The short answer (July 2026): The six tools practitioners actually use split into three categories: subscription apps (Midjourney starting at $10/month, Adobe Firefly starting at $9.99/month), pay-as-you-go APIs (OpenAI gpt-image-2 from $0.005/image, Black Forest Labs FLUX.2 from $0.014/megapixel, Google Imagen 4 at roughly $0.02/image via Vertex AI), and open-weights models you run locally (Stable Diffusion 3.5, FLUX.2 [dev], Ideogram 4.0). Read the license before you self-host, though: only Stable Diffusion 3.5 is free for commercial use out of the box (under a revenue cap); FLUX.2 [dev] and Ideogram 4.0's open weights are non-commercial unless you buy a separate license. Midjourney V8.2 leads on artistic output, Ideogram 4.0 leads on text rendering inside images, Adobe Firefly leads on commercial IP safety, and Stable Diffusion 3.5 costs nothing per image once your GPU is paid for. Pick by workflow and output type, not benchmark rankings alone.
Last verified: July 27, 2026, against official docs and pricing pages.
What you're actually choosing between
Three different business models define this category, and your choice of access model matters as much as your choice of tool.
Subscription apps (Midjourney, Firefly standalone) give you a monthly credit or GPU-hour allowance. Right for daily generation with no infrastructure overhead.
Pay-per-image APIs (gpt-image-2, FLUX.2, Imagen 4) charge per generation with no seat fee. Right for developers building products or for infrequent high-quality bursts.
Self-hosted open weights (Stable Diffusion 3.5, FLUX.2 [dev], Ideogram 4.0) cost nothing per image after hardware. Right for high-volume workflows, privacy-sensitive projects, and teams that need to fine-tune on their own data. Check the license first: Stable Diffusion 3.5 is free for commercial use under a $1M revenue cap, but FLUX.2 [dev] and Ideogram 4.0's open weights are licensed for non-commercial use only unless you pay for a commercial agreement.
| Tool | Access model | Entry price | Current model | Key strength |
|---|---|---|---|---|
| Midjourney | Subscription (web) | $10/mo | V8.2 | Artistic quality, native 2K output |
| OpenAI gpt-image-2 | API, pay-per-image | $0.005/image (low quality) | gpt-image-2 | Instruction following, image editing |
| Adobe Firefly | Subscription or CC bundle | Free (25 credits/mo) | Image Model 5 | Commercial IP safety, CC integration |
| FLUX.2 (BFL) | API, per-megapixel | $0.014/MP (klein 4B) | FLUX.2 [max] | Photorealism, open weights option |
| Ideogram 4.0 | Web sub + API + open weights | Free self-host (non-commercial) | 4.0 (9.3B params) | Text rendering inside images |
| Stable Diffusion 3.5 | Open weights, self-host | Free (hardware cost only) | SD 3.5 Large (8.1B) | Zero per-image cost, full fine-tuning |
Midjourney V8.2: best for artistic quality
Midjourney operates through a web app at midjourney.com. The Discord bot still works, but the web UI is the primary interface in 2026. You log in, type a prompt, press Enter, and get four variations. No slash commands required.
The current default model is V8.2, the default since July 24, 2026 (V8.1 arrived in alpha on April 14, 2026). Practical improvements over V7: native 2048x2048 output with no separate upscale step, generation speed roughly 4-5x faster than earlier versions, and a Raw mode that removes the model's default aesthetic when you need tighter prompt adherence.
Plans and what each actually includes:
| Plan | Monthly price | Annual price (per mo) | Fast GPU hours/mo | Key add |
|---|---|---|---|---|
| Basic | $10 | $8 | 3.3 hours | Entry access |
| Standard | $30 | $24 | 15 hours | Unlimited Relax Mode |
| Pro | $60 | $48 | 30 hours | Stealth Mode |
| Mega | $120 | $96 | 60 hours | Maximum throughput |
Relax Mode is the inflection point. Basic subscribers burn through 3.3 Fast GPU hours and stop. Standard and above can queue unlimited slow-speed images after Fast hours run out. For most individual creators, Standard is the practical minimum. Pro adds Stealth Mode, which keeps your outputs out of the public gallery. Additional Fast GPU time costs $4/hour on any plan.
There is no free trial as of July 2026. Annual billing saves 20%.
Midjourney has no public developer API. Teams building products should use FLUX.2 or gpt-image-2 instead.
OpenAI gpt-image-2: best for instruction-following and image editing
OpenAI removed DALL-E 3 from the API on May 12, 2026. The current image model is gpt-image-2. GPT Image 1 is scheduled to deprecate October 23, 2026, so new API projects should target gpt-image-2 directly.
gpt-image-2 accepts text prompts, reference images, or both. Its primary advantage over generative-only tools is editing: pass it a product photo and ask it to swap the background, add a logo, or adjust lighting, and it follows complex multi-step instructions accurately. That makes it the default choice for marketing automation and e-commerce image pipelines.
Pricing per image, at common resolutions:
| Quality | Square (1024x1024) | Portrait/landscape (1024x1536) |
|---|---|---|
| Low | $0.006 | $0.005 |
| Medium | $0.053 | $0.041 |
| High | $0.211 | $0.165 |
The Batch API halves all prices. Portrait and landscape orientations cost materially less than square at medium and high quality, so orient your pipeline accordingly if cost matters.
Access is through the Images API endpoint or the Responses API for multimodal tasks. Both are available to any developer with an OpenAI API key; no ChatGPT subscription is required, though OpenAI does require API organization verification before granting gpt-image-2 access.
Adobe Firefly Image Model 5: best for commercial IP safety
Firefly's differentiator is training data provenance. Adobe trained its models on licensed Adobe Stock content and public-domain images. For organizations with formal legal review, that gives Firefly outputs clearer commercial standing than Midjourney or FLUX outputs trained on scraped web data. Firefly also attaches Content Credentials (C2PA metadata) to every generated image, supporting provenance tracking in publishing workflows.
The current model is Firefly Image Model 5, generally available since March 2026. Custom models reached public beta the same month: you upload brand images, Adobe trains a fine-tuned variant, and you generate brand-consistent outputs without running your own ML infrastructure.
Standalone Firefly plans:
| Plan | Monthly price | Generative credits/mo |
|---|---|---|
| Free | $0 | 25 |
| Standard | $9.99 | 2,000 |
| Pro | $19.99 | 4,000 |
| Pro Plus | $49.99 | 10,000 |
| Premium | $199.99 | 50,000 |
Creative Cloud Pro (the plan formerly called All Apps, now $69.99/mo billed annually) includes 4,000 Firefly credits and access to Firefly inside Photoshop, Illustrator, and Express. Adobe's newer Creative Cloud Standard tier is $54.99/mo. If you already pay for a Creative Cloud plan, that bundled allowance is your cheapest entry point, and the integration with Generative Fill in Photoshop is seamless.
The Firefly web app also surfaces models from other vendors including Google and Runway, making it a practical multi-model hub for teams already in the Adobe ecosystem.
FLUX.2: best for API-first photorealism
Black Forest Labs uses per-megapixel pricing: cost scales with output resolution. A 1024x1024 image is 1 MP. A 2048x2048 image is 4 MP, so it costs more, though not a clean 4x, because BFL prices the first megapixel higher and each additional megapixel cheaper. There is no subscription and no seat fee at bfl.ai.
The FLUX.2 model family, current as of 2026:
| Model | Price | Best for |
|---|---|---|
| FLUX.2 [klein] 4B | from $0.014/MP | Drafts, high-volume pipelines |
| FLUX.2 [klein] 9B | from $0.015/MP | Slightly higher quality than 4B |
| FLUX.2 [pro] | from $0.03/MP | Production-quality default |
| FLUX.2 [flex] | from $0.05/MP | LoRA fine-tuning at inference time |
| FLUX.2 [max] | from $0.07/MP | Highest quality, slowest |
| FLUX.2 [dev] | Free non-commercial (open weights, local) | Self-hosting, fine-tuning |
FLUX.1 [schnell], the original fast model from the prior FLUX.1 generation, is released under Apache 2.0 and free to self-host indefinitely, including commercial use. Editing operations (inpainting, outpainting, object replacement) are billed per megapixel like generation rather than as a flat per-operation fee, running from roughly $0.014/MP on the klein tiers to $0.045/MP and up on [pro].
If you are already on Google Cloud, Imagen 4 is a comparable pay-per-image alternative: roughly $0.02/image (Fast tier) to $0.04/image (Standard) via Vertex AI, with Batch API cutting prices 50%. Imagen 4 pricing is not prominently listed in Google Cloud's main pricing page; verify current rates in your Vertex AI console before building a cost model.
Ideogram 4.0: the only tool that reliably renders text inside images
Ideogram 4.0 launched June 3, 2026, as a 9.3-billion-parameter diffusion transformer released with open weights. Read the license before you build on it: the weights ship under a non-commercial agreement (personal, research, and evaluation use), and commercial self-hosting requires a separate paid license from Ideogram, with outputs barred from training competing models. Its defining capability is best-in-class text rendering inside generated images, the model's headline feature. For any image that requires readable words (poster, product label, social graphic, book cover, signage), Ideogram 4.0 is the clear first choice.
The open-weights NF4 checkpoint (roughly 4.8 GB) runs comfortably on a single 24 GB CUDA GPU via the Hugging Face Diffusers library. A larger FP8 build offers higher fidelity but needs a newer FP8-capable GPU (Ada or Hopper class).
A hosted web product is also available (check ideogram.ai for current subscription rates, which have shifted across plan structures). The 4.0 API tiers run $0.03/image (Turbo), $0.06/image (Default), and $0.10/image (Quality); the older 2a Turbo model is cheaper still at $0.025/image. Character reference features add cost on top of base generation.
JSON-first prompting is new in 4.0. You pass a structured object specifying foreground elements, background, typography, and bounding-box layout constraints rather than a prose description. Useful for programmatic generation where consistent layout across a batch matters more than prompt flexibility.
Ideogram 4.0 also ranked second overall in designer-preference evaluations across all image generators at launch, behind only proprietary frontier models.
Stable Diffusion 3.5: unlimited images for the cost of electricity
Stability AI released SD 3.5 in October 2024 under a community license that allows commercial use free of charge for individuals and organizations under $1M in annual revenue (above that, an Enterprise License applies). Three model variants:
- Large (8.1 billion parameters): best quality, requires a GPU with at least 16 GB VRAM for reliable 1-megapixel generation.
- Large Turbo: same architecture, distilled to 4-step generation. Faster at a small quality cost.
- Medium (2.5 billion parameters): runs on consumer cards with 8-12 GB VRAM.
The standard setup in 2026 is ComfyUI, which exposes every pipeline step as a node graph. Install Python 3.10 or later, clone ComfyUI from its GitHub repository, download the SD 3.5 Large checkpoint from Stability AI's Hugging Face page, and load the default SD 3.5 workflow. Most community workflows are distributed as JSON node graphs for ComfyUI.
The per-image economics: a GPU drawing 300W at roughly $0.13/kWh (U.S. average) costs about $0.04/hour to run. At 2 images per minute, that is under $0.001 per image. Breakeven against FLUX.2 [pro] at $0.03/image is roughly 30-40 images per session once warmup time is counted.
The real reason to self-host is volume and customization. You can train LoRA adapters on your own data, use custom ComfyUI nodes, and process thousands of images with no per-image cost accumulating.
How to pick
Your image needs readable embedded text: Ideogram 4.0, by a wide margin.
Your organization has a legal review process for image provenance: Adobe Firefly Pro or Pro Plus, with Content Credentials enabled.
You are building an API pipeline: FLUX.2 [pro] at $0.03/MP for photorealistic work; gpt-image-2 medium at $0.041/image (portrait orientation) for instruction-following tasks; Imagen 4 Fast at roughly $0.02/image if your infrastructure already lives in Google Cloud.
You want the best artistic output and prefer a subscription: Midjourney Standard ($30/mo) gives you Relax Mode plus V8.2 quality. Upgrade to Pro ($60/mo) only if you need Stealth Mode.
You generate high volumes and have a capable GPU: Self-hosting Stable Diffusion 3.5 Large or FLUX.2 [dev] reduces per-image cost to near zero. The hardware investment pays back inside a few months of moderate use compared to hosted API rates.
In 2026 most tools converge on photorealism for clean prompts. The real differentiators are text rendering (Ideogram), commercial license certainty (Firefly), per-image API economics (FLUX.2, gpt-image-2), and zero-marginal-cost volume (SD 3.5).
FAQ
Can I use Midjourney images commercially?
Yes, on paid plans. Midjourney's Terms of Service grant paid subscribers a broad commercial license for their outputs. Check the current terms at midjourney.com before using outputs in work with significant legal exposure, as terms have been updated periodically.
What happened to DALL-E 3?
OpenAI removed DALL-E 3 from its API on May 12, 2026. The current replacement is gpt-image-2, which handles both generation and editing and offers significantly better instruction-following on complex prompts. GPT Image 1 deprecates October 23, 2026; new API projects should target gpt-image-2 directly.
Is there a genuinely usable free AI image generator in 2026?
Adobe Firefly's free tier gives you 25 credits per month, enough for light experimentation. Stable Diffusion 3.5 Medium is free to self-host, including for commercial work under a $1M revenue cap, if you have a GPU with 8-12 GB VRAM and a few hours for setup. Ideogram 4.0's open weights are also free to run locally, but only for non-commercial use unless you buy a commercial license.
Do any of these generators allow NSFW content?
Midjourney, Firefly, and Ideogram's hosted products all prohibit NSFW content in their terms of service. Self-hosted models (SD 3.5, FLUX.2 [dev], Ideogram 4.0 open weights) run without the content filters that hosted APIs apply, subject to your jurisdiction's laws and each model's license terms.
Can I fine-tune any of these on my own brand images?
Yes, several. Stable Diffusion 3.5 and FLUX.2 [dev] support full fine-tuning locally. FLUX.2 [flex] supports LoRA at inference time via the hosted API. Adobe Firefly custom models (public beta as of March 2026) let you train on brand assets through the web UI without any ML infrastructure on your side.
The image generation market updates faster than most software categories. Stay current by reading AI Weekly free, 3 issues a week, read by 40,000+ practitioners.