Krea 2 Turbo · 5 GPUs measured first-party · text-to-image · Updated October 2026
What GPU Do You Need for Krea 2 Turbo?
Krea 2 Turbo is the aesthetics-first image model, Krea's distilled generator tuned for photorealistic, opinionated output rather than raw speed. It's also the heaviest text-to-image workload we've ever measured: ~46GB peak VRAM at bf16, with 9.5 images per minute on a B200 the best result across our 5-GPU run. These are, to our knowledge, the only published benchmarks of it.
Benchmarked weights: krea/Krea-2-Turbo
NVIDIA B200
9.5 images/min on Krea 2 Turbo, the ceiling. Measured on our bench. 192GB of VRAM, $40,000 at launch.
- 9.5 images/min on Krea 2 Turbo
- 192GB, clears the Krea 2 Turbo floor
- Rentable by the hour rather than bought
- 1000W board rating
- Datacenter or workstation hardware, not a retail purchase
NVIDIA RTX PRO 6000 Blackwell Workstation Edition
4.71 images/min on Krea 2 Turbo, lowest launch price that still fits. Measured on our bench. 96GB of VRAM, $8,565 at launch.
- 4.71 images/min on Krea 2 Turbo
- 96GB, clears the Krea 2 Turbo floor
- Rentable by the hour rather than bought
- 600W board rating
- Datacenter or workstation hardware, not a retail purchase
What GPU Do You Need for Krea 2 Turbo?, images/min by GPU
Measured on our own bench. A card absent from this chart has not been run on this model yet, or cannot fit it.
Efficiency: images/min per 100W drawn
Power is the average pulled during the run, sampled at 1Hz. The fastest card is often not the one here, and for anything left running this is the number that shows up on the bill.
Value: images/min per $1,000 of MSRP
Launch price, not street price, so it ages. A speed leaderboard always crowns the most expensive card; this is the counterweight.
Krea 2 Turbo. Measured image generation speed by GPU
| GPU | Images/min | s per image | img/W·min | Avg power |
|---|---|---|---|---|
| NVIDIA B200 | 9.5 | 6.32 | 0.012 | 800.9 W |
| NVIDIA H200 | 7.25 | 8.27 | 0.011 | 658.6 W |
| NVIDIA H100 80GB HBM3 | 6.85 | 8.76 | 0.011 | 638.1 W |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 4.71 | 12.74 | 0.01 | 491.1 W |
| NVIDIA A100 80GB SXM4 | 3.35 | 17.9 | 0.01 | 345.9 W |
What you're paying the VRAM for. Krea's pitch is look, not throughput: a model tuned hard toward the polished, photographic aesthetic its creative-tool users want, without the prompt-wrestling. Our numbers frame the cost of that: 6.3 seconds per image on a B200, 8.3 on an H200, and a ~46GB resident pipeline that rules out every consumer and workstation card short of 48GB. This is a rented-silicon model, full stop, and at roughly 570 images per B200-hour, a curated shoot's worth of output costs a few dollars.
Reading our chart. The architecture gap is stark again: the A100 80GB, plenty of memory, crawls at 3.35 images/min, a third of the B200's pace, and even the RTX PRO 6000 (4.71/min) trails the Hopper cards. Like TRELLIS.2, this workload scales with compute generation, not just VRAM. If you're choosing where to run it: B200 for throughput, H100/H200 as the value middle, nothing older than Hopper unless waiting is free.
About Krea 2 Turbo. Krea 2 Turbo: from wikeeyang, on Hugging Face since July 2026. 20,665 downloads in the last 30 days.
How it compares. H100 80GB HBM3: Krea 2 Turbo 6.85 images/min, FLUX.1 dev 9.06 (12B), Qwen-Image 2.1 4.59 (7B), FLUX.1 Kontext dev 4.26, Qwen-Image-Edit 3.68 (20B). 1 of 4 beat Krea 2 Turbo here.
Cost on a rented GPU. 1,000 images of Krea 2 Turbo: $3.81 on a RTX PRO 6000 Blackwell Workstation Edition ($1.08/hr, 3.5 hours), $10.49 on a B200 ($5.98/hr, 105 min, 2.8x the cost).
Krea 2 Turbo: cost per 1,000 images on rented GPUs
| GPU | Cheapest rate | Speed (images/min) | Cost per 1,000 images |
|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 4.71 | $3.81 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 3.35 | $4.71 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 6.85 | $5.20 |
| NVIDIA H200 | $3.59/hr | 7.25 | $8.25 |
| NVIDIA B200 | $5.98/hr | 9.5 | $10.49 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for Krea 2 Turbo. 6-30 images/min: 3 (B200, H200, H100 80GB HBM3); under 6 images/min: 2 (RTX PRO 6000 Blackwell Workstation Edition, A100 80GB SXM4). 30 images/min means two seconds or less per picture.
Time per image. Krea 2 Turbo at 1024px and 8 steps: 6.3s per image on the B200, 17.9s on the A100 80GB SXM4. A batch of 100 takes 11 min on the fastest card and 30 min on the slowest.
VRAM for Krea 2 Turbo. Measured peak 43.2GB, so 48GB is the smallest common card size; smallest card it ran on: H100 80GB HBM3 (80GB).
Power on Krea 2 Turbo. Most efficient: B200, 801W, 1.41 kWh per 1,000 images. At $0.15/kWh: $0.21 per 1,000 images.
Our verdict
Krea 2 Turbo: 9.5 images/min peak, ~46GB measured floor, the heaviest image model on our bench and a pure rented-GPU workload. You run it for its signature aesthetic, not its speed; budget Hopper-or-newer silicon and let the per-image quality justify the seconds.
FAQ
- What GPU does Krea 2 Turbo need?
- 48GB-class or bigger, we measured ~46GB peak with the full bf16 pipeline resident. That means rented datacenter cards (B200, H200, H100, A100 80GB) or 48GB workstation silicon; no consumer card applies.
- How fast is it?
- 9.5 images/min on a B200 (6.3s per image), 7.25/min on an H200. Older architecture collapses the pace: the A100 80GB manages 3.35/min despite having the memory, this workload wants new compute, not just capacity.
- Why run it over FLUX?
- Aesthetics. Krea 2 Turbo is tuned for a polished, photorealistic look straight out of the prompt. The trade is weight and speed (FLUX.1 Schnell generates ~8× faster in our data). Draft on Schnell, finish on Krea is a legitimate pipeline.
- Is renting it affordable?
- Reasonable for curated output: about 570 images per B200-hour. For bulk generation the economics favor Schnell decisively; Krea's rate buys look, not volume.
- Where else can I find benchmarks of it?
- As far as we know, nowhere. These are the only published Krea 2 Turbo GPU measurements. That's a core purpose of this site: running the workloads nobody else measures, on a fixed methodology.