other object, reflection, shadow and camera angle.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 18
Lighting correction
Correct a dim indoor photograph to soft neutral daylight. Preserve facial features, clothing colours and composition; do not beautify, sharpen excessively or change the room.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 19
Old-photo restoration
Restore the supplied family photograph by reducing scratches and dust while retaining age, facial identity and original composition. Do not colourise unless asked.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 20
Style conversion
Transform the supplied street photograph into a screen-print-inspired illustration while preserving the building silhouettes, perspective and subject placement.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 21
Consistent character
Create four storyboard panels featuring the same fictional courier: same coat, hair, backpack and face. Change only the action and camera distance in each panel.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 22
Pet portrait
Create a dignified portrait of the supplied dog in a quiet garden. Preserve markings, ear shape and coat pattern; do not add clothing or human expressions.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 23
Travel postcard
Create a modern travel postcard illustration of a coastal town, with no landmarks copied as a claim of location and no text. Use a limited two-colour print palette.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 24
Fashion flat lay
Create a top-down flat lay of a neutral capsule wardrobe with clear garment separation, natural fabric folds and no logos. Leave the centre open for layout.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 25
App icon
Create a simple app icon for a fictional reading app: one open book shape, two colours, strong silhouette, no letters, no trademarked symbols and transparent background.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 26
Scientific visual
Create a non-literal educational illustration of the water cycle with empty labelled zones, arrows and clear separation. Do not present it as a measured scientific diagram.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 27
Interior concept
Create a calm small reading corner with a chair, lamp and bookshelf, realistic scale, one window light source and enough empty wall for later art.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 28
Night scene
Create a rainy night bus stop with practical lights, wet pavement and a safe, non-threatening mood. Avoid readable advertisements and invented public notices.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 29
Collage
Create a paper collage of leaves, ticket stubs and geometric shapes, tactile edges, muted colours and no readable text. Keep each layer visually distinct.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
Brief 30
Print check
Create a 4K-ready concept for a square product card with a plain background and a single centred object. Keep fine details simple enough to survive downscaling and print review.
Record: input files, exact model/version, account or plan, output size, first result, revision count, failures and reviewer notes.
How to score without fooling yourself
Use a five-point scale for each brief: instruction coverage, identity or object preservation, composition, artefact control and delivery suitability. Define a point before looking at results. For example, instruction coverage earns five only when all requested elements are present and no forbidden element appears. Preservation earns five only when the reference subject survives unchanged in the named areas. Delivery suitability includes crop, resolution, legibility and whether a human designer can finish the asset. Do not award points for “vibes” unless the brief explicitly asks for an aesthetic.
Keep three totals separate: raw first attempt, best revision and time-to-acceptable. The last metric must include waiting, failed generations, downloading, manual cleanup and typography fixes. If you have no timed run, do not publish a speed claim. If a service changes during the test, split the report by date and model version. If you use a subjective panel, disclose panel size, instructions and disagreements. Never replace those details with “we tested it” or a made-up percentage.
A repeatable workflow for your own team
Start with an asset brief, not a model preference. Write the audience, job, subject, required elements, forbidden elements, crop, destination and approval owner. Separate facts from taste: “the bottle must remain upright” is a pass condition, while “the image should feel premium” needs a definition such as restrained colour, controlled reflections and no clutter. This makes a comparison useful to a designer, marketer or product team rather than turning it into a gallery of opinions.
Create a folder for each tool and a manifest with one row per generation. Include the prompt exactly as sent, attached input names, date in UTC, interface or API, model/version, selected aspect ratio, output size, revision number and reviewer. Keep failed outputs too. A failed attempt explains why the next prompt changed and stops a polished final image from hiding the cost of getting there.
For reference edits, mark protected regions before you start: face, hands, product label, logo, building silhouette or background geometry. Review only those regions first, then inspect the rest. For text, zoom to 100 percent and transcribe every visible character into the review record. For products, compare shape, count, colour and orientation against the source. For people, ask the subject for consent and avoid publishing a realistic depiction as a factual event unless it is clearly labelled.
Run a second pass only when the first pass identifies a specific defect. “Fix the image” is hard to audit; “keep the face and replace only the left background sign with a blank wall” is a controlled edit. Log whether the tool changed anything outside the request. That measure—unrequested change—is often more valuable than a subjective beauty score because it predicts how much cleanup a real workflow will require.
Finally, choose a winner per task, not one winner for the whole category. A tool can be the best fit for a fast concept and the wrong fit for a sensitive identity-preserving edit. Publish the decision rule with the result: for example, “we chose the tool that met all protected-region checks with the fewest revisions.” If nobody ran the protocol, publish the protocol and say so. Readers can then reproduce the work instead of trusting a number with no evidence behind it.
Source-checked product notes
OpenAI's Images in ChatGPT documentation says ChatGPT Images 2.5 supports creating and editing, selection-based edits, aspect-ratio changes and image uploads; it lists availability on all tiers and says the official DALL·E GPT has been retired. Treat that as current product documentation, not as a promise that every user sees identical limits.
Google's model page names the stable model code gemini-3.1-flash-image, supports text, image, video and PDF inputs, and lists image generation and search grounding. Google's image-generation guide documents editing, 0.5K through 4K options, video-to-image support and limitations such as imperfect adherence to requested output counts.
Midjourney's version page says V8.2 is the current default, released July 24, 2026, and describes an Edit Model replacing older reference and retexture tools. It does not establish a universal winner for text, faces, products or speed.
A safe publishing and buying checklist
- Write the brief and pass criteria before generating anything.
- Remove private faces, documents, addresses and client assets unless you have permission and understand the vendor's handling.
- Check the current plan, credit, API and commercial-use terms on the vendor's own site.
- Inspect hands, faces, logos, text, borders and small details at full size.
- Keep a human approval step for public, political, medical, financial or identity-sensitive images.
- Label illustrative comparisons honestly. A prompt pack is not a benchmark, and one attractive output is not evidence of a general capability.
For more task-focused prompts, see the ChatGPT image prompt library, the Nano Banana 2 prompt guide and the Nano Banana Pro guide.
Frequently asked questions
Did this page run the 30 prompts in all three tools?
No. The CSV supplied that idea as a research lead, but no reproducible run, outputs, scoring sheet or account records were available. This page gives you a controlled 30-brief protocol instead of inventing a result.
Which tool is best?
It depends on the job. ChatGPT is convenient for conversational creation and edits, Nano Banana 2 is documented by Google as a high-efficiency image model with image and video inputs, and Midjourney V8.2 is its current default with an edit model. Choose against your review criteria.
Can I compare pricing in a single table?
Only with a timestamp and region. Plans, credits, API rates and included features change. Check each vendor's live billing page immediately before purchase rather than treating a static comparison as a quote.
Acceptance thresholds matter
Before generating anything, define the minimum acceptable result: the subject count must be correct, the crop must fit its destination, no unapproved text may appear, and protected reference areas must survive. Record one rejection reason such as wrong count, identity drift, artefact, unusable crop or policy concern. This prevents a beautiful but unusable picture from winning by accident and gives your team a useful reason to rerun a brief.
Keep the acceptance rule separate from aesthetic preference. A reviewer can prefer warmer colour while still agreeing that the requested bottle shape is present. When those decisions are mixed together, a tool may score well because a reviewer likes its style even though it failed the actual job. Clear thresholds make the comparison transferable to a real campaign, product screen or printed asset.