Skip the trial-and-error with ImageFX prompts — get cinematic first try

How ImageFX prompts actually work

ImageFX is Google’s experimental image generator at labs.google/fx, tuned for photographic and cinematic output rather than illustration or stylised work. The model responds strongly to cinematographic vocabulary — lens specifications, film stocks, lighting language. Below is a short editorial primer on what makes an ImageFX prompt different from prompts for other models, and what to do with the output of this tool.

ImageFX rewards lens and lighting language more than other models

Where Midjourney leans atmospheric and Stable Diffusion leans technical-tag, ImageFX leans toward the language of photography. Specifying a lens (35mm, 85mm portrait, 24mm wide), a film stock (Portra 400, Kodak Gold, Cinestill 800T), or a lighting direction (golden hour, blue hour, harsh midday, single-source rim-lit) shifts the output more than the same details would on other models.

This is because ImageFX was trained heavily on photo-and-film captioned data. The model learned that “Portra 400” means a specific colour palette and grain structure; that “rim-lit” means a light source behind the subject; that “shallow depth of field with warm bokeh” is a single coherent visual style. Prompts that use this vocabulary produce noticeably more cohesive output than prompts that avoid it.

Generic prompts produce flat results

The failure mode on ImageFX is generic-prompt output — vague illustration-style images that look like stock graphic-design renders rather than photographs. This happens when the prompt has no cinematic anchors. Adding even one specific photographic cue (a lens, a time of day, a film aesthetic) usually shifts the output meaningfully.

The tool generates prompts with at least one strong cinematographic anchor by default. If the output still feels flat, the most useful edit is usually adding a specific lighting direction — side-lit by a single overhead source or back-lit, rim-lit edges — rather than adding more subjects or modifiers.

ImageFX vs Imagen — same family, different prompts

ImageFX and Imagen 3 share architecture but are tuned differently. Imagen handles structured multi-component prompts (subject / style / lighting as labelled sections); ImageFX prefers flowing cinematographic descriptions. A prompt that works on ImageFX often works on Imagen too, but the inverse is less reliable — Imagen-style structured prompts can read as over-specified to ImageFX.

The tool defaults to ImageFX-tuned output. For Gemini Imagen workflows, the Gemini generator produces a different format that better matches that model’s parsing.

What to edit before submitting

Three quick adjustments to most generated ImageFX prompts. Pick one lighting source. If the prompt names two lighting directions (“golden hour, side-lit by overhead light”), pick one. Conflicting light sources confuse the model. Match film stock to mood. Portra 400 reads warm and natural; Kodak Gold reads vintage; Cinestill reads cinematic-night. Pick the stock that matches your intent rather than leaving the default. Drop generic descriptors. Words like “beautiful,” “stunning,” or “cinematic” without a specific anchor are evaluative noise. Replace with the visible thing that makes it cinematic (single hard shadow, three-quarter angle, blown-out highlights).

For prompts in other models’ formats, see the Midjourney generator, Stable Diffusion generator, Flux generator, or the generator hub for the full comparison.