Skip to content

Type a term, e.g. API, Shopware or SEO.

    ← All posts
    Artificial Intelligence

    By René Mattis · vona ·

    Create AI Images for Free: Local, Free Tiers or via API

    High-quality AI images for print, web and social media for free or a few cents each: local models, free tiers and the API compared.

    Good AI images do not have to be expensive. There are three routes, and they differ considerably in cost, effort and legal certainty:

    • Locally on your own computer: genuinely free, apart from hardware and electricity. You need a suitable graphics card or a recent Mac and a model whose licence allows business use.
    • Free tiers from providers: quick to try, but tightly limited. Many free tiers do not allow business use or use your input for training.
    • Directly through the providers’ interface (API): not free, but cheap. An image in social media size costs from around 3 cents at Google, and from just under 2 cents in batch, with no subscription.

    Which route suits you depends on how many images you need, in what quality, and whether customer or product data is involved. The details, as of 29 October 2026:

    What resolution do you really need?

    Before looking at costs, it is worth checking the size you need. Image models usually charge by the tiers 1K, 2K and 4K, and the tier determines the price.

    Use Typical size Sufficient
    Social media (feed 4:5, story 9:16) 1080 × 1350 or 1080 × 1920 pixels 1K to 2K
    Website (hero image, blog) 1600 to 2400 pixels wide 2K
    Print A4 at 300 dpi 2480 × 3508 pixels (around 8.7 megapixels) 4K or upscaling

    For social media and the web, 1K or 2K is almost always enough. For print, you need either a model with 4K output or you upscale a smaller image. Real-ESRGAN is a good option for this: it is free, and its licence (BSD-3) allows business use. Be careful with popular add-on models for upscaling programs: some, such as “UltraSharp” or “Remacri”, are under a licence that excludes commercial use. For print files, you then convert the image to the printer’s colour profile in your layout program, because image models deliver RGB.

    Route 1: Local and genuinely free

    With open models, image generation runs entirely on your computer as a local model. There are no fees, and no images or drafts leave the building. That is particularly interesting if you work with unreleased products or client motifs.

    What you need: According to the manufacturers, current compact models run on NVIDIA graphics cards with 8 to 16 GB of graphics memory. On a Mac with Apple silicon, the free app Draw Things is the easiest way in; it works offline and supports the common open models. On Windows, macOS and Linux, InvokeAI (a clear interface with a canvas) and ComfyUI (node workflows, more control, steeper learning curve) are good choices. Ollama has also been able to generate images since early 2026, but according to the developers only experimentally on the Mac and via the command line.

    The biggest trap is the licence. “Open weights” does not automatically mean “free to use”. For a company, there are three groups:

    Licence Models (examples) Usable for business?
    Apache 2.0 or MIT Z-Image Turbo, FLUX.2 Klein 4B, Qwen-Image-2512, HiDream-O1-Image Yes, with no revenue limit
    Community licences with a revenue limit FLUX 3 Image Klein (up to 5 million US dollars annual revenue), Krea 2 and Stable Diffusion 3.5 (up to 1 million US dollars) Yes, below the limit. The revenue of the whole group counts, and there are additional obligations such as reviewing the images
    Non-commercial or regionally restricted FLUX.2 Dev and Klein 9B, Qwen-Image 2.1 (research only), HunyuanImage (licence does not apply in the EU) No, or only with a separate licence from the manufacturer

    Particularly tricky: with some non-commercial licences, the generated images may be used commercially, but the model itself may not be run for revenue-generating purposes. For a company, that is effectively no permission at all. So read the licence on the model page before using a model, not just the summary in a forum.

    Who the local route is worthwhile for: anyone who needs images regularly, already has suitable hardware and is willing to spend a little time learning. For photorealism and text in images, the best cloud models are often still ahead; for social media motifs, backgrounds and illustrations, open models are frequently good enough.

    Route 2: Providers’ free tiers

    Many providers let you generate AI images without a subscription. For private experiments, that is ideal; for business images, the small print matters. We read the terms of use of the well-known services:

    Service Free Usable for business? Note
    Adobe Firefly daily generations (Adobe does not state a number) Yes, for features without a beta label according to Adobe, no training on customer content; partner models cost credits
    Canva Free up to 20 AI uses a month Yes training only with consent
    Google Gemini (app) daily quota not excluded visible watermark, download only in 1K; with “Keep Activity” on, input is used for training
    ChatGPT Free limited (OpenAI does not state a number) Yes training on by default, can be switched off
    Grok limited Yes visible watermark, training on by default
    Ideogram, Leonardo weekly or daily credits Yes images on the free tier are publicly visible; at Leonardo, ownership stays with the provider
    Microsoft Copilot, Qwen Chat, Recraft Free monthly or daily credits No personal use only according to the terms; on Recraft Free, the images belong to Recraft

    And Google AI Studio? According to Google, the web interface of Google AI Studio is free, and for users in the EEA the rule is: even with free use, input is not used for training. Images from AI Studio only carry the invisible SynthID watermark. However, Google does not state a fixed free quota for the image models, and the programming interface offers no image models at all on the free tier. AI Studio is therefore well suited to trying out prompts, not as a reliable free source for a whole campaign.

    Our assessment: For individual motifs, Adobe Firefly and Canva are the most solid free routes, because business use is expressly allowed and your content is not used for training. Free tiers with training or a public gallery are not suitable for client projects, unreleased products or photos of people. And for print, none of the free routes demonstrably offers enough resolution; most providers do not state any figures at all.

    Route 3: Directly via API, for a few cents per image

    If you need images regularly, the providers’ direct interface is usually the cheapest option. You pay per generated image, with no subscription and no minimum volume. List prices according to the providers’ pricing pages (as of 29 October 2026, in US dollars, excluding tax):

    Model 1K 2K 4K Batch
    Nano Banana 2.1 (Google) $0.034 $0.050 $0.113 −50%
    Nano Banana 2 Lite (Google) $0.034 – – −50%
    Nano Banana Pro (Google) $0.134 $0.134 $0.24 −50%
    FLUX 3 Image Klein (Black Forest Labs) $0.009 $0.014 $0.036 no discount
    FLUX 3 Image (Black Forest Labs) $0.048 $0.10 $0.607 no discount
    Recraft V4.1 $0.035 (vector $0.08) Pro $0.21 – –

    OpenAI charges for its new GPT Image 2.5 image models by token and does not state a fixed price per image; the cost depends on size and quality level. At Google, each image costs half via the batch interface if the result does not have to be available immediately, for example for a series overnight.

    Two sample calculations with Nano Banana 2.1 at list price:

    • 100 social media motifs in 1K: around $3.40, in batch around $1.70.
    • 20 motifs in 4K for print: around $2.30, in batch around $1.15.

    Allow for a multiple of that for corrections, because the first image is rarely right. Even then, you stay far below the cost of most subscriptions.

    What to look out for when choosing a provider:

    • Training: Google and OpenAI do not use data from the paid API for training. At Black Forest Labs, the API terms expressly allow input and output to be used for training, including for EU access. For confidential motifs, that rules it out.
    • Data location: OpenAI offers processing in Europe for images, subject to approval. At Google, EU data residency is available via Vertex AI, not via the simple Gemini API.
    • Cost control: Set a monthly spending cap with the provider. According to Google’s documentation, it takes effect with a delay of a few minutes.

    Your own interface instead of a platform subscription

    The interface on its own is just programming access. For someone without programming skills to generate images in everyday work, you need a front end: a small script for series, a simple web interface or your own tool. The advantage over credit-based platforms: you only pay for the model, and the prompts can be tailored to your workflow. That also reduces the number of failed attempts. How to write good instructions is covered in our prompt engineering tips.

    We built this for LANAKILA, the sportswear brand of our owner, with RENEKI. The desktop app for macOS and Windows talks directly to Google’s image models. Prompt building blocks for camera, lighting and background can be combined with a click, projects can be shared in the team, and finished images go straight into the online shop. What the switch from a platform subscription to our own solution meant per image, with real figures, is in the article AI Product Images: Why the Subscription Cost More Than Our Own Solution. RENEKI is also available for other companies; we set it up and adapt it to your team, your image models and your shop.

    Your own solution pays off when images are produced regularly, several people work with them or the images are part of a fixed workflow. For a few motifs a quarter, a free tier or AI Studio is often enough.

    • Copyright: In Germany, images generated purely by AI are generally not protected by copyright, because there is no human creation. Others may therefore be able to reuse them. Conversely, and more importantly: prompts that deliberately imitate protected characters, brands or the style of specific works can infringe third-party rights.
    • Labelling: The transparency obligations of the EU AI Act have applied since 2 August 2026. Anyone who publishes realistic-looking AI images of real people, places or events must label them as artificially generated. Most providers also embed invisible markers (SynthID, C2PA); at Microsoft, removing them is expressly prohibited. More on our EU AI Act page.
    • Product images: An AI image must show the real product. Colours, cut and details that are wrong are annoying and can be misleading under competition law.

    Which route suits you?

    1. An occasional motif for social media: Adobe Firefly or Canva on the free tier, trying out prompts in Google AI Studio first.
    2. Regular images, confidential motifs, suitable hardware available: locally with Z-Image Turbo or FLUX.2 Klein 4B in Draw Things, InvokeAI or ComfyUI.
    3. Regular high-quality images, including for print: directly via API, for example Nano Banana 2.1 in 4K, with series in batch.
    4. Many images in a team, with a fixed workflow: your own interface built on the API, like RENEKI.

    Matching tools from our directory

    In our AI tools directory, we have checked the features, prices and data protection of the tools against the providers’ own information. Relevant to this topic:

    • Google AI Studio: test Google’s image models and create API keys
    • Adobe Firefly: image AI with its own models for commercial use
    • FLUX (Black Forest Labs): image models from Freiburg, via API or partly with open weights
    • ComfyUI: image AI locally as a node workflow
    • Ollama: run open models locally, image generation experimental on the Mac
    • Midjourney: subscription image generator, strong on artistic styles

    An overview of current image models with sources is available under AI models for image and video.

    If you want to find out which route is the most cost-effective for your company, or would like your own tool like RENEKI: under AI consulting we review existing AI processes and costs, and under app development we build custom tools. Or get in touch directly for a free initial consultation.

    As of 29 October 2026. API prices according to the pricing pages of Google (Gemini API, updated 28 October 2026), Black Forest Labs and Recraft; free tiers and terms of use according to the providers’ help, pricing and legal pages; licences according to the model pages on Hugging Face. Prices and terms change frequently; the providers’ information is authoritative.

    ← Back to overview

    Newsletter

    News from the vona workshop.

    AI tools we actually use, our weekly AI review & insights from our projects – short, practical and without spam.