My builder scorecard ยท August 2026
The model is only as good as the job you give it.
Personal scores from repeatable creative briefs. They are not public Elo, provider claims, or a permanent ranking.
Overall
Prompt adherence
Style & design
Editing & reference
Where I'd use each one
A still-image workflow, in plain English. The score alone is not the choice.
The first complete brief
My default when I want a polished, usable draft with the full visual hierarchy in place.
Narrative image sequences
Where I test character and scene continuity before moving a concept into a series.
Reference-led edits
Best suited to assembling several useful references into one controlled new frame.
Layered visual direction
For reference-heavy briefs, edits, and scenes that need more than surface-level style.
Art direction
The one I use when the image needs taste, texture, and a less default AI visual language.
Designed compositions
A strong alternative for structured prompts with explicit subject, style, action, and context.
Precise briefs
A serious option when literal instruction-following and grounded details matter most.
Fast creative exploration
For expressive still-image concepts when I want an unexpected interpretation to react to.
Cheap rough variations
The practical pick for volume, early directions, and low-cost visual exploration.
I started comparing these models because I kept losing evenings to beautiful images that were unusable for the actual thing I was making. I am not looking for a model that can win a screenshot competition after twenty lucky retries. I am looking for the one that survives a real brief: a product visual, a character-led marketing slide, a weird art direction, a reference edit, then three more versions without collapsing into AI wallpaper.
That is why GPT Image 2 has become the one I usually open first. That is a habit from my workflow, not a claim that it wins every image.
Even at its lower settings, it often feels like a cheat code. The first draft arrives with the hierarchy intact. The composition understands the point. The tiny instruction buried halfway through the prompt has a better chance of existing. It does not mean every generation is perfect. It means I spend less time negotiating with the model before I have something worth directing.
This is a builder's notebook, not a universal leaderboard
The scorecard above is deliberately personal. I use repeatable briefs from the work I am actually doing for DreamCraft: editorial product images, character sequences, stylised campaign art, constrained edits, and reference-driven scenes. I allow sensible rerolls, then score the output I would genuinely keep.
Those numbers are not public Elo. They are not provider claims. They are my current answer to one practical question: which tool gets me to a usable decision fastest?
GPT Image 2 wins my first-draft test
GPT Image 2 is the most dependable complete-brief model in my current workflow. It tends to understand the whole ask, not merely render individual keywords. Layout, visual hierarchy, atmosphere, the annoying detail that normally disappears, and the relationship between objects all arrive more often in the same image.
That matters more to me than an abstract crown. When I am building a feature or planning a creative, I want the first result to create momentum. I want a thing I can sharpen, reject, remix, or ship. GPT Image 2 keeps giving me that starting point.
Why the next few places are interesting
Nano Banana Pro is the serious precision alternative. Google positions its Pro tier for harder creative work, richer real-world grounding, and high-fidelity output. In my tests it is particularly strong when the brief needs to be taken literally and the output cannot simply be pretty.
Krea 2 is different. It is not trying to be a neutral answer machine. Its value is aesthetic steering. If the goal is to escape the default polished AI look and find a visual language with more personality, Krea 2 is one of the most exciting places to iterate. It rewards exploration, then becomes more controllable as you add the right composition, material, lighting, and style constraints.
FLUX.2 remains a strong creative tool with a different visual character. Its own prompting guidance is refreshingly concrete: make the subject, action, style, and context explicit; put important constraints early; describe the result you want instead of leaning on negative prompts. It is the kind of advice that improves every model, but it is especially useful when you want a deliberate, designed frame.
Seedream 5 is the Seedream entry I keep returning to. It is the one I reach for when the work leans into references, edits, and richer visual instruction. The result feels less like asking a model for an aesthetic and more like directing a specific frame.
The new names that deserve a real test
Wan 2.7 Image belongs in this comparison. It is an image model, not the older Wan video entry people still associate with the name. Its pitch is exactly the kind of work I care about: tighter instruction-following, visual coherence, and better control when a character or scene has to survive more than one shot.
Kling Image 3 is another serious still-image option. Kling's own image guide centres its current model on natural-language generation, editing, and multi-reference control. That gives it a very practical lane in a DreamCraft workflow: take a good frame, combine the useful parts of several references, and refine instead of starting over.
Grok Imagine is in the scorecard strictly for image generation and editing. I am not judging its video side here. For stills, it is an interesting creative option when I want a fast, expressive interpretation of a natural-language brief, but it is not yet the default I would choose for every controlled production job.
The Lite lesson
Nano Banana 2 Lite is not here to lose a beauty contest. It is built for speed and scale. That makes it useful for rough concepts, bulk variations, and moments when the cost of thinking is higher than the cost of generating. But when I compare the output with the strongest models on a difficult creative brief, I can see the trade-off.
That trade-off is the whole reason a unified workspace is interesting. A good creative tool should not make you worship one model. It should help you choose the right one for the shot.
Image generation is finally less about finding one winner and more about having taste, a clear brief, and the freedom to route that brief to the right engine.
What this changes for DreamCraft
I have exciting plans for DreamCraft. I want it to be the place where the best image is not trapped behind one expensive model, five confusing dashboards, or the same generic AI slop everyone else is paying for.
The ambition is simple: give people a clean way to create genuinely great images at a fraction of the usual cost, with the right model, the right controls, and a workflow that respects their taste. Not more AI noise. Better pictures, faster decisions, and more room to make something that actually feels like yours.
Model capabilities, pricing, and names move fast. This is a dated record of my own tests and workflow rather than a permanent market ranking.