How Higgsfield Does Image Generation: One Call, Dozens of Models

Image generation is the foundation of Higgsfield. It is the cheapest and fastest thing the platform makes, and it is the seed for almost everything else: the videos, the 3D models, and the explainer clips all start from an image. What makes Higgsfield’s image side unusual is that it is not one model — it is a large catalog of models from many providers (Google, OpenAI, Bytedance, Black Forest Labs, xAI, Kling, Recraft, and Higgsfield’s own Soul family), all reached through a single call.

This article explains that catalog: the one call behind every image, the two modes (text-to-image and image-to-image), how to pick a model by the job, and how images connect to the rest of Higgsfield.

The one call behind every image

Every image comes from one tool, generate_image, with the same handful of inputs regardless of which model runs:

• a model id (or let Higgsfield choose — see Auto below),
• a prompt describing the image,
• optionally a reference image (for editing or style),
• optional params like aspect_ratio, resolution, and how many variations to return (count, 1 to 4).

You do not have to set most of these. If you leave aspect_ratio or resolution unset, the server applies the model’s default and tells you what it used, so a bare “make me a red bicycle” still produces a sensible image. Under the hood it submits an async job and returns the picture in a few seconds.

Two modes: from text, or from an image

Almost every model works two ways:

• Text-to-image. A prompt alone. “A red bicycle leaning against a white wall.”
• Image-to-image. A prompt plus a reference image, passed as a media input. This is how you edit an existing image (“change the wall to brick”), transfer a style, or carry a face or product into a new scene. Editing-focused models such as Flux Kontext, Seedream, and OpenAI’s Hazel are built for this.

The difference matters for cost and control: when you already have an image that is close, editing it is faster and more predictable than re-rolling a fresh prompt until it lands.

Pick a model by the job

The catalog is big, but you rarely choose by hand — the assistant or the app picks for you, and there is an Auto model that routes by prompt intent. Still, knowing the families tells you what Higgsfield is good at:

If you take one thing from this table: choose by what you are making, not by which model is newest. A logo wants Recraft or Hazel; a fashion portrait wants Soul 2.0; a quick concept wants a budget model or Auto.

Resolution and aspect ratio

Most models render at 1K by default and can go to 2K or 4K (a couple, like Seedream at high quality, reach roughly 6K). The Soul models use quality tiers shown as 1.5K or 2K instead. Higher resolution costs more credits, so raise it only when you actually need the pixels.

Aspect ratios are broad and shared across most models: square 1:1, landscape 16:9 and 3:2, vertical 9:16 and 4:5, ultrawide 21:9, and more. The vertical ratios are what you want when the image will feed a Short or a phone-first video.

Soul: consistent people across many images

The Soul family is Higgsfield’s answer to the hardest part of AI images — keeping the same person looking the same across many generations. Soul 2.0 produces realistic people out of the box, and it accepts a soul_id: the id of a trained identity. Train a Soul once from a set of photos and every future Soul image can carry that exact face. This is the image-side half of the avatars feature; the overview guide covers training. For a one-off character you skip the id and let the prompt (or a reference image) define the look.

The call in practice

Through a Higgsfield-connected assistant, you just describe the image:

Make a photoreal image of a red bicycle leaning against a white wall, 4K.

The assistant picks a model and calls the tool. Conceptually:

generate_image
  model: "nano_banana_pro"
  prompt: "a red bicycle leaning against a white wall, photoreal"
  resolution: "4k"
  aspect_ratio: "3:2"

It submits the job, waits, and returns the image URL. If you had omitted resolution and aspect_ratio, the response would tell you the defaults it applied (for example, 1k and 3:4) so there is no guessing about what you got.

Cost and speed

Images are the bargain of the platform. A one-thousand-pixel Nano Banana Pro image priced at just 2 credits when this guide was written, and it renders in seconds. Costs rise with resolution, quality tier, and batch size (some ad models generate many images per job), and you can ask for up to four variations in one call. As always, you can preview a specific job’s cost before committing by asking the assistant to estimate it, which spends nothing.

How images feed the rest of Higgsfield

Image generation is rarely the end of the line. The same picture you make here is the input to most other Higgsfield tools:

• Image to video — animate a still into a clip.
• Image to 3D — lift a character or object into a GLB mesh.
• Style key — one locked image drives the look of every clip in a narrated explainer.
• Ads — a product image plus an avatar becomes a marketing video.

Because of this, a good habit is to get the image right first — the character, the props, the framing, the style — and only then move it into video or 3D. Fixing something in a cheap, fast image beats fixing it in an expensive, slow video.

Tips

• Start with Auto if you do not know which model fits; switch to a specific one once you know the job.
• Edit, do not re-roll. If an image is close, use image-to-image to change the one thing that is wrong instead of generating from scratch.
• Use a Soul for recurring faces, a one-off reference for a single scene.
• Raise resolution late. Draft at 1K, then regenerate or upscale the keeper to 2K or 4K, so you are not paying 4K prices for throwaways.
• Match the model to text needs. For anything with readable words, logos, or diagrams, pick a text-strong model (Nano Banana Pro, Hazel, Recraft) rather than a general one.

Recap

Higgsfield image generation is one call, generate_image, in front of a large catalog of models from many providers, working from text or from a reference image. You pick a model by the job — people, product ads, logos, cinematic stills, or just Auto — set resolution and aspect ratio only when you care, and pay a few credits for a result that arrives in seconds. Get the image right first, because it is the seed for the video, 3D, and ads that follow.