Benchmark status: this page is a draft shell. Peak-VRAM (measured) cells stay TBD until AigenRig Bench Protocol v1.0 runs are published. The “Expected class” column is provisional, not a measurement.

How Much VRAM Do You Need for Flux, SDXL & Wan Video?

Last updated: 2026-10-30

Quick answer (provisional until measured)

Workload (Protocol v1.0 defaults)	Model file (as tested)	Spec	Expected class	Peak VRAM (measured)
Flux main	flux1-dev-fp8.safetensors	1024² · 20 steps	Mid / high VRAM	TBD
SDXL	sd_xl_base_1.0.safetensors	1024² · 20 steps	Mid VRAM	TBD
Wan 480p	wan2.1_t2v_1.3B_fp16	832×480 · 81 frames	Mid VRAM	TBD
Wan 720p	wan2.1_t2v_14B_fp8_scaled	1280×720 · 81 frames	High VRAM	TBD

Do not treat marketing “supports Flux” claims as peak-VRAM proof. We publish the ComfyUI peak from the same seed and step counts every time.

Why VRAM beats raw TFLOPS for local AI

Weights + activations must fit (or you pay for offload latency).
Video (Wan) keeps many frames in flight; 720p jumps hard vs 480p.
FP8 / scaled checkpoints change the floor; always name the exact file in the table.

Offload vs bigger card

If a card only finishes with --lowvram or sequential offload, it gets a separate protocol row, never mixed into the main leaderboard. That keeps “fits in VRAM” honest.

Buying guide (fill after first three cards)

12GB class: what completes without offload (from 3060 row)
24GB class: what becomes comfortable (from 3090 row)
24GB+ / 4090 class: Wan 720p headroom (from 4090 row)

Affiliate note

As an Amazon Associate I earn from qualifying purchases. Partner cloud links may also pay AigenRig. Details: /disclosure/.