When Local Image Gen Beats Every API
$319 vs $0.03–0.21/img the number that matters here
By Stan K.
A used RTX 3060 12GB costs $319 and generates for electricity (≈$0.03/hour); the APIs charge $0.03–0.21 per image — before rejections. The Apache shortlist that actually grants commercial rights (Qwen-Image, Z-Image Turbo, FLUX.2-klein-4B, FLUX.1-schnell), the workloads where local wins outright, and the honest break-even math.
A used RTX 3060 12GB costs $319 in sold medians — and generates images for electricity, about $0.03/hour on a 170W-class card. The APIs charge $0.03–0.21 per generated image, before the rejection multiplier eats in. That spread decides the whole question, and it is pure arithmetic — provided the license under your local model actually grants you commercial rights. That’s the first filter.
The Apache shortlist
Only models whose license is verifiably Apache 2.0 — no revenue caps, no territory bans, no research-only clauses — belong in a commercial local pipeline:
| Model | Params | Official requirement | On our 3060 12GB |
|---|---|---|---|
| Qwen-Image + Qwen-Image-Edit | 20B | — | runs at GGUF Q3 — our production anchor in ComfyUI |
| Z-Image Turbo | 6B | 16GB VRAM (“fits comfortably”); sub-second on an H800 at 8 NFE (9 steps) | runs via community quants — our experience, not an official figure |
| FLUX.2-klein-4B | 4B | ~13GB VRAM (3090/4070 class) | the official figure sits just above our card |
| FLUX.1-schnell | — | no official VRAM figure published | few-step by design, Apache from day one |
Every model here survived the license audit in The License Traps Hiding in “Open” Image Models — the list of ones that didn’t (a US ban, a $1M revenue cap, a research-license successor) is longer than this table.
Where local wins outright
- Mass thumbnails. Hundreds of images a month at a 1-in-4 keeper rate is 4× generations per keeper — a meter punishing pattern. Locally, a missed render costs a re-roll; on a $0.21 API it costs $0.21. The per-keeper math is the whole argument, worked through in The Real Cost of a Usable Image.
- Privacy. Client materials, unreleased products, anything under NDA — nothing leaves the box. No API term of service can leak what it never receives.
- Iterating without a meter. Prompt exploration is rejection by design. Local turns the most wasteful phase of image work into the cheapest one.
- Zero marginal cost. Electricity at ≈$0.03/hour is noise against any per-image price. The 300th image of the month costs the same as the 1st: effectively nothing.
Where the API still wins
- Top-of-arena quality. GPT Image 2.6 Sunburst tops the AA board at Elo 1262; Nano Banana 2.1 sits at 1177 for $0.03/img (AA-listed). A quantized 12GB rig does not compete with that tier — we don’t pretend it does in our FLUX vs Qwen notes.
- 4K and heavy resolution. Nano Banana Pro at 4K is $0.24/img, FLUX 3 Klein at 4K is $0.036/img — renting resolution beats buying VRAM you use twice a year.
- Embedded in a product without GPUs. If there’s no card in the loop, $0.008–0.036/img on FLUX 3 Klein is the value play among verified APIs — the full table is in the Image Gen hub.
The honest break-even
~1,519 paid imageswhat pays off a $319 used 3060 at $0.21/img
At GPT-Image-level pricing ($0.21/img), $319 ÷ $0.21 = ~1,519 generated images — at a 1-in-4 keeper rate, roughly 380 usable ones, and electricity is a rounding error. At 300 images/month that’s $63/month of avoided API spend: the card pays for itself in about five months. Against the cheapest verified API ($0.03/img) the same volume is $9/month and payback stretches to ~35 months — the cheap APIs only lose at real volume or rejection-heavy iteration.
That last sentence is the honesty clause: against $0.03/img, a 3060 is not a money-saver at low volume — it’s a buy for the 4× rejection multiplier, the privacy, and the iteration freedom. If your monthly image spend is a coffee, rent. If it’s a car payment, buy — and the full hardware math (what else that card runs, rent-vs-own at every tier) is in One GPU Beats Ten Subscriptions and the Local GPU AI leaderboard.