TL;DR: A used GPU pays for itself against Midjourney Standard ($30/month) in under a year, and local generation costs about a hundredth of a cent per image in electricity. Against pay-per-image APIs it’s a different story — FLUX schnell costs $0.003 per image on fal.ai or Replicate, so hardware never pays back at hobby volume. Self-host for volume, iteration speed, and privacy; keep the API if you make a few hundred images a month.

	Midjourney	FLUX / gpt-image-1 APIs	Self-hosted (ComfyUI)
Best for	Polished results with zero setup	Metered use inside apps and scripts	Unlimited volume, LoRAs, private inputs
Price / Cost	$10–$30/mo subscription	$0.003–$0.19 per image	~$0.0001/image power; $260–$805 GPU
The catch	Caps measured in GPU minutes	Costs scale linearly, forever	VRAM floor is real; licenses vary by model

Honest take: if you already own an 8GB+ GPU and generate images weekly, a Midjourney subscription is the one bill you can cancel today. If you generate 200 images a month through FLUX schnell at sixty cents, buying a $700 card to save it is hobby math, not savings.

This is the image-generation entry in our cost series, using the same method as the self-hosted Whisper vs cloud STT math and the TTS cost breakdown: real prices with verification dates, a hardware floor, and an explicit “don’t do this” section. All prices checked October 2026.

What does cloud image generation cost in 2026?

Cloud image generation spans a 60x price range: $0.003 per image for FLUX schnell on the metered platforms up to ~$0.19 for a high-quality gpt-image-1 render, with Midjourney selling subscriptions instead of images.

The metered APIs, as of October 2026:

Service	Model	Price per 1024×1024 image
fal.ai	FLUX.1 schnell	~$0.003 (billed $0.003/megapixel)
Replicate	FLUX.1 schnell	$3 per 1,000 images (~$0.003)
fal.ai	FLUX dev	~$0.025/megapixel
Replicate	FLUX dev	$0.025–$0.03 per image
OpenAI	gpt-image-1	~$0.02 low / ~$0.07 medium / ~$0.19 high quality

Two footnotes on that table. fal.ai bills by megapixel, so a 512×512 test render costs a quarter of the listed price. And OpenAI’s gpt-image-1 is token-billed ($40 per 1M image output tokens), so the per-image figures are OpenAI’s own quality-tier estimates, not a flat rate — budget against the high end if you care about output quality.

Midjourney doesn’t sell per-image at all. As of October 2026 the Basic plan is $10/month for ~3.3 fast GPU hours — roughly 200 images, since a standard job burns about a fast minute — and Standard is $30/month for 15 fast hours plus unlimited slower “relax” generations. Pro ($60) and Mega ($120) add stealth mode and more fast time. The cap is measured in GPU minutes, not images, so heavy upscaling and variations drain it faster than the headline numbers suggest.

The spread is the story: a 1,000-image month costs $3 on schnell, $25–$30 on FLUX dev, $30 on Midjourney Standard (if it fits in relax mode), and $70–$190 on gpt-image-1. Pick the wrong tier for your usage pattern and you overpay by an order of magnitude before self-hosting even enters the conversation.

What does self-hosted image generation cost to run?

On hardware you already own, a locally generated image costs about a hundredth of a cent in electricity — effectively zero, with no cap and no meter.

The math for a used RTX 3060 at its full 170W board power, generating one SDXL-class image in 15 seconds at the US average residential rate ($0.17/kWh):

$ python3 -c "watts=170; sec=15; rate=0.17; print(f'\${watts/1000*sec/3600*rate:.5f} per image')"
$0.00012 per image

Even a worst case — FLUX.2 dev at ~42 seconds per image on a 450W RTX 4090 — lands under a tenth of a cent. Generate 2,000 images a month and power is still under a quarter; our always-on server electricity breakdown has per-region rates if yours differ. The real cost is the card:

• ~$260–$296 — used RTX 3060 12GB (street price, September 2026). Runs SDXL comfortably — our 8GB VRAM Stable Diffusion guide covers the config — plus Z-Image-Turbo via FP8 quants and FLUX.2 klein 4B.
• ~$679–$805 — RTX 5060 Ti 16GB (street, September 2026; MSRP $429 — nobody pays MSRP in this market). The 16GB tier runs Z-Image-Turbo at BF16 and FLUX.2 dev at aggressive GGUF quants, generating a 1024px image in a handful of seconds.
• ~$1,456–$1,558 — used RTX 3090 24GB (October 2026). The 24GB tier where FLUX.2 dev at GGUF Q4 fits resident. Only worth it if you specifically want dev-class detail locally.

Speed is no longer the local weakness it was in 2024. Community figures put Z-Image-Turbo (Apache 2.0, 6B, 8 steps) at roughly 2.3 seconds per 1024×1024 image on an RTX 4090 and 4–5 seconds on an RTX 4060 Ti — iteration-loop speed, not batch-job speed.

The honest line items that aren’t electricity: an evening to set up ComfyUI and download 10–20GB of weights, and ongoing disk space as model files accumulate. If you’ve never run ComfyUI, start with the ComfyUI vs Automatic1111 vs Forge comparison before buying anything.

Where is the break-even point?

A used RTX 3060 pays for itself against Midjourney Standard in about 9–10 months — and never pays for itself against FLUX schnell API pricing at any sane volume.

Using the $278 street-price midpoint:

You’re replacing	Monthly cost	GPU pays back in
Midjourney Mega ($120/mo)	$120	~2.5 months
Midjourney Standard ($30/mo)	$30	~9–10 months
Midjourney Basic ($10/mo)	$10	~28 months
gpt-image-1, 500 high-quality images/mo	~$95	~3 months
FLUX dev API, 1,000 images/mo	~$25–$30	~10–11 months
FLUX schnell API, 1,000 images/mo	~$3	~7.5 years — don’t

Three readings:

1. Subscriptions are what self-hosting kills. Midjourney Standard and above, or any real gpt-image-1 volume, pays for a card inside a year. The subscription also caps you; the card doesn’t.
2. schnell API pricing is nearly unbeatable on cost. At $0.003 per image you’d need ~93,000 images to recoup $278. Against that baseline, the reasons to go local are LoRA training on your own subjects, img2img on photos that shouldn’t leave your machine, no content filter on legal-but-flagged work, and latency-free iteration — not dollars.
3. The 16GB tier changes the question, not the math. A $742-midpoint RTX 5060 Ti against Midjourney Standard is ~25 months — long. The case for it is quality parity: Z-Image-Turbo and FLUX.2 dev quants get you prompt-adherence and detail that the $0.003 schnell tier doesn’t match.

For a few big batch jobs a year with no GPU, rent instead of buying either: an RTX 3090 on Vast.ai starts around $0.07/hr (market price, floats), so a 5,000-image overnight run costs under a dollar in compute.

When should you NOT self-host image generation?

Skip self-hosting when any of these hold:

• Your output is commercial and you haven’t read the model license. FLUX.2 dev and klein 9B ship under Black Forest Labs’ non-commercial license — self-hosting them for client work is a violation no matter whose GPU it runs on. The commercially safe local picks are Z-Image-Turbo and FLUX.2 klein 4B (both Apache 2.0) and SDXL; our Flux 2 guide walks the license table. Midjourney and OpenAI bundle commercial terms into the price — for some client situations that clarity is worth $30/month.
• You generate under ~300 images a month and schnell quality suffices. That’s under a dollar on fal.ai or Replicate. Hardware spend for that is a hobby purchase; call it one.
• You need best-in-class prompt following or in-image text. gpt-image-1 still leads open-weight models on complex instruction adherence and typography. If your workload is “render this exact diagram with these exact labels,” the $0.19 render that works beats five free ones that don’t.
• Your best GPU has less than 8GB VRAM. That’s the practical floor for current SDXL-class work, and quantized workarounds below it trade away the speed that makes local generation pleasant. Check the GPU buying guide on runaihome.com before forcing it.

What to actually buy

Prices as of October 2026, all taken from the comparison above:

Your situation	The pick	Price	Where
Own an 8GB+ GPU already	ComfyUI + Z-Image-Turbo or SDXL	$0	Setup guide
Replacing a $30+/mo subscription on a budget	Used RTX 3060 12GB	$260–$296	Check price
Want near-dev quality locally, commercially safe	RTX 5060 Ti 16GB + Z-Image-Turbo	$679–$805	Check price
Occasional big batch runs, no GPU	Rented RTX 3090	from $0.07/hr	Vast.ai
Low volume, no privacy constraint	FLUX schnell via fal.ai or Replicate	~$0.003/image	fal.ai/pricing

FAQ

Is self-hosted image quality as good as Midjourney in 2026?

Close enough that volume users should test before paying another month. FLUX.2 dev and Z-Image-Turbo hold their own on photorealism; Midjourney still wins on zero-effort aesthetic polish, and gpt-image-1 on complex prompt adherence and text rendering. For LoRA-customized subjects — your product, your character — local wins outright, because the cloud services can’t do it.

How much VRAM do I need to generate images locally?

8GB is the working floor for SDXL-class models with the right settings, 12GB is comfortable (and adds FLUX.2 klein 4B plus FP8 Z-Image-Turbo), 16GB runs Z-Image-Turbo at full BF16, and 24GB fits FLUX.2 dev at GGUF Q4. Below 8GB, rent time on Vast.ai instead of fighting your card.

Does self-hosting image generation actually save money?

Against subscriptions, yes — a $278 used RTX 3060 beats Midjourney Standard in under a year and electricity rounds to a hundredth of a cent per image. Against FLUX schnell’s $0.003/image API pricing, no: you’d need roughly 93,000 images to break even. The dividing line is whether you’re replacing a subscription or a metered API.

Sources

• fal.ai pricing — FLUX dev $0.025/MP, schnell $0.003/MP (checked October 2026)
• Replicate pricing — FLUX dev $0.025–$0.03/image, schnell $3 per 1,000 images (checked October 2026)
• OpenAI API pricing — gpt-image-1 token rates and per-quality image estimates (checked October 2026)
• Midjourney pricing overview, eesel.ai — plan prices and fast-hour allocations (third-party tracker, updated 2026)
• Z-Image-Turbo community benchmarks — per-card generation times (2026)

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe