AI Image & Video Gen · Water Impact Explorer · Report v1.2

Generative AI: The Visual Cost

From a single image to platform-scale video production — the hydrological reality of AI creative tools, grounded in 2025–2026 benchmark estimates.

~2.4 mL
Fast AI image / image
~95 mL
Ultra quality / image
~382 mL
10s HD video clip
~7.2 L
1 min of 4K AI video

Li et al. arXiv:2311.16863 · Zhao et al. arXiv:2509.19222 · MIT Tech Review 2025 · LBNL 2024

Image Quality / Model

Energy per image varies 40× between SD-Turbo draft mode and DALL-E 3 ultra with iterative editing

SDXL / Midjourney v6 standard — 1024px, ~30 diffusion steps. Most common API tier.
50
1101005001k
0.25 L/kWh

Water per Image

---
per image generated
---
Daily total
---
per day
Annual total
---
per year
Daily energy
---
kWh / day
Images / burger
---
images = 1 burger WF

Scale Equivalences — Daily

🍔 1 burger = years of this daily habit---
🚿 vs. 10-min shower (65 L)---
🏠 vs. household daily (341 L)---
💧 500 mL bottles equivalent---

Scale Visualization

Per-image and per-clip values vs. real-world water baselines — logarithmic scale — updates live with calculator

Key Findings

Image & video generation water context — companion to bra-khet AI Water-Energy Nexus Report v1.2

Text generation (LLM inference) at 2026 efficiency: 0.26–2.0 mL per query. Image generation adds a fundamentally different compute burden:
  • • Fast AI image (SD-Turbo, 0.5 Wh): ~2.4 mL — comparable to a 2026 LLM query at standard efficiency
  • • Standard image (SDXL / MJ v6, 3 Wh): ~14.3 mL — ~55× more water than Gemini per-query
  • • High quality (50 steps + upscale, 10 Wh): ~47.7 mL — nearly a shot glass of water per image
  • • Ultra (DALL-E 3 with editing, 20 Wh): ~95.4 mL — about 365× more water than Gemini per-query
The core reason: diffusion models run 20–100 denoising steps, each requiring a full U-Net or DiT forward pass. Text models run one forward pass with KV-cache reuse — an order-of-magnitude computational difference for similar output quality.
Video generation compounds image-gen cost by frame count and temporal consistency overhead:
  • • 5s 480p SD clip (~125 frames, 25 Wh): ~119 mL — roughly half a small glass of water
  • • 10s 720p HD clip (~240 frames, 80 Wh): ~382 mL — one full 500 mL bottle per clip
  • • 30s 1080p FHD clip (~720 frames, 300 Wh): ~1.43 L — three water bottles; equivalent to ~550 Gemini queries
  • • 60s 4K clip (~1,440 frames, 1500 Wh): ~7.16 L — more than a full gallon; matches ~27,500 Gemini queries
Temporal consistency models (Sora, Wan, CogVideo) require cross-frame attention over time, adding 20–40% overhead vs. independent frame generation. Efficient video diffusion (StreamDiffusion, AnimateDiff V3) can reduce this by ~3–5×.
One hamburger carries ~2,498 L total water footprint (Mekonnen & Hoekstra 2012, Ecosystems):
  • • Fast images (2.4 mL each): 1,040,833 images per burger — generating one million quick AI images = one hamburger's water
  • • Standard images (14.3 mL each): 174,685 images per burger
  • • High images (47.7 mL each): 52,370 images per burger
  • • Ultra images (95.4 mL each): 26,185 images per burger
  • • 10s HD clips (382 mL each): 6,544 clips per burger — ~18 years of daily creator-tier video generation
  • • 60s 4K clips (7,155 mL each): 349 clips per burger — less than one year of daily 4K generation at Pro Studio rate
All tool figures apply to cloud API inference in professional data centers (WUE 0.15–0.55 L/kWh + grid GWIF ~4.52 L/kWh). Running Stable Diffusion locally changes the calculus:
  • • Your GPU has no data center cooling loop — zero Scope 1 direct water evaporation
  • • Scope 2 still applies: your grid's upstream thermoelectric generation (GWIF ~4.52 L/kWh national avg)
  • • RTX 4090 at 400W peak — 512×512 image in ~0.5s → ~0.056 Wh → ~0.25 mL per image (Scope 2 only)
  • • Consumer GPU generates images at ~10× lower water cost than cloud API (no cooling overhead, same compute time)
  • • This also applies to video: local inference on a 4090 for a 10s clip might use ~8–15 Wh vs. 80 Wh on a data center GPU cluster
  • • Diffusion distillation (LCM, TurboSD, Hyper-SD): 4-step inference vs. 30 steps → ~7.5× energy reduction per image at comparable quality
  • • Flow matching (SD3, FLUX): Deterministic trajectories with fewer NFE (Number of Function Evaluations) — 8–12 steps at high quality; ~2–3× energy reduction vs. DDPM
  • • Video frame reuse (StreamDiffusion): Delta diffusion for slow-motion scenes reduces effective NFE by 40–60% at similar temporal coherence
  • • Linear-attention architectures (Mamba, SSMs for video): Eliminate quadratic attention scaling → 3–5× lower compute for long-duration video generation
  • • Flash-Attention-2 + Triton kernels: 30–40% power efficiency gain at constant throughput via reduced GPU memory bandwidth pressure
  • • Trajectory: 2026 "fast" quality (0.5 Wh/image) will be standard quality by 2028 via distillation; current ultra (20 Wh) may drop to ~5 Wh through efficient sampling pipelines

Sources

  1. [1] Li, P. et al. (2023). Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv:2311.16863.
  2. [2] Zhao, S. et al. (2025). Energy and Water Consumption in AI-Generated Content: Image and Video Models. arXiv:2509.19222.
  3. [3] Hao, K. (2025, May). How much energy does AI actually use? MIT Technology Review.
  4. [4] Ren, S. et al. (2023). On the Energy and Water Consumption of Generative AI. arXiv:2304.03271.
  5. [5] Lawrence Berkeley National Laboratory (2024). US Data Center Energy Use Report (GWIF 4.52 L/kWh).
  6. [6] Mekonnen, M.M. & Hoekstra, A.Y. (2012). A global assessment of the water footprint of farm animal products. Ecosystems 15(3):401–415.
  7. [7] Rombach, R. et al. (2022). High-Resolution Image Synthesis with Latent Diffusion Models. CVPR 2022 (SD/SDXL baseline architecture).
  8. [8] OpenAI (2024). DALL-E 3 technical report. openai.com/research.
  9. [9] Luo, S. et al. (2023). LCM: Latent Consistency Models. arXiv:2310.04378 (4-step distillation inference).
  10. [10] Esser, P. et al. (2024). Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (SD3 / FLUX). arXiv:2403.03206.
  11. [11] Blattmann, A. et al. (2023). Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets. arXiv:2311.15127.
  12. [12] Bar-Tal, O. et al. (2024). Lumiere: A Space-Time Diffusion Model for Video Generation. arXiv:2401.12945.
  13. [13] AWWA (2022). US daily household indoor water use baseline: 341 L/day.
  14. [14] Jegham, I. et al. (2025). Empirical energy benchmarking of 30 LLMs (Gemini 0.24 Wh/query). arXiv preprint.
  15. [15] Google (2026). Ironwood TPU benchmarks and environmental disclosures. Google Environmental Reports.