New: the Video Studio is here, for whole AI films.Get it now →

GPU check

The fastest GPUs for local AI images

14 NVIDIA cards ranked by how long an image actually takes on them, grouped by how much memory they carry. Real generations from owners' own machines over the last 180 days, not spec sheets.

6 GB of VRAM

Ranked on MiniMax H3 Image, the workload these 2 cards have in common.

Every card here was measured at 1024x1024 and 20 steps, with VRAM reclaim enabled, so the comparison holds the work fixed.

  1. 1RTX 3060 Laptop GPU6 GB2 min 15 s23 runs on 3 machines
  2. 2RTX 20606 GB5 min 10 s14 runs on 3 machines

8 to 10 GB of VRAM

Ranked on Boogu Image Edit, the workload these 3 cards have in common.

Every card here was measured at 1344x768 and 25 steps, with no accelerators enabled, so the comparison holds the work fixed.

  1. 1RTX 308010 GB3 min 4 s28 runs on 3 machines
  2. 2RTX 4060 Laptop GPU8 GB5 min 6 s12 runs on 4 machines
  3. 3RTX 30508 GB19 min36 runs on 3 machines

12 GB of VRAM

Ranked on Qwen Image 2.1 Edit, the workload these 5 cards have in common.

Every card here was measured at 1344x768 and 20 steps, with no accelerators enabled, so the comparison holds the work fixed.

  1. 1RTX 4070 Ti12 GB1 min 34 s203 runs on 3 machines
  2. 2RTX 407012 GB1 min 37 s55 runs on 3 machines
  3. 3RTX 507012 GB1 min 45 s275 runs on 7 machines
  4. 4RTX 5070 Ti Laptop GPU12 GB2 min 24 s42 runs on 3 machines
  5. 5RTX 306012 GB3 min 50 s103 runs on 11 machines

16 to 32 GB of VRAM

Ranked on MiniMax H3 Reference Edit, the workload these 4 cards have in common.

Every card here was measured at 1440x1440 and 20 steps, with VRAM reclaim enabled, so the comparison holds the work fixed.

  1. 1RTX 409024 GB1 min 18 s15 runs on 4 machines
  2. 2RTX 5070 Ti16 GB1 min 31 s384 runs on 4 machines
  3. 3RTX 509032 GB2 min 15 s253 runs on 4 machines
  4. 4RTX 5060 Ti16 GB2 min 48 s93 runs on 3 machines

Common questions

What is the fastest GPU with 6 GB of VRAM for local AI image generation?
The RTX 3060 Laptop GPU, at 2 min 15 s on MiniMax H3 Image, measured across 3 machines. It is the quickest of the 2 cards in that memory band.
What is the fastest GPU with 8 to 10 GB of VRAM for local AI image generation?
The RTX 3080, at 3 min 4 s on Boogu Image Edit, measured across 3 machines. It is the quickest of the 3 cards in that memory band.
What is the fastest GPU with 12 GB of VRAM for local AI image generation?
The RTX 4070 Ti, at 1 min 34 s on Qwen Image 2.1 Edit, measured across 3 machines. It is the quickest of the 5 cards in that memory band.
What is the fastest GPU with 16 to 32 GB of VRAM for local AI image generation?
The RTX 4090, at 1 min 18 s on MiniMax H3 Reference Edit, measured across 4 machines. It is the quickest of the 4 cards in that memory band.
Does more VRAM make local AI image generation faster?
More memory buys two things: speed, and models that will not load at all without it. The quickest card at 6 GB of VRAM is the RTX 3060 Laptop GPU, at 2 min 15 s on MiniMax H3 Image. The quickest at 16 to 32 GB of VRAM is the RTX 4090, at 1 min 18 s on MiniMax H3 Reference Edit. Those are two different presets, because the larger models do not fit in the smaller band, so read them as two measurements rather than one ratio.

Times are medians over real generations reported by the App in the last 180 days. A card only appears in a band on a workload it shares with the others: the same preset, the same clip length, the same output resolution, the same step count and the same accelerator settings, with at least ten runs across three machines behind every number. That is stricter than it sounds and it is the point: owners of bigger cards run bigger settings, so ranking on the preset name alone measures what people choose rather than what the card does. Prices are deliberately absent: they move, and these timings do not. Adult packs and retired presets are left out. Check your own machine The full list of what runs and what does not is in System requirements