57 models, side by side.

Half a credit an image. The volume tier, for when you need twenty of something.

Black Forest Labs' workhorse, and one of the cheapest reference-capable models here.

FLUX.2 with more control per pass, priced a step above Pro.

The top of the FLUX.2 ladder, for when the render itself is the deliverable.

Half a credit, JPEG out. The other volume tier.

Higgsfield's own photographic model. Half a credit at 720p or 1080p, and no references.

OpenAI's instruction-follower, fast build. Six quality rungs, and the cheapest is a fortieth of the dearest.

The same model and the same price, tuned for the edit that has to change one thing and nothing else.

The previous GPT Image, kept on the shelf for work already locked to it.

xAI's image model. Two credits, and one picture in for an edit rather than a reference set.

The sharper Grok: 2K output, better text rendering, and a resolution that is the whole decision.

Plain-language image edits and text-to-image generation with up to eight references, at 1K or 2K with Fast or High thinking.

Google's latest fast image model: fourteen references, sharp text, and output up to 4K.

The fixed-1K Nano Banana variant, with seven references and fourteen aspect ratios.

The previous Nano Banana, available in Older versions for saved work and recipes.

The default image model: cheap, fast, takes seven references, and gives away its top resolution.

Seedream with more capability per pass, priced by size rather than given away.

Dense text, structured layouts and photographic detail from Alibaba: generate from text or edit one image at a fixed price.

The only way to get an editable vector out of this catalog. Logos, icons, flat illustration.

Recraft at 2048×2048: more path detail, at six times the price of the standard tier.

Six credits: re-syncs a person in existing footage to a new audio track.

A portrait and a voice track become a talking clip — people, animals or cartoons alike, at 720p.

Drives a still image with the motion of a video. Prompt optional, 720p or 1080p.

Kuaishou's flagship: seven references, a reference video, and 4K output.

Kling 3.0 without the reference channels: a prompt, a first and last frame, and 4K.

A photo plus an audio track becomes a talking clip. The prompt is optional.

Pruna's quality-focused clip model. Up to twenty seconds, native speech, and first-to-last-frame control.

Takes a still and a driving video, and moves the still the way the video moves.

Swaps the subject inside existing footage for one of your own.

Pruna's general clip model. Cheap, flexible durations, one reference audio.

The economical Seedance: same references and ceiling, capped at 720p.

Seedance 2.0 tuned for turnaround rather than for price.

Thirty reference images, ten reference videos, ten reference audios, and takes up to thirty seconds.

The previous Seedance generation, and the only one in its family that renders 4K.

The volume tier of the Veo family. Audio is compulsory; 1080p only runs at eight seconds.

Veo without references, at roughly a third of the price. The single-shot tool.

Google's flagship: generates dialogue and sound from the same prompt as the picture.

Google's other video model: sound with every take, and the only one here that edits a clip you already have.

Dub a long performance with a continuous voice-over.

fal's own retuning of MiniMax H3. Fifteen seconds, native audio, and almost nothing to configure.

Alibaba's Wan 3.0. The one clip model here that runs a full thirty seconds at 1080p, from a prompt or one first frame.

The same Wan 3.0 surface at roughly a third of the wait, for about 1.4 times the rate.

xAI's clip model. Needs an image to start from, and always generates audio.

Runway's model. 1080p only, five or ten seconds, six aspect ratios.

Alibaba's clip model. Prompt-only, five aspect ratios, up to fifteen seconds.

ByteDance's performance model: a portrait and an audio track, at 1080p with real body motion.

Black Forest Labs' clip model. Picture and sound from one plain-language prompt, for up to twenty seconds.

Higgsfield's motion transfer: up to eight images of your subject, performing a source video's motion.

Higgsfield's object swap: the source video stays, and the thing in it becomes the one in your images.

Thirty-three English-accent voices, priced for volume.

Four voices, sixteen languages, tuned for realtime delivery.

The directable one: emotion selector, speed control, twenty-four languages.

Expressive voice-over with inline audio tags, George, and your saved ElevenLabs voices.

Expressive voice-over directed by inline audio tags, with twenty-six preset voices and saved ElevenLabs voices.

Thirty voices across twenty-five languages, from Google.

Prompt-to-track, and the one music model here that sings.

A minute of instrumental scoring per take, from a model trained on licensed music.