Home › Image & video models › FLUX vs SDXL
FLUX vs SDXL: VRAM requirements, official column vs honest gap column
Neither vendor publishes a VRAM spec. Both were checked today the same way: HF API metadata, the repos’ READMEs, vendor docs (details on the individual pages). What is official is the parameter counts — so that is what this table prints.
| Quantity | FLUX.1 [dev] / [schnell] | SDXL base 1.0 |
|---|---|---|
| Core model params (official repo metadata) | 11,901,408,320 / 11,891,178,560 (safetensors BF16, huggingface.co/api/models/black-forest-labs/FLUX.1-dev → safetensors BF16 params 11,901,408,320, retrieved 2026-10-24; schnell retrieved same day) | 2,567,463,684 (huggingface.co/api/models/stabilityai/stable-diffusion-xl-base-1.0 → safetensors F32 params 2,567,463,684, retrieved 2026-10-24; the paper’s “2.6B parameters in the UNet” matches) |
| fp16/bf16 weight floor (our arithmetic: params × 2 B) | 22.17 GiB / 22.15 GiB | 4.78 GiB (fp32 ≈ 9.56 GiB) |
| Weight-floor ratio (arithmetic) | FLUX.1 [dev] weights ≈ 4.6× SDXL base fp16 weights | |
| Text encoders officially named? | Not in the publicly reachable BFL docs (cards are gated) → no public data (we checked, 2026-10-24) | Yes — CLIP ViT-L + OpenCLIP ViT-bigG named in SDXL paper (arXiv:2307.01952), full text retrieved 2026-10-24: “These changes result in a model size of 2.6B parameters in the UNet”; its architecture table lists “Text encoder: CLIP ViT-L & OpenCLIP ViT-bigG” and publishes no memory requirement; their size is not published → no public data |
| Full-pipeline VRAM spec (the number people want) | none published (checked 2026-10-24) | none published (checked 2026-10-24) |
| Measured-anchors on this site | deliberately empty | deliberately empty |
The left columns are as true as the official counts they derive from; the right-hand gaps are why we refuse any “N GB minimum” headline for either model. Cheapest cited anchors for testing (modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending. Re-check 2026-10-22 (us): source table values unchanged since its stated 2026-09-29 check): RunPod RTX 5090 32GB @ $0.99/h (≥12 GB, SDXL-class headroom) · RunPod RTX 5090 32GB @ $0.99/h (≥24 GB, FLUX weight floor). Same-series pages: FLUX.1 schnell VRAM · FLUX.1 dev VRAM · Kontext dev VRAM · SDXL VRAM · FLUX.1 overview · method.
FAQ
Does FLUX need more VRAM than SDXL?
On weights alone the answer is arithmetic, not opinion: FLUX.1 [dev] ≈ 22.17 GiB bf16 vs SDXL base fp16 ≈ 4.78 GiB — ≈ 4.6× the weight floor. For the full pipeline nobody official has published a number for either model → no public data (we checked both, 2026-10-24).
Can a 24 GB card run SDXL or FLUX?
fp16/bf16 weights only: SDXL fits with ≈ 19.2 GiB spare under our nominal-GB heuristic; FLUX leaves ≈ 1.8 GiB for encoders, VAE and activations, which is why bf16 FLUX is normally discussed on 32 GB — a phrase we print as reasoning, not as a spec. Measured full-pipeline peaks: no public data for either (we checked, 2026-10-24).
Which should I rent to test?
Same table, same day: SDXL testing starts at RunPod RTX 5090 32GB @ $0.99/h (cheapest ≥12 GB); FLUX at RunPod RTX 5090 32GB @ $0.99/h (cheapest ≥24 GB); modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending. Re-check 2026-10-22 (us): source table values unchanged since its stated 2026-09-29 check.
Why won't you just print a GB verdict?
Because both vendors publish nothing and the community numbers are unverifiable one-rig anecdotes. We print official counts, label our arithmetic as arithmetic, and label every gap as a gap. Post a measured peak with your handle and date and it may land on these pages as a cited anchor.