Using FLUX 1.1 Pro and FLUX.2 for Your Ad Visuals: Complete Prompt Guide (2026)

Because it gives you almost nothing to turn but the prompt. There is no sampling step count, no CFG or guidance scale and no negative prompt on any FLUX model, so the quality of the render depends on how well the sentence is written, and the same prompt behaves consistently from one generation to the next. That is also why FLUX follows a precise brief better than most of its competitors, and why its skin and material textures read as photographic.

The ratings below are the SociaLover editorial team's assessment on advertising use cases, re-checked in September 2026: not an independent benchmark, and not a reproducible measurement. FLUX.2 matches these ratings line for line, with text rendering as its clearest gain over 1.1 Pro. Treat them as a starting point and run your own brief through each model.

One brief, four vendors

FLUX.2, Nano Banana 2, GPT Image 2 and Seedream 4.5

The biggest change is reference images. FLUX 1.1 Pro generates from text only, so the same product came out slightly different in every render; FLUX.2 accepts up to four reference images in SociaLover, which is how a product photographed once stays identical across a whole set of scenes. For advertising that single feature matters more than any quality gain, because a campaign is a series, not one image.

Resolution moved in two directions. In SociaLover, FLUX.2 renders in 1K (4 tokens) or 2K (8 tokens), a clean draft-and-deliver pair. On the BFL API, FLUX 1.1 Pro Ultra still goes further, up to 4 megapixels, which remains the reason to keep it around for a print-size hero image or a portrait that has to hold up on a large screen.

Text rendering improved between the two generations, in our editorial re-check from a clear weakness to 7.8 out of 10, but FLUX.2 is still not the model you pick when a headline has to come out of the generation: GPT Image 2 and Nano Banana Pro do that better, and Ideogram 3 (not in SociaLover) better still. Read how FLUX.2 compares with GPT Image 2 and Nano Banana Pro before routing a creative that carries copy.

What did not change is the way you talk to it. Both generations read long, plain prose sentences; neither has a negative prompt, a CFG scale or a weighting syntax; the seed still gives you controlled variations. So FLUX 1.1 Pro still makes sense in two cases: a pipeline already built on the BFL API that you do not want to rewrite, and the raw mode on Ultra for faces that have to look unretouched. For everything else in SociaLover, FLUX.2 is the default.

In five blocks, in this order: subject, setting, light, photographic style, finish. FLUX follows instructions very precisely, so a poorly structured prompt produces a precisely mediocre result. Here is the structure that works on both generations:

The same skeleton, filled with the vocabulary of each industry, is what our 50 ad prompts by industry are built on. In SociaLover you paste the sentence into Creative Designer, pick FLUX.2 and the ratio, and add the product photo as a reference.

The prompt, the output geometry, the seed and a handful of switches, and nothing inherited from Stable Diffusion. FLUX 1.1 Pro is not a Stable Diffusion checkpoint: there is no sampling step count, no CFG or guidance scale, and no negative prompt. Here is what each parameter does on the Black Forest Labs API, and how to set it for advertising. In SociaLover these choices are reduced to the ratio, the resolution (1K or 2K) and up to four reference images; the seed and the switches below are API-level parameters. See the Black Forest Labs API documentation for the current list.

Note on upscaling: no FLUX model has an upscale parameter. If a close-up face needs more definition, generate it, then send the result through Image Enhancer. Asking the model for "8K" in the prompt does not change the output resolution; the size is set by the resolution setting or by width and height.

Fifteen prompts, one per ad situation, all written on the five-block skeleton above and all usable as-is on FLUX.2 in SociaLover or on FLUX 1.1 Pro through the API. Replace the bracketed placeholders with your own product, keep everything else, and generate three to five seeds before judging the prompt.

Lifestyle recipe

Clean product recipe

Fashion recipe, three scenes

Left, the lifestyle recipe: a person actually using the product, daylight through a window, authentic setting. Middle, the clean-product recipe (gradient background, dramatic side light, subtle shadow below), which is a lighting formula rather than a category, and transfers from a phone to a sneaker without a rewrite. Right, the fashion recipe held constant while only the surface changes: floating, plinth, draped fabric. That last one is exactly the 'lock the frame, vary the scene' technique described below. These are creatives produced on the platform, not the literal output of the prompt strings above. For the full product route, read our guide to product photography with FLUX.

The step-by-step version of that product route, from phone photo to export, is in product photography with FLUX.

Generation, editing and definition are three different jobs, and FLUX splits them across three tools: FLUX.2 (or FLUX 1.1 Pro on the API) generates from text, FLUX Kontext edits an existing render from an instruction, and Image Enhancer adds definition afterwards, since no FLUX model has an upscale parameter.