Want to Make AI Images? Which Should You Try First — A Year’s Worth of New Software and Services, Researched

September 29, 2026

The image at the top of each article on this blog is now something I make with generative AI on my own computer. Given how fast things move, which services are actually easy to use right now?

Both local image generation, where you make images on your own computer, and cloud image generation services, where you pay a monthly fee, have grown in number over the past year. Which should you try first? This article researches what’s available as of September 2026, covering both local image generation and cloud image generation services that make images for you over the internet.

This is content as of September 2026.

What I Looked Into — Scope and Sources

What I looked into is the image-generation models available as of September 2026, along with the paid services. I referenced the developers’ official blogs, model distribution pages, and each company’s pricing pages. Links are at the end of the article.

I haven’t included a comparison of image quality. This blog hasn’t run every one of these side by side to compare. What this article covers is limited to what can be confirmed with numbers and documentation: scale, storage requirements, pricing, and licensing.

Services like MiniMax that make videos are sometimes used to produce images too, but I haven’t covered that here.

What’s the Difference Between Local Image Generation and Cloud Image Generation Services?

Using AI to make images breaks down into two broad approaches. One is downloading a model’s files to your own computer and running it there. The other is sending instructions to a machine a company has set up and receiving the finished image back. The first is local; the second is cloud.

With local generation, what you need is your graphics card’s working memory. This is called VRAM, and whether the model fits into it determines whether it runs at all. With a 24GB graphics card, the idea is that you can run models up to whatever fits in 24GB.

With cloud generation, that constraint doesn’t apply. Instead, you pay per image or a flat monthly fee.

What the Model Distribution Sites’ New Releases Show

I looked at Hugging Face, a model distribution site, and its ranking of image-generation models by how much attention they’re getting right now (as of September 6, 2026). Two things stood out among what was released or updated between August and September 2026.

• LLaDA-Image (released September 4, 2026, 6 billion parameters, Apache 2.0). This handles both generation and editing in a single model, and the fast version reportedly produces an image in 4 steps
• A sudden wave of Krea 2 derivatives. The base Krea 2 model was released in June 2026, but fine-tuned and speed-optimized versions of it kept coming out one after another in the week leading up to September 6

Outside the distribution sites, AMD announced that ROCm 7.1.1 makes ComfyUI generation up to 5.4x faster (compared with ROCm 6.4: 2.6x for SDXL, 5.2x for FLUX.1 schnell, 5.4x for WAN 14b).

Lining all this up, I got the impression that the direction things are heading has become clear. Fast versions that produce an image in 4-8 steps are becoming standard, generation and editing are converging into single models, and support for hardware beyond NVIDIA is falling into place.

Local Image-Generation AI Models — Scale and Licensing

Here are the models you can download and run on your own computer. For the VRAM column, I’ve used the officially stated figure where one exists, and the distributed file size where it doesn’t. Steps are the number of computation passes needed to produce one image — fewer means faster.

Model	Scale	License	Commercial use	VRAM guideline	Steps
SDXL 1.0	6.5GB class	CreativeML Open RAIL++-M	Allowed	8GB+	20-30
Stable Diffusion 3.5	large / medium	Stability Community License	Depends on scale	12GB+	20-30
FLUX.1 schnell	12 billion	Apache 2.0	Allowed	12GB+	4
FLUX.1 dev	12 billion	FLUX non-commercial	Conditional	12-24GB	20-30
FLUX.2 dev	32 billion	FLUX non-commercial	Conditional	19.3GB (4-bit version) + more	20-30
FLUX.2 klein 4B	4 billion	Apache 2.0	Allowed	About 13GB	4
FLUX.2 klein 9B	9 billion	FLUX non-commercial	Conditional	Not officially stated	4
Z-Image Turbo	6 billion	Apache 2.0	Allowed	Fits in 16GB	8
Ideogram 4.0	9.3 billion	Ideogram Non-Commercial Model Agreement	Not allowed (separate fee)	Runs on one 24GB graphics card	20
Krea 2 Raw / Turbo	12 billion	Krea 2 Community	Allowed with conditions	Not officially stated	Turbo is 8
Anima Base v1.0	2 billion	CircleStone Labs Non-Commercial	Not allowed	Not officially stated	—
LLaDA-Image / Turbo	6 billion	Apache 2.0	Allowed	Not officially stated	50 / 4

Which One Gets Used the Most?

I noted the cumulative download counts shown on the distribution site as of September 6, 2026.

Model	Downloads	Last updated
SDXL 1.0	1.77 million	October 2023
FLUX.1 dev	761,000	June 2025
FLUX.1 schnell	713,000	August 2024
Z-Image Turbo	677,000	January 2026
Qwen-Image	308,000	August 2025
Stable Diffusion 3.5 medium	153,000	October 2024
Krea 2 Turbo	76,800	July 2026
Ideogram 4.0 (fp8)	59,000	June 2026

It doesn’t look like the newest model is necessarily the most-used one. SDXL, from 2023, holds a number an entire order of magnitude higher. My guess is this comes from how much more third-party tooling has accumulated around SDXL, but I haven’t dug fully into that.

How Small Can Quantization (Lowering Precision to Shrink a Model) Get?

There’s a technique for reducing storage size by lowering the numeric precision of a model’s values, called quantization. Here are the actual file sizes of the versions distributed for FLUX.2 dev.

Version	File size
Unmodified (BF16)	64.4GB
8-bit (Q8_0)	35GB
6-bit (Q6_K)	27.4GB
5-bit (Q5_K_M)	24.1GB
4-bit (Q4_K_S)	19.3GB
3-bit (Q3_K_S)	15.8GB
2-bit (Q2_K)	12.9GB

Going from 64.4GB down to 19.3GB looks like it should fit in the memory of a 24GB graphics card. But the distribution page states that the component that interprets the text prompt needs to be downloaded separately. For FLUX.2 dev, that’s Mistral-Small-3.2-24B, which is itself large. That part is meant to be offloaded to the computer’s system memory, so simple subtraction doesn’t get you to “it fits."

The same structure shows up elsewhere: Ideogram 4.0 separately requires Qwen3-VL-8B, and FLUX.2 klein 9B separately requires an 8-billion-parameter Qwen3. The “VRAM guideline" column in the table above doesn’t include this extra part, which is worth keeping in mind.

How Much Do Cloud Image Generation Services Cost?

Here are the paid cloud image generation services. Prices are the values listed on each company’s official page as of September 2026, in US dollars.

Service	Billing type	Price	Commercial use
Nano Banana Pro (Google)	Per image	$0.134/image at 1K/2K, $0.24/image at 4K	Allowed
Seedream 4.0 (BytePlus)	Per image	$0.03/image. New accounts get 200 free images	Terms not confirmed
Midjourney	Monthly	$10 / $30 / $60 / $120	Allowed (companies with over $1 million in total revenue need $60+)
Adobe Firefly	Monthly + credits	$9.99 (2,000) / $19.99 (4,000) / $34.97 (10,000) / $139.91 (50,000)	Allowed
ChatGPT	Monthly	Free / $8 / $20	Allowed (output rights go to the user)
Canva	Monthly	Pro runs around $15-18/month	Terms not confirmed
Comfy Cloud	Monthly + credits	$20 (4,200) / $35 (7,400) / $100 (21,100)	Allowed (subject to the terms of whichever model you use)

Converting per-image pricing into a monthly image count makes it easier to get a feel for the numbers. At Seedream 4.0’s $0.03/image, 300 images a month comes to $9. At Nano Banana Pro’s $0.134/image, the same 300 images comes to $40.20. That gives you a baseline for comparing against the monthly-fee services.

What Does Adobe Firefly’s “Rights-Cleared" Claim Mean?

Adobe officially states that Firefly’s output is “safe for commercial use." Their explanation is that training used only material with cleared rights and material whose copyright has expired. Enterprise contracts reportedly come with compensation coverage if a rights issue arises.

That said, this claim only covers Adobe’s own models. The Firefly app also lets you choose models from Google or OpenAI, and choosing one of those takes you outside Adobe’s guarantee. The official page itself states that whether output from another company’s model suits your intended use is left to the user’s own judgment.

Are “Open Source" and “Open Weight" the Same Thing?

This is where I got the most confused while researching. Models you can download for free tend to get lumped together as “open source," but the conditions aren’t all the same.

• Open source — published under a license with very few restrictions, like Apache 2.0 or MIT. FLUX.1 schnell, FLUX.2 klein 4B, Z-Image Turbo, and LLaDA-Image fall here
• Open weight — the files can be downloaded, but usage comes with conditions attached. FLUX.2 dev, FLUX.2 klein 9B, Ideogram 4.0, Anima, and Krea 2 fall here

Does FLUX’s Non-Commercial License Draw a Different Line for the Model Versus Its Output?

I read the original text of FLUX’s non-commercial license. It draws a distinction between the model and the images that come out of it.

• The model itself — what’s permitted is research, experimentation, personal learning, and hobby projects. Use in revenue-generating activity, or use in a service that interacts directly with users, is explicitly excluded
• The output — it states, “You may use Output for any purpose (including for commercial purposes)," meaning the resulting images can be used commercially. The only thing prohibited is using that output to build a competing model

As I read it, the line being drawn is that you may sell the image itself, but running the model as part of a revenue-generating activity steps outside the non-commercial scope. Using such an image on a site that carries ads or affiliate links may run up against that line. This blog’s own policy is to use the schnell version if we use FLUX at all. This isn’t legal advice, so if you’re concerned, check the original text yourself.

The License Conditions for Ideogram 4.0 and Anima

Ideogram 4.0 is known for rendering text cleanly inside images. It’s published in two tiers: the program that runs it is Apache 2.0, but the model file itself sits under a non-commercial agreement. The official license page states that commercial use requires Self-Serve or Enterprise, with Self-Serve billed annually at $300/month for a range of 10,000-100,000 images a month.

Anima Base v1.0, specialized for anime-style output, is small at 2 billion parameters and easy to approach. Its license is non-commercial, though, and since it’s built on an NVIDIA model as its base, that model’s conditions apply as well.

The distribution page for Kroma, a fine-tune of Krea 2, spells this out clearly: “this fine-tune is MIT, but it grants no rights to the underlying base weights." Using a derivative version doesn’t loosen the conditions.

Which Image-Generation AI Runs on Your PC’s Graphics Card?

Organized by capacity. The speed and power-draw figures in the left column are values I measured myself on 5 graphics cards under matched conditions on September 6, 2026. That measurement was done with a text-generation model, though — not with image generation. The two right-hand columns are what I researched.

VRAM	My blog’s matching card (write speed [tok/s] / power while generating [W])	Looks like it runs	Tight on memory
8GB	None on hand	SDXL, FLUX.1 schnell, a scaled-down Z-Image Turbo	FLUX.2 dev, Ideogram 4.0
12GB	RTX 3060 12GB (54.5 / 180W)	SDXL, Stable Diffusion 3.5 medium, the FLUX.1 family	FLUX.2 klein 4B falls just short at about 13GB
16GB	RTX 5060 Ti 16GB (74.8 / 161W) RX 9060 XT 16GB (54.5 / 159W)	Z-Image Turbo, FLUX.2 klein 4B, a scaled-down Krea 2 Turbo	FLUX.2 dev’s 4-bit version (19.3GB)
24GB	RTX 3090 24GB (119.9 / 344W)	Ideogram 4.0, FLUX.2 dev’s 4-bit version (with the text component offloaded to system memory)	The unmodified FLUX.2 dev (64.4GB)
128GB unified memory	Radeon 8060S (41.0 / 85W)	Capacity is essentially not a constraint	Speed

For small PCs with 128GB of unified memory, a third-party review reports that ComfyUI recognized roughly 90GB as usable capacity. That’s well above what a 24GB graphics card offers. But the same report also noted that speed didn’t reach that of an RTX 3060.

Here are two examples of 16GB graphics cards (both are cards I use for measurements on this blog).

As an Amazon Associate we earn from qualifying purchases.

How Fitting in a Graphics Card’s Memory Relates to Speed for Image-Generation AI

My own measurements on this blog show the same pattern. The Radeon 8060S’s write speed came to about a third of the RTX 3090’s. Power draw, on the other hand, was 85W versus 344W — a quarter as much. The high-capacity Radeon 8060S is slow, and the fast RTX 3090 uses a lot of power.

So what changes when there’s enough capacity? I have an earlier article that measured exactly this. Running the same setting 12 times in a row in ComfyUI, SDXL, which fits in VRAM, came out at a flat 6.0 seconds from the second run onward, with 0.0% variance. A larger model that doesn’t fit swung from 32.1 seconds to 72.1 seconds, a variance of ±25.2%. For the model that didn’t fit in VRAM, the model was being loaded and unloaded 14 times over 146 seconds.

Whether it fits or not changes both speed and stability. When you look at the “looks like it runs" column in the table above, picking something that fits with room to spare, rather than something that barely fits, looks like the easier choice to work with.

The detailed method and numbers are in the article below.

Local or Cloud — Three Forks in the Road

How Many Images a Month Are You Making?

While the count stays low, per-image billing works out cheaper. At 30 images a month, that’s $0.90 with Seedream 4.0, or $4.02 even with Nano Banana Pro. There’s no reason to buy a graphics card at that point. Once the count climbs past a few hundred and keeps up every month, local image generation comes into view as an option.

Is There Material You Don’t Want Going Out Onto the Internet?

With a cloud service, both your instructions and any reference images you provide pass through another company’s machine. If your source material includes people’s faces, unreleased products, or internal documents, local image generation, which stays entirely on your own computer, has the advantage. This is a separate axis from image count or price.

What Are You Going to Use the Images For?

If it’s just for your own enjoyment, licensing rarely becomes an issue. Once you’re using it for work, publishing it on a site with ads, or turning it into a product, the licensing conditions covered above start to matter. For local image generation, Apache 2.0 models are the easiest to work with; among cloud image generation services, some specifically sell themselves on handling rights clearance for you.

What I Haven’t Confirmed in This Article

• I haven’t compared image quality. This blog hasn’t run every model side by side, so I haven’t written about which one produces the nicest-looking images
• I haven’t measured image-generation speed. The speed and power-draw figures in the body of this article were measured with a text-generation model
• For the cloud image generation services, I only read the pricing pages — I haven’t actually paid and used them myself
• I read the original license texts, but this isn’t legal advice. If you’re concerned, consult the original text and a professional
• The claim that a 128GB unified-memory machine could use about 90GB comes from a third-party review, not a measurement done on this blog

Summary — What This Comes Down to as of September 2026

Local image generation has moved toward fast versions that produce an image in 4-8 steps as the norm, and the required capacity keeps coming down. With a 16GB graphics card, Z-Image Turbo and FLUX.2 klein 4B come into range, and both are Apache 2.0. That said, being free to download doesn’t guarantee commercial use — Ideogram 4.0, for example, requires a separate fee. Cloud image generation services range from $0.03 to $0.24 per image, so you need to multiply out your monthly image count before comparing against a flat monthly plan.

So which should you try first, going by what this article found? If your computer doesn’t have a graphics card, or you’re only making a few dozen images a month, starting with a pay-per-image online service is the easy route. If you have a 16GB graphics card, you can start local image generation with Z-Image Turbo or FLUX.2 klein 4B, both published under Apache 2.0. If you’re working with material you don’t want going out onto the internet, such as people’s faces or internal documents, local image generation is the better fit.

What surprised me while researching this is that the choice narrows down over two unglamorous factors, capacity and licensing, before image quality even comes into play. The fact that capacity and speed move independently was the same story as running a text-generation model on your own computer. In both cases, the first thing to check is whether it fits within your own graphics card’s capacity.

How large a model can actually run on my own computer — the record of where I got tripped up measuring that is in the article below.

Gear Used in This Article

References

• Black Forest Labs, “FLUX.2" https://bfl.ai/blog/flux-2
• Black Forest Labs, “FLUX.2 [klein]" https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence
• FLUX.2-dev model page and license text https://huggingface.co/black-forest-labs/FLUX.2-dev
• FLUX.2-klein-4B https://huggingface.co/black-forest-labs/FLUX.2-klein-4B
• FLUX.2-dev quantized versions (file sizes) https://huggingface.co/city96/FLUX.2-dev-gguf
• Z-Image-Turbo https://huggingface.co/Tongyi-MAI/Z-Image-Turbo
• Ideogram 4.0 technical overview https://ideogram.ai/blog/ideogram-4.0/
• Ideogram licensing https://ideogram.ai/licensing/
• Krea 2 Turbo https://huggingface.co/krea/Krea-2-Turbo
• Kroma https://huggingface.co/lodestones/Kroma
• LLaDA-Image https://huggingface.co/inclusionAI/LLaDA-Image
• Anima (ComfyUI documentation) https://docs.comfy.org/tutorials/image/anima/anima
• Google, “Nano Banana Pro" https://blog.google/innovation-and-ai/products/nano-banana-pro/
• Gemini API pricing https://ai.google.dev/gemini-api/docs/pricing
• BytePlus ModelArk, “seedream-4.0" https://docs.byteplus.com/en/docs/ModelArk/1824718
• Adobe Firefly plans https://www.adobe.com/products/firefly/plans.html
• OpenAI terms of use https://openai.com/policies/terms-of-use/
• Comfy Cloud pricing https://comfy.org/pricing
• AMD’s ROCm and ComfyUI (GIGAZINE) https://gigazine.net/gsc_news/en/20260107-amd-comfyui-rocm/