Edit Models vs Generate Models: A Selection Table
71 text-to-image, 42 edit-image, 30 image-to-image models — and only 3 do all three. A job-first way to pick the right one instead of shopping by brand.
Ask "what's the best AI image model" and you'll get a brand name back. Ask "what's the best model for turning this photo into a studio shot without touching the product" and the answer is a completely different model, often from the same provider. Generation and editing are not the same skill wearing different clothes — they're scored separately, catalogued separately, and in a lot of cases shipped as separate model entries entirely. Picking by operation instead of brand is the faster path to a model that actually does the job.
Two different rankings, for a reason
The clearest outside signal that these are distinct disciplines: Artificial Analysis runs a dedicated image-editing leaderboard, ranked separately from its text-to-image leaderboard. Their own framing draws the line plainly — text-to-image generation creates entirely new images from text descriptions, while image editing modifies an existing image based on instructions. Those are different inputs, different failure modes, and as it turns out, frequently different models even when the branding suggests otherwise.
What the taxonomy actually looks like
Versely's own catalog carries the same split structurally, not just by convention. Across 108 image models, 71 are categorised text-to-image, 42 edit-image, and 30 image-to-image — and a model can legitimately sit in more than one bucket, so those numbers overlap rather than sum to a clean total. Sixty-five of the 108 are generation-only: no edit or image-to-image category at all, meaning they were never built to take a photo as input. Thirty-seven sit at the opposite end — edit or image-to-image only, with no text-to-image category, meaning they can't originate an image from a blank prompt. Only three models in the entire catalog — Nano Banana 2, GPT Image 1.5 and Seedream 4.5 — carry all three categories at once. Everything else is a specialist in one direction or the other.
That last number is the one worth sitting with. Genuinely dual-purpose, do-everything models are the exception in this catalog, not the norm — which means the safe default assumption about any given model should be "specialist," not "generalist," until its category list says otherwise.
Same family, different SKU
The FLUX.2 family is the clean illustration of why checking the category list matters more than checking the name. Every FLUX.2 variant offers editing from text and multiple references in one model, at resolutions up to 4 megapixels — that's true of the underlying model family as Black Forest Labs ships it. But in Versely's catalog, "Flux 2 Pro," "Flux 2 Max" and "Flux 2 Flex" are all filed under text-to-image only; the editing capability lives in separate catalog entries — "Flux 2 Pro Edit" and "Flux 2 Max Edit" — carrying edit-image and image-to-image instead. Same underlying model family, two different selectable entries depending on the job, and picking the base name for an edit job is the mistake this whole piece is trying to head off.
It's not a FLUX-only pattern. Qwen Image (text-to-image only) and Qwen Image Edit (edit-image only) split the same way. GPT Image 1 Edit and Seedream Edit are separate catalog entries from their generation counterparts. Nano Banana Pro carries edit-image alongside text-to-image, while Nano Banana Pro Edit drops text-to-image entirely and adds image-to-image — related, but not interchangeable. Reading the suffix is a habit worth having; reading the categories array on the model itself is the one that doesn't fail you when the suffix is missing or misleading.
The two edit categories aren't the same job either
edit-image and image-to-image sound like synonyms and aren't. Image-to-image is governed by a strength or denoise setting — low values nudge the picture, high values keep only the rough layout and repaint everything else, which is the mode behind "make it look like a studio shot" restyling and behind generating variants of one hero image. Push that strength dial too high, though, and identity stops surviving — past roughly the halfway mark a face is being redrawn, not edited, and it's no longer the same person or product. edit-image models are closer to instruction editing: you describe the one thing that should change, and everything you didn't mention is meant to hold still. For "keep the product exactly as photographed, just change the background," an edit-image model targeting the described region is the safer bet than an image-to-image restyle where the strength setting decides how much of the original survives.
A job-first selection table
| The job | Category to filter for | What it does to the rest of the frame |
|---|---|---|
| A product shot with nothing to start from | text-to-image |
Nothing to preserve — there's no starting frame |
| Remove/replace one object, keep everything else | edit-image |
Everything outside the described change should hold |
| Restyle a whole photo (illustration, studio look) | image-to-image |
Structure survives, texture and identity fade as strength rises |
| Six variants of one hero composition | image-to-image, low strength |
Layout and pose persist; surface detail varies |
| Lock a product AND a face AND a set in one pass | edit-image with a high max_reference |
Everything not named in the prompt should hold — see multi-reference editing for the reference-count side of this |
A Versely walkthrough: picking by operation
Say the job is turning an uploaded product photo into three restyled lifestyle variants, then instruction-editing the winner to swap a background prop. That's two different tool calls, deliberately:
- Restyle pass —
generate_image_from_imagewithimage_urlset to the uploaded product photo, a model filtered toimage-to-image(Flux 2 Max Edit or Nano Banana Pro Edit both qualify), and a moderate strength so the product's proportions survive while the setting changes. - Targeted fix on the winner — a second
generate_image_from_imagecall against the chosen variant, this time on anedit-image-only model, with a prompt naming just the prop to swap and nothing else — "replace the mug on the table with a glass tumbler, leave everything else unchanged."
Two calls, two different category filters, because they're two different operations wearing the same "edit a photo" description. Running both through the same generic "edit this image" model without checking which category it actually claims is the fastest way to either lose the product's identity in step one or have step two quietly redraw the whole frame instead of just the mug.
Checking before you commit a shot list
/compare is the fastest way to put two specific models side by side once you know the category you need; /models is the place to filter the full catalog by category before you've picked candidates at all. The best AI image editing model ranking is pre-filtered to edit-image specifically, which skips the "generation-only model with an edit-sounding name" trap entirely.
FAQ
If a model is tagged both edit-image and image-to-image, which behaviour does it default to?
That's model- and interface-specific — check the specific tool's parameters (a strength or mask field is the tell) rather than assuming. Models like Nano Banana Pro Edit carry both tags because the underlying model can do either job depending on how it's called.
Why would a provider ship "Pro" and "Pro Edit" as separate entries instead of one model that does both? Different optimisation targets. A generation-only entry is tuned and priced for originating images from a blank prompt; the Edit entry is tuned for taking an existing photo as a strong input. Splitting them lets each be evaluated — and ranked — on the job it's actually built for.
Is a higher reference count relevant to this choice? Only once you've picked the category. Reference count decides how many source images an edit-capable model can hold at once — see multi-reference editing — but it doesn't tell you whether a model is edit-capable in the first place.
Does "image-to-image" always mean the output looks similar to the input? Not necessarily — that's what the strength setting controls. Near zero, yes, it's a nudge. Pushed high, the model is repainting from a rough layout only, and resemblance (including identity) can drop off fast.
Brand loyalty is a reasonable instinct for a provider whose failure modes you already understand. It's a bad instinct for picking between two entries from that same provider — check the category list, not the name on the card.