# Electricity per image, card by card *What one generated picture costs in electricity on each graphics card this site has measured, at the same dated national price as the tokens page, computed only from seconds and board joules already published: a 1024×1024 FLUX.2 \[klein\] 4B picture at 8 steps runs from about three cents a thousand on a laptop card to nine on a used 12 GB card, and the whole shelf from under a cent a thousand to a dollar and a half. Board power, not the wall, one picture at a time, and the arithmetic shown once so a reader can redo it at their own rate.* *Published 2026-11-03 (UTC) · A small (human) team and a fleet of AI agents.* **the short version:** A thousand FLUX.2 [klein] 4B pictures at 1024×1024 and 8 steps, one at a time, cost 3.16 cents of board electricity on an NVIDIA GeForce RTX 5090 Laptop GPU (24 GB) under Dynamic Boost, 4.77 cents on a ZOTAC GeForce RTX 5080 Solid CORE OC (16 GB) at its stock 360 W, and 8.33 to 8.90 cents on three Ampere cards at 350 W: an EVGA GeForce RTX 3090 Ti FTW3 Ultra (24 GB), an EVGA GeForce RTX 3090 XC3 Ultra (24 GB) and an NVIDIA GeForce RTX 3080 Ti Founders Edition (12 GB). Drawn back to back for ten minutes, 3.39 cents on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB) held to 300 W. Four times the steps costs 3.5 to 3.9 times the electricity, four times the pixels 3.0 to 3.3 times, a heavier model 1.5 to 2.5 times. The cheapest published picture is 0.79 cents a thousand, a 512×512 klein picture at 8 steps on the laptop card held to 95 W, its share of a batch of three; the dearest is 150 cents a thousand, Z-Image-Turbo at 2048×1280 and 32 steps on the 12 GB card with the model partly off the card. The price is 18.31 cents a kilowatt-hour, the U.S. residential average for July 2026; nothing was measured for this page. Board power, not the wall, except one print-lab batch of three 512×512 plates at 32 steps, read both ways: 1.27 watt-hours at the card, 1.87 at the meter. 2,567 words · about 12 minutes (at 220 words/min) · 2 tables · data kit: no https://research.strata2signal.com/electricity-per-image/ --- ## What is being counted, and at what price {#what-is-being-counted-and-at-what-price} An image model draws a picture in sampling steps. The benches behind this page summed the power the card's own driver reports, once or twice a second, over each picture, and printed joules or watt-hours; a joule is a watt for a second, and 3,600,000 joules are a kilowatt-hour, the unit on an electricity bill. This page turns those into money, the twin of [*Electricity per million tokens, card by card*](/electricity-per-million-tokens/), which did the same for writing. The price is **18.31 cents a kilowatt-hour**, the average price of electricity to an American home in July 2026 by the U.S. Energy Information Administration (*Electric Power Monthly*, table 5.6.A, the edition released 2026-09-24), the rate the tokens page and [*What it costs to run a high-end AI rig at home*](/what-it-costs-to-run/) use; the agency's August reading, 18.16 cents, is 0.8 per cent lower. The figures are **cents per thousand pictures**, because one picture costs thousandths of a cent. ## The arithmetic, once {#the-arithmetic-once} Take a ZOTAC GeForce RTX 5080 Solid CORE OC 16 GB at its own 360 W limit drawing one FLUX.2 [klein] 4B picture at 1024×1024 and 8 steps: its page's kit prints 3.593 seconds and 937 joules. 937 ÷ 3,600,000 is 0.000260 kilowatt-hours, which at 18.31 cents is 0.004766 cents for the picture, or **4.77 cents a thousand**. Every cell on this page is that: *joules per picture ÷ 3,600,000 × 18.31 × 1,000*. A watt-hour, where a source prints one, is 3,600 joules. To use your own rate, replace the 18.31. ## The table {#the-table} Every row is from [*The same eighteen pictures, card by card*](/the-same-eighteen-pictures-card-by-card/) and its kit's tables: one pinned ComfyUI stack, FLUX.2 [klein] 4B in bf16 at 1024×1024 and 8 steps. The first three figures are the grid cell, the median of three timed pictures drawn one at a time; the last three, the page's ten-minute run, back to back. A desktop card's first row is its stock limit, but the EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB's is its everyday 350 W, measured beside a serving card; the NVIDIA GeForce RTX 5090 Laptop GPU 24 GB ran under Dynamic Boost, 138 to 150 W. Across rows the grid's label holds, same software, different computer, but for the NVIDIA GeForce RTX 3080 Ti Founders Edition 12 GB and the EVGA GeForce RTX 3090 XC3 Ultra 24 GB, measured in one computer a day apart. | Card and setting | One picture, s | One picture, J | Cents per thousand | Ten minutes, s per picture | Ten minutes, J per picture | Cents per thousand | |---|---:|---:|---:|---:|---:|---:| | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96 GB, 600 W | 1.506 | Too short to read | — | Stopped at 83 °C | — | — | | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96 GB, 300 W | No cell at this size | — | — | 2.295 | 666.7 | 3.39 | | ZOTAC GeForce RTX 5080 Solid CORE OC 16 GB, 360 W | 3.593 | 937 | 4.77 | 3.608 | 991.6 | 5.04 | | ZOTAC GeForce RTX 5080 Solid CORE OC 16 GB, 300 W | No cell at this size | — | — | 3.611 | 977.6 | 4.97 | | NVIDIA GeForce RTX 5090 Laptop GPU 24 GB, Dynamic Boost | 4.501 | 622 | 3.16 | 4.482 | 663.0 | 3.37 | | NVIDIA GeForce RTX 3080 Ti Founders Edition 12 GB, 350 W | 5.725 | 1,750 | 8.90 | 5.814 | 1,908.4 | 9.71 | | EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB, 350 W | 5.021 | 1,638 | 8.33 | 4.972 | 1,666.1 | 8.47 | | EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB, 300 W | No cell at this size | — | — | 5.194 | 1,518.5 | 7.72 | | EVGA GeForce RTX 3090 XC3 Ultra 24 GB, 350 W | 5.180 | 1,719 | 8.74 | 5.292 | 1,795.8 | 9.13 | | EVGA GeForce RTX 3090 XC3 Ultra 24 GB, 300 W | No cell at this size | — | — | 5.527 | 1,623.7 | 8.26 | **The blanks.** The grid's kit prints a dash under four power samples, and at two a second the workstation card's 1.506-second picture has three; its stock ten-minute run met the bench's 83 °C stop at 31 seconds. The 300 W pass re-ran only the 2048×1280 cells, hence no 300 W cell at this size. **What the table says.** The laptop card has the cheapest one-picture cell, 3.16 cents a thousand, a third under a 5080's 4.77, on about half a 5080's power (145.4 W mean over ten minutes against 270.1). The three Ampere cards sit within seven per cent of each other, at 1.7 to 1.9 times a 5080's cost. Over ten minutes the workstation card at 300 W and the laptop at 150 W spent the same per picture, 666.7 against 663.0 joules, within sensor noise as the source says. Held to 300 W for ten minutes, a 3090 Ti and a 3090 saved 8.9 and 9.6 per cent of the energy for about four and a half per cent more time a picture; a 5080, 60 W under its limit, saved 1.4 per cent. ## Steps, pixels and the model {#steps-pixels-and-the-model} The same grid ran every card at 8 and 32 steps, at 512×512, 1024×1024 and 2048×1280, on FLUX.2 [klein] 4B, Z-Image-Turbo and FLUX.1 [dev] in fp8; this table is its 32-step cells at each card's higher setting. A paged cell is one where the NVIDIA GeForce RTX 3080 Ti Founders Edition 12 GB could not hold the whole transformer and streamed part of it from system memory each step; the source sets no paged cell in a ratio with a whole one, and neither does this page. | Card and setting, 32 steps | FLUX.2 [klein] 4B, 512×512, cents per thousand | FLUX.2 [klein] 4B, 1024×1024, cents per thousand | FLUX.2 [klein] 4B, 2048×1280, cents per thousand | Z-Image-Turbo, 1024×1024, cents per thousand | FLUX.1 [dev], 1024×1024, cents per thousand | |---|---:|---:|---:|---:|---:| | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96 GB, 600 W | Too short to read | 12.68 | 36.36 | 18.53 | 28.58 | | ZOTAC GeForce RTX 5080 Solid CORE OC 16 GB, 360 W | 5.47 | 18.02 | 52.07 | 28.60 | 44.97 | | NVIDIA GeForce RTX 5090 Laptop GPU 24 GB, Dynamic Boost | 3.78 | 12.24 | 34.92 | 18.32 | 28.21 | | NVIDIA GeForce RTX 3080 Ti Founders Edition 12 GB, 350 W | 10.77 | 32.66 | 92.52 | 53.02, paged | 69.99, paged | | EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB, 350 W | 9.69 | 29.26 | 80.30 | 44.75 | 60.91 | | EVGA GeForce RTX 3090 XC3 Ultra 24 GB, 350 W | 10.31 | 31.91 | 89.89 | 50.82 | 66.29 | *Board electricity, from the joules-per-picture tables of the same page's kit; a ratio down a column crosses computers and carries the grid's label.* **Steps.** From 8 to 32 steps at 1024×1024 on klein, four times the steps costs 3.5 to 3.9 times the electricity on the five cards with both cells; the shortfall is the once-per-picture work, prompt and decode, 0.55 to 1.30 seconds a picture on the four cards the source reads it for. **Pixels.** At 32 steps on klein, four times the pixels (512×512 to 1024×1024) costs 3.0 to 3.3 times as much, and 2.5 times the pixels (1024×1024 to 2048×1280) 2.7 to 2.9 times on all six cards: a 2048×1280 step costs about three 1024×1024 steps, the source's finding 6. **The model.** At 1024×1024 and 32 steps Z-Image-Turbo costs 1.5 to 1.6 times klein on the whole cells, FLUX.1 [dev] 2.1 to 2.5 times. ## The laptop, and why its figures are so low {#the-laptop-and-why-its-figures-are-so-low} The five cheapest published pictures on the shelf are all the NVIDIA GeForce RTX 5090 Laptop GPU 24 GB's, at 512×512 and 8 steps, from the render leg of [*A laptop, asked the desktop's questions*](/a-laptop-asked-the-desktops-questions/). Held to a fixed 95 W on the [print lab](https://printlab.strata2signal.com)'s 8-step recipe it drew a picture every 1.640 seconds for 0.0431 watt-hours, which is 155.2 joules and **0.79 cents a thousand**; under Dynamic Boost at a flat 150 W, 1.246 seconds for 0.0500 watt-hours, 0.92 cents, faster and dearer. The same kit's desktop seats on the same recipe, an older ComfyUI build in another computer, read 0.1369 watt-hours on an EVGA GeForce RTX 3090 XC3 Ultra 24 GB at 350 W (2.51 cents a thousand), 0.1309 on an EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB at 350 W (2.40) and 0.1398 on a Gigabyte GeForce RTX 3080 10 GB at 250 W (2.56); the laptop's kit names those seats by class, the eighteen-pictures kit's Who ran where by board. Two things make the figure. **The power cap**: at 95 W the board's own ceiling held its clocks down for 93 to 100 per cent of every timed window, and in ten minutes it made 294 pictures to a 3090's 311 at 350 W on 29 per cent of the board power, as its page reads it. **A batch of three**: a recipe row is one picture's share of three plates drawn in one submit, as the print lab draws; one at a time for ten minutes at 95 W, the same card read 190.5 joules, 0.97 cents a thousand, 23 per cent more. ## The dearest picture on the shelf {#the-dearest-picture-on-the-shelf} The dearest published picture is Z-Image-Turbo at 2048×1280 and 32 steps on the NVIDIA GeForce RTX 3080 Ti Founders Edition 12 GB at its stock 350 W: 85.428 seconds and 29,500 joules, **0.15 cents for the one picture, 150.04 cents a thousand**. The cell is paged: at that size as little as 68.7 per cent of the transformer stayed on the card through the sampler. The same cell on the EVGA GeForce RTX 3090 XC3 Ultra 24 GB, in the same computer and slot a day later with the model whole, read 82.082 seconds and 28,464 joules, 144.77 cents a thousand; the source refuses that ratio, and so does this page. Held to 300 W the 12 GB card drew the same picture in 88.218 seconds for 26,199 joules, 133.25 cents: eleven per cent less energy for three per cent more time, the cap and a cooler card together, the source says. The FLUX.1 [dev] cell at that size has no figure on four of the six cards, where the ninety-second rule stopped the warm-up; of the two that finished, the ZOTAC GeForce RTX 5080 Solid CORE OC 16 GB drew it on 24,934 joules, 126.82 cents a thousand. ## The one wall reading {#the-one-wall-reading} One picture on this shelf was read at the wall as well as the card, on [*What it costs to run a high-end AI rig at home*](/what-it-costs-to-run/): a print-lab picture, three plates of FLUX.2 [klein] 4B at 512×512 and 32 steps drawn as one batch, on an RTX 3090 (24 GB) held at 250 W, which [*The cost to purchase and run a very capable home AI rig*](/home-inference-and-diffusion-rig-cost/) names as one of the two EVGA GeForce RTX 3090 XC3 Ultra 24 GB cards of that box. A bench of twelve on 2026-09-12, 03:34Z to 03:38Z, read 18.3 seconds a batch and **1.27 watt-hours at the card**, and 367 W at the box's meter against 102 W idle that minute, **1.87 watt-hours at the wall**: 23.25 cents a thousand batches at the card and 34.24 at the meter, the "about 34 cents" that page prints. The wall reads 1.47 times the board on that picture, 0.60 watt-hours more. The box idled at 102 W that minute, but that idle holds the card's own idle draw, which the 1.27 already counts, so these readings cannot split the 0.60 between the box and the supply's losses, and this page does not. It is the shelf's only board-to-wall reading, one picture on one box; the wall reads more on every row above. ## What idling would cost, and why it is not here {#what-idling-would-cost-and-why-it-is-not-here} No source prints what a card draws holding an image model and doing nothing, only each card empty before its passes (the grid kit's dated windows), so unlike the tokens page this page prints no idle bill. That page's line, *watts × 8,760 ÷ 1,000 × 18.31* cents a year, takes any reading a reader has. ## What this page does not say {#what-this-page-does-not-say} - **Board power, not the wall**, on every row but the one print-lab meter row; the card's own counter, fans included. - **Across computers the figure carries the grid's label**, and the laptop's recipe rows and the print-lab batch are other instruments on other builds: they sit beside the grid's cells here and are never divided by them. - **No idle bill, no 768-pixel cell, no 4- or 20-step cell.** None is published with energy for any card. - **No cap-ladder row.** [*What 150 Watts Buys*](/what-150-watts-buys/) printed each image lane's busy draw on the NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96 GB at 600 W and its slowdown at 500 and 450 W, not joules per picture. - **The debug log ran** on every grid instance, 30 to 60 lines a step; a normal ComfyUI should match or beat these figures by an amount the source did not measure. - **The price is a national average**, not our bill or yours. ## How to check our work {#how-to-check-our-work} Every figure is on a released page linked beside it; this page measured nothing and prints no number its sources do not print, except the cents, the formula above applied to each joule or watt-hour figure; the ratios and percentages in the text, one figure set against another; and the wall reading's 0.60 watt-hours, 1.87 − 1.27. Redo any cell with a calculator: joules per picture, divided by 3,600,000, times your rate in cents, times a thousand. ## Who ran this, and thanks {#who-ran-this-and-thanks} The figures are five pages' work and not this one's, the pages in Sources below; the U.S. Energy Information Administration publishes the price monthly where anyone can read it. **NVIDIA** made the silicon, the Founders Edition board, the laptop part, the workstation card and the driver every board watt here was read from; **ZOTAC**, **EVGA** and **Gigabyte** the other boards, and **MSI** the laptop that carries its board. The pictures stand on **ComfyUI** and the **comfy-kitchen** and **comfy-aimdo** libraries beneath it, on **PyTorch** and **NVIDIA's CUDA**, on **Astral's uv** and on **Network UPS Tools**. The models are Black Forest Labs' **FLUX.2 [klein] 4B** and **FLUX.1 [dev]** and Alibaba's Tongyi Lab's **Z-Image-Turbo**, with Alibaba Cloud's **Qwen 3 4B**, Google's **T5-XXL** and OpenAI's **CLIP ViT-L/14** as their text encoders; **Hugging Face** hosted every pinned file. Each source page names the licence it found for its models; this page ran nothing and adds none. A small human team owns the cards, set the rules before the runs and signed the figures on each source page; a fleet of AI agents ran the harnesses and did the division here under that team's rulings. Thanks to the readers who asked, on every forum this workshop reads, what a picture actually costs on the card they have: this page is that division. ## Sources {#sources} - [*The same eighteen pictures, card by card*](/the-same-eighteen-pictures-card-by-card/), 2026-09-26: findings 1 and 3 to 6, How it was measured, and the kit's `data/TABLES.md`: every grid cell and ten-minute run. - [*A laptop, asked the desktop's questions*](/a-laptop-asked-the-desktops-questions/), 2026-09-22: Five things the 95 W pass found, and the render kit's `RESULTS.md`, R1 and R4: the recipe rows and the desktop seats. - [*What it costs to run a high-end AI rig at home*](/what-it-costs-to-run/), 2026-10-06: the price, and the Numbers row "One picture, three plates, warm". - [*The cost to purchase and run a very capable home AI rig*](/home-inference-and-diffusion-rig-cost/), 2026-09-14: What kind of performance to expect, the card that drew it. - [*Electricity per million tokens, card by card*](/electricity-per-million-tokens/), 2026-10-30: the price's sourcing and the idle arithmetic.