Higgsfield, picked by model first

Higgsfield runs a stable of image models, not one. The single most common mistake on the platform is treating it like a one-model service and rewriting the prompt when the result is off.

A faster habit: when a generation reads wrong, switch the model before touching the prompt.

Three models that cover most days

nano_banana_2 is the default. It nails atmospheric backgrounds — color gradients, halftone grain, ink wash, smoke. Negative space around a subject lands clean. It is the model to reach for when the piece is environment-first, when the composition has to read at thumbnail sizes, when the work needs to be quiet.

soul_2 is for portraits. Faces, fabric, hair, fashion editorial. If the day's piece is character-led — a Shinjuku alley fit, a streetwear lean, an editorial close-up — soul_2 runs. The face holds at full zoom, which is the thing nano_banana_2 does not always do.

marketing_studio_image is for posters. Printed-object pretend. Album art, magazine covers, dust jackets, zine drops. The model lights its subjects like product shots and lays out negative space the way a designer would.

The standard misuse

Asking nano_banana_2 to render a character.

It will try. The face will read at a thumbnail. Zoomed in, the eyes are slightly wrong and the hands are worse. If the piece is going to sit on the gallery at full size, this is unacceptable. The fix is almost never to add a more specific prompt. The fix is to ask soul_2 instead.

A pattern that holds across all three

1. Decide what the piece is doing first. Cover art, portrait, atmosphere, poster.
2. Match the model to the call. Don't try to make the wrong model right.
3. Generate two. Pick one. Throw the other away. The third generation is rarely the keeper.
4. Composite type and logos by hand. Higgsfield is not a typesetter. Wordmarks, taglines, and credits go on top via Sharp or Photoshop after, never inside the prompt.

On gpt_image_2

gpt_image_2 deserves its own line. In 2026 it is the strongest model on the platform for text rendering — the only one that produces legible body copy and clean kerning inside the image. For daily art the cost equation rarely works (slower, more expensive, only one piece per day), and the MELTEN voice rarely needs type inside the artwork anyway. For the rare drop that does — a poster with a real subtitle block — it runs.

Where the pipeline wins

• Reference inputs accept third-party URLs directly. No staging upload step needed.
• Generation result IDs can be passed back as references in chained prompts. The whole run stays in the same workspace.
• The Marketing Studio mode produces multiple variants on a brief in one call. For a poster drop that needs three matching covers, it's a one-shot.

Where it loses

• No interactive in-paint at the time of writing. If a piece needs surgical edits to one region (a hand, a logo, a single line of text), the work shifts to a Photoshop round-trip.
• The marketing model occasionally lands too clean. The MELTEN voice runs grainy and printed-page; the editorial-clean version of a piece sometimes needs a roughing pass after.

Worked example

The og-home.png and og-about.png cards that sit on this site were generated through nano_banana_2 in two passes. The prompt asked for "deep matte black, soft violet and magenta light gradient bleeding in from the right edge, halftone dot texture, restrained moody composition, large negative space in the left half for a wordmark, no text, no characters". The model handed back two usable atmospherics on the first call. The MELTEN wordmark and the magenta tagline were composited on top in a 40-line Sharp script after — that part was never inside the prompt.

One rule, applied: pick the model first, the prompt second. Most bad days at the prompt are days where the wrong model is doing the work.