Text to Image Models: Which to Use, What They Cost, and How to Run Them

You need to pick from the text to image models on offer this week, and the pages you find rank them on looks or explain diffusion instead of telling you what an image costs. On LumozAI, 25 image models sit behind one endpoint at https://lumozai.com/api/v4/run, priced from $0.002 to $0.095 per generation, and you switch between them by changing one field in the request.

This page is for the person who has to commit to a model and a bill. It gives you the price of all 25, what each one is good for, the four things that decide what you actually pay, what a month costs at 500, 5,000 and 500 images, and the shortest working code to run one model or compare three on the same key.

Key facts at a glance

What a text to image model is, and why edit models belong in the same list

A text to image model takes a written prompt and returns a new image. Most models in this catalog use latent diffusion or flow matching: the model starts from noise and removes it step by step until it resolves into a picture matching the description. How good the hands look, whether the sign in the background spells the word correctly, and how long you wait all depend on which model you picked.

The catalog also lists editing and reference-image models next to the generators, and that's deliberate. An edit run costs the same as a fresh generation. FLUX 3 Image takes up to 10 reference images and charges $0.050 whether you describe a new scene or ask it to change a jacket from navy to olive. Flux 2 Pro takes eight references at $0.090, Krea 3 takes eight at $0.060. So when you budget for images, budget for edits too: they come out of the same balance at the same per-run rate.

The price table: 25 models by cost per image

Every price below is the per-generation rate on the image model catalog. The per-1,000 column is that rate multiplied out, which is the number you want when you're estimating a batch.

Under $0.030 per image

$0.030 to $0.045 per image

$0.047 and up per image

The cheapest run in this table costs about 47 times less than the most expensive: SDXL Lightning at $0.002 against Grok Imagine 2 at $0.095. That spread is the whole argument for matching the model to the job instead of defaulting to the most expensive one you've heard of.

What each model is good for

Grouped by the job, not by rank. A model that's wrong for your job at $0.002 is worse than the right one at $0.050.

Cheapest for volume

SDXL Lightning at $0.002 is the floor, and it's built for it: four-step inference, sub-second runs. Chippy at $0.015 and p-image-edit at $0.015 sit just above with more capability. Chippy is the high-throughput option, and the site's own benchmark puts it at 0.8s. p-image-edit handles multi-image edits in under a second, which makes it the one for processing an existing catalog rather than generating from scratch.

Use this tier for thumbnails, placeholders, storyboards, and anything a human will look at for two seconds.

Typography inside the image

qwen-image at $0.030, Nano Banana 2.1 at $0.035, and GPT Image 2.5 Flare at $0.045. These are the models you pick when the image has to spell something correctly: a product label, a shop sign, a poster headline. qwen-image is built for complex text rendering. Nano Banana 2.1 is specified for crisp on-image typography and legible signage, and takes up to 14 reference images. GPT Image 2.5 Flare goes further into non-Latin scripts, including native Bengali.

If your creative has a headline in it, this is the tier. Don't try to get there with a $0.002 model and a clever prompt.

Photorealism and portraits

Realify 2.0 at $0.050, imagen-4 at $0.050, and Lumoz Image at $0.055. Realify 2.0 is specified for hyper-realistic human anatomy and takes up to 10 reference images, so it's the one for headshots and people. Imagen 4 is the landscape and architecture pick; the site's own benchmark credits it with natural lighting and physics. Lumoz Image is the generalist at the top of this group, covering photorealism, typography and creative variations in one model.

Editing and reference images

FLUX 3 Image at $0.050 takes up to 10 references. Flux 2 Pro at $0.090 takes eight. Krea 3 at $0.060 takes eight and edits at 1K or 2K. The reference count matters more than the price here: if your job needs six reference images to hold a character consistent, the cheaper models that cap at one or two aren't options at any price.

Stylized

Grok Imagine 2 at $0.095 is the most expensive model in the catalog, and it's specified for dynamic stylized digital art, surreal concept renders and bold creative visuals. Pay it when the look is the product.

Resolution changes the price

Not every model charges a flat rate. FLUX 3 Image and Realify 2.0 both run $0.03 at 768sq, $0.05 at 1k, $0.07 at 1.5k, $0.09 at 2k and $0.15 at 4K. That's a fivefold spread on one model, so check the resolution tier before comparing it to anything else in the table. A 4K FLUX 3 Image run costs nearly three times a 1k Lumoz Image run. If your brief says 4K, budget $0.15 per image for that model, not $0.050.

The four things that actually decide your bill

Quality, measured as retry rate

The per-image price is not what you pay. What you pay is the per-image price times the number of attempts it takes to get an image you'll ship.

This is the part price tables leave out. A model that nails your prompt first time at $0.050 is cheaper than a model that needs four tries at $0.015, because the cheap one just cost you $0.060 and three extra minutes. Photorealism, typography and prompt adherence are separate qualities, and a model can be strong in one and weak in another. Pick the model that's strong in the quality your job depends on, then measure how often you regenerate.

Speed

Latency matters when a person is waiting. The site's own benchmark lists Chippy at 0.8s, FLUX Schnell at 0.9s, Lumoz Image at 1.8s, Imagen 4 at 2.4s, FLUX 3 at 3.2s and Realify 2.0 at 3.6s. Treat those as the site's numbers, not an independent test.

If you're generating in the background for a batch job, a 3.6s run costs you nothing but wall time. If a user is watching a preview render, 3.6s is a different product.

Price per image

At 5,000 images a month, SDXL Lightning costs $10 and FLUX 3 Image at 1k costs $250. Same volume, same endpoint, 25 times the bill. That gap is the whole reason to pick by job instead of by reputation.

Licensing

Outputs carry full commercial ownership and publishing rights, subject to each base model's open license terms. That last clause is the one to read. Each base model has its own license, and some are more permissive than others. If you're publishing commercially at scale, check the base model's terms for the one you picked, not just the platform's.

What a month costs at your volume

Three worked examples, no calculator needed.

500 cheap-tier images. 500 × $0.002 = $1.00 on SDXL Lightning. At Chippy's $0.015 it's $7.50. That's a side project's whole image budget for a month, and it's less than the $10 minimum top-up.

5,000 mid-tier images. 5,000 × $0.030 = $150 on qwen-image. At Nano Banana 2.1's $0.035 it's $175. At Lumoz Image's $0.055 it's $275. This is the range where a model choice starts to show up as a real budget line.

500 flagship hero images. 500 × $0.050 = $250 on FLUX 3 Image at 1k. At 4K, the same 500 images cost $750. At Lumoz Image's $0.055 it's $275. Hero images are few, so paying the top rate for them is usually the right call.

The point of running all three: mix tiers in one account. Run your volume work on SDXL Lightning, your typography work on qwen-image, and your hero shots on FLUX 3 Image. It's one balance and one key, so there's no reason to pick a single model for everything.

Run one model: the smallest thing that works

One endpoint, one model, one request. Send model_id and prompt to https://lumozai.com/api/v4/run with your key in the Authorization header.

curl -X POST https://lumozai.com/api/v4/run \
  -H "Authorization: Bearer $LUMOZ_API_KEY" \
  -H "Accept: application/json" \
  -d "model_id=flux-3-image" \
  -d "prompt=A ceramic coffee mug on a walnut desk, morning window light, shallow depth of field" \
  -d "aspect_ratio=1:1" \
  -d "resolution=1k"

The response gives you a URL to the generated image:

{
  "success": true,
  "data": {
    "request_id": 982348712938172,
    "output": "https://lumozai.com/storage/outputs/result_flux-3-image_9823.png",
    "output_type": "image"
  }
}

That's the shape for every image model on the platform. Parameters vary by model: FLUX 3 Image takes resolution, grounding and output_format, while Lumoz Image takes num_outputs. The endpoint, the auth header and the response shape don't change.

One thing to know before you build retry logic: failed or interrupted runs are refunded automatically, so a retry costs you time rather than money. Retry freely on a 500, and don't build a fallback path to avoid a model because you're afraid of paying for its failures.

Run several: one key, one balance, swap the model string

Every model in the table draws from the same balance and uses the same key. Comparing three models on one prompt is three requests, not three signups.

import os, requests

API_KEY = os.environ["LUMOZ_API_KEY"]
PROMPT = "Storefront sign reading OPEN LATE, neon, rainy street at night"

prices = {"sdxl-lightning": "$0.002", "qwen-image": "$0.030", "flux-3-image": "$0.050"}

for model_id, price in prices.items():
    r = requests.post(
        "https://lumozai.com/api/v4/run",
        headers={"Authorization": f"Bearer {API_KEY}", "Accept": "application/json"},
        json={"model_id": model_id, "prompt": PROMPT, "aspect_ratio": "16:9"},
    ).json()
    print(f"{model_id} ({price}): {r['data']['output']}")

Three runs, three prices, one bill. The output URLs land next to each other and you can look at all three before you commit to a model.

Compare before you commit: the Studio, then the MCP connector

The fast route is the Studio. Run one prompt across several models side by side in the browser, no code and no key to wire up. Look at the outputs, note which one handled your prompt, then write code for that pick. Comparing in a browser first saves you from building a pipeline around the wrong model.

If you'd rather generate from where you already work, the MCP connector runs the same models from Lumoz LLM or an AI editor. The server is at https://lumozai.com/api/mcp and connects to Lumoz LLM in the browser and on the desktop, plus Cursor and Windsurf, with an OAuth sign-in. The MCP setup page has the connector URL and the config snippets.

One step after that: most generated images aren't finished when they come out. A product shot usually needs its background gone. The Background Remover microservice runs at $0.005 per run on the same balance, so the cutout costs a tenth of a mid-tier generation. The transparent image maker runs it in the browser if you want to see the output first.

Questions people ask

Start with one prompt

Create an account, get your API key, and run the same prompt across three price tiers in the Studio: SDXL Lightning at $0.002, qwen-image at $0.030 and FLUX 3 Image at $0.050. You'll know which tier your work needs in about five minutes, and the whole test costs less than a dime.