Benchmark status: this page is a draft shell. Every score cell is TBD until AigenRig Bench Protocol v1.0 measured rows are inserted. No numbers on this page are estimates or vendor claims.
Best GPU for AI Image & Video Generation (2026): Tested in ComfyUI
TL;DR
Pick by VRAM first, then by measured seconds and watts under the same ComfyUI workflows. This page will rank cards using AigenRig Bench Protocol v1.0 (Flux 1024², SDXL 1024², Wan 480p / 720p). Until the first 3060 → 3090 → 4090 run set is finished, treat every score cell as TBD.
How we test (short)
- Seed
42, 1 warmup + median of 3 timed runs, “Prompt executed in…” from ComfyUI. - Main table: no
--lowvram; offload configs get their own row. - Full rules: /methodology/ (AigenRig Bench Protocol v1.0).
Scorecard (fill after runs)
| GPU | Flux s/img | SDXL s/img | Wan 480p s | Peak VRAM | Power W | $/100 local | Notes |
|---|---|---|---|---|---|---|---|
| RTX 3060 12GB | TBD | TBD | TBD | TBD | TBD | TBD | Day-1 card |
| RTX 3090 (used) | TBD | TBD | TBD | TBD | TBD | TBD | Day-1 card |
| RTX 4090 (RunPod) | TBD | TBD | TBD | TBD | TBD | TBD | Cloud reference |
| RTX 5090 | TBD | TBD | TBD | TBD | TBD | TBD | When available |
Local cost formula: (seconds/3600) × (W/1000) × $/kWh × 100. Cloud: (seconds/3600) × $/hr × 100.
Who should buy what (after data)
- Budget / learning ComfyUI: [fill from 3060 row]
- Serious local Flux + light Wan: [fill from 3090 row]
- Heavy Wan 720p / headroom: [fill from 4090 / 5090 rows]
- No tower, burst workloads: cloud GPU hour math in the cloud article (coming soon)
Affiliate note
As an Amazon Associate I earn from qualifying purchases. aigenrig.com may also earn commissions or credits from partner links (including Runpod and Vast.ai) at no extra cost to you. See /disclosure/.
Related
- How much VRAM for Flux, SDXL & Wan
- Cloud GPU vs local GPU cost breakdown (coming soon)
- Bench Protocol / methodology