“Should I buy a GPU for Flux.2 or for Stable Diffusion 3.5?” is a question with a much sharper answer than most comparisons admit, because the two models are not in the same weight class and their published files say so plainly.
Quick answer
- Flux.2 (quality-first, willing to wait): 24GB is the floor, not the luxury tier — an RTX 4090, or a used RTX 3090 if price leads. You will be running the 4-bit build with the text encoder off the card.
- SD 3.5 Large (fast, flexible, ControlNet-heavy): RTX 5070 Ti or RTX 4070 Ti Super. Its released checkpoint is 16.5GB, so a 16GB card is tight at full precision and a 24GB card is roomy.
- SD 3.5 Medium (iteration speed, hobbyists): RTX 4060 Ti 16GB, or an RTX 3060 12GB. The checkpoint is 5.11GB — this is the model that does not fight your hardware.
- One card for both: RTX 4090 or a used RTX 3090. 16GB does not cover both models honestly, whatever the FP8 headlines say.
NVIDIA GeForce RTX 4090
24GB GDDR6X24GB holds SD 3.5 Large at full precision with LoRA room, and the 4-bit Flux.2 build with the encoder offloaded. The smallest card that covers both.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Who this is for
You’re either choosing a GPU specifically to run one of these two flagship image models, or you already have a card and need to know which model your hardware can realistically support in 2026. I’m assuming you care about local generation (privacy, batch work, custom LoRAs) rather than just hitting an API.
If you’re still on Flux.1 Dev and not sure whether Flux.2 is worth the upgrade, that’s covered below too — the VRAM jump is larger than the parameter counts suggest.
VRAM side-by-side
The full-precision column below is not estimated — it is the size of the file each project publishes, checked 2026-09-13. The lower-precision columns are that file halved and quartered.
| Model | Params | Released weights | FP8 | 4-bit |
|---|---|---|---|---|
| Flux.2 Dev | 32B | ~64 GB (7 BF16 shards) | ~32 GB | ~16 GB |
| Flux.1 Dev | 12B | 23.8 GB | ~12 GB | ~6 GB |
| SD 3.5 Large | 8B | 16.5 GB (sd3.5_large.safetensors) | ~8 GB | ~4 GB |
| SD 3.5 Medium | 2.5B | 5.11 GB | ~2.6 GB | ~1.3 GB |
In every case the text encoders are separate files loaded on top, which is why the “fits in X GB” claims you see quoted are usually a gigabyte or several optimistic.
The gap is the story. Flux.2’s transformer is roughly four times the size of SD 3.5 Large’s entire checkpoint. NVIDIA’s May 2026 FP8 work is often read as bringing Flux.2 to 16GB cards, and it does not: half of 64GB is 32GB, which clears an RTX 5090 only if nothing else wants memory. The precision that reaches consumer hardware is 4-bit, and Black Forest Labs’ own local-deployment example says as much — a 4-bit model with a remote text encoder on an RTX 4090.
There is a neat coincidence buried in that table worth pulling out. Flux.2 at 4-bit and SD 3.5 Large at full precision occupy almost exactly the same memory — about 16GB of weights each. On a 24GB card you can run either. The difference is that one of them is running as its authors released it and the other has been compressed fourfold to get there.
For how activation overhead and encoder placement move these figures in a real pipeline, see the how much VRAM for Flux guide. The same principles apply to SD 3.5, at a quarter of the scale.
Speed side-by-side
Times below are modelled from memory bandwidth (methodology) for a 1024x1024 generation, not measured on our own hardware — treat the ordering between cards as reliable and the absolute seconds as indicative. The Flux.2 column is the 4-bit build with the encoder offloaded, matching the Flux.2 buyer’s guide; SD 3.5 runs at its released precision.
| GPU | VRAM | Flux.2 (4-bit) | SD 3.5 Large | SD 3.5 Medium |
|---|---|---|---|---|
| RTX 5090 | 32GB | ~6 s | ~3 s | ~1 s |
| RTX 4090 | 24GB | ~9 s | ~4 s | ~1.5 s |
| RTX 3090 (used) | 24GB | ~14 s | ~6 s | ~2 s |
| RTX 5080 | 16GB | ~11 s, nothing spare | too tight at 16.5GB | ~2 s |
| RTX 5070 Ti | 16GB | ~13 s, nothing spare | too tight at 16.5GB | ~2.5 s |
| RTX 4060 Ti 16GB | 16GB | ~28 s, swapping | too tight at 16.5GB | ~4 s |
| RTX 3060 12GB | 12GB | no | at FP8 only | ~5 s |
Two patterns matter. First, SD 3.5 Large is roughly two to three times faster than Flux.2 on the same card — it is computing a quarter of the parameters, and the 4-bit dequantization step costs Flux.2 something back on top. Second, the 16GB tier is where the two models part company: it runs Flux.2’s 4-bit build with nothing left over, and it cannot hold SD 3.5 Large’s 16.5GB checkpoint at full precision at all. 16GB is not the “does both” tier that it is usually sold as.
Check NVIDIA GeForce RTX 5090 on Amazon→Buy on Shopee SG→Which model should drive your buy?
This is where most “comparison” articles go vague. The decision follows from the file sizes above more cleanly than you would expect:
- Quality-first, willing to wait: Flux.2 wins on prompt adherence, complex composition, and especially text rendering. Buy for Flux.2 — and that means 24GB, not the 16GB usually quoted. An RTX 4090, or a used RTX 3090 at a third the price.
- Speed-first, iteration matters more than peak quality: SD 3.5 Large at two to three times the throughput. A 24GB card holds its 16.5GB checkpoint with LoRA room; 16GB does not.
- VRAM-tight (12GB card, can’t upgrade): SD 3.5 Large at FP8 (~8GB) is your ceiling and it is a good one. Flux.2 is out — its 4-bit build alone is 16GB.
- Budget under $500: RTX 4060 Ti 16GB. SD 3.5 Medium for daily work at 5.11GB, SD 3.5 Large at FP8 for hero shots. Flux.2 technically loads at 4-bit and then swaps constantly; plan around the other two.
- Multi-modal / ControlNet-heavy: SD 3.5’s ecosystem is dramatically more mature, and control modules are extra resident weights — which is the argument for 24GB all over again.
If you’re building a serious ComfyUI workflow with several models resident at once, the calculus shifts again — the best GPU for ComfyUI guide covers the loadout question, and the short version is that 24GB is the floor for node graphs that keep a control stack loaded.
Flux.2 isn’t always the better choice
Worth saying plainly: Flux.2 gets called “the new standard” a lot, and that’s not wrong on absolute quality, but it ignores three things.
It’s slow. A ten-second generation breaks the prompt-iteration loop in a way a three-second one does not. If you generate 200 images a day refining a concept, SD 3.5 Large gets you to the final image faster even though Flux.2’s individual outputs are better.
The LoRA ecosystem is still catching up. Civitai had thousands of Flux.1 LoRAs within months. Flux.2 LoRAs exist but the long tail of style/character/concept training that makes SD ecosystems sticky is still building.
And the VRAM floor is brutal. On an RTX 3080 10GB or any 12GB card, Flux.2 is not a quality compromise — it simply does not load, because even the 4-bit build is 16GB. In that lane, Flux.1 Dev on the same hardware is the better model to reach for: at FP8 it is around 12GB, which those cards actually hold.
Common mistakes
- Buying a 16GB card because “FP8 makes Flux.2 fit.” It does not. FP8 halves 64GB to 32GB, which is past every consumer card once the text encoder loads. What fits 16GB is the 4-bit build, with nothing spare — and 24GB is what Black Forest Labs’ own example uses.
- Assuming SD 3.5 = SDXL hardware. It doesn’t. SD 3.5 Large’s checkpoint is 16.5GB against SDXL’s roughly 7GB. If you sized a build for SDXL, check before you upgrade the model — a 16GB card that was comfortable becomes the wrong side of the line.
- Ignoring FP8 because “it’s lower precision.” On Blackwell and Ada Lovelace, FP8 quality loss on both these models is hard to see. For SD 3.5 Large it is the difference between fitting a 12GB card and not, which makes it the most valuable setting in the menu.
- Buying a 5090 for SD 3.5 Medium. A 5.11GB model will not make that GPU sweat; you will be limited by everything except the card. Buy a 4070 Ti Super and pocket the difference.
For a broader view across all SD-family models on consumer GPUs, the best GPU for Stable Diffusion overview covers the older SDXL / SD 1.5 considerations that still matter for the LoRA back-catalog. And if Flux.2 specifically is your target, the best GPU for Flux.2 buyer’s guide works through all seven cards and what the 4-bit path costs in image quality.
Final verdict
| Use case | Pick | Real budget |
|---|---|---|
| Flux.2 with ControlNet room | RTX 5090 | ~$4,900 |
| Flux.2 daily driver (4-bit, encoder offloaded) | RTX 4090 | ~$2,200 |
| Both models, lowest price that works | RTX 3090 | ~$820 used |
| SD 3.5 Large primary, Flux.2 not needed | RTX 5070 Ti | ~$1,050 |
| SD 3.5 Large at FP8 + Medium | RTX 4060 Ti 16GB | ~$425 |
| SD 3.5 Medium only | RTX 3060 12GB | ~$250 used |
NVIDIA GeForce RTX 3090
24GB GDDR6X24GB used for around $820 — holds SD 3.5 Large at full precision and the 4-bit Flux.2 build. The cheapest card that covers both models honestly.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
The one-sentence verdict: if you can only buy one card for both models in 2026, buy 24GB rather than 16GB, and a used RTX 3090 gets you there for a quarter of a 5090’s price — then build around whichever model you actually use 80% of the time, not the one that benchmarks prettier.
Frequently asked questions
How much VRAM does Flux.2 actually need?
More than the usual answer. Its transformer ships as roughly 64GB of BF16 weights, so FP8 is about 32GB and the 4-bit build about 16GB, with the text encoder on top of each. In practice that makes 24GB the floor, running 4-bit with the encoder off the card.
Can a 12GB GPU run SD 3.5 Large?
At FP8 yes, comfortably — that build is around 8GB and leaves room for LoRAs. At full precision no: the released checkpoint is 16.5GB, so it needs a 24GB card to run with any headroom.
Is Flux.2 always higher quality than SD 3.5 Large?
On prompt adherence and text rendering, generally yes. On stylized art with mature LoRA support, SD 3.5 often wins on practical output. The gap is smaller than benchmark posts suggest.
Should I upgrade from Flux.1 Dev to Flux.2?
Only if you have 24GB. Flux.1 Dev’s weights file is 23.8GB against Flux.2’s roughly 64GB, so on 12GB and 16GB cards Flux.1 at FP8 remains the better daily experience by a wide margin.
Which model is faster on the same GPU?
SD 3.5 Large is roughly two to three times faster than Flux.2 on identical hardware. SD 3.5 Medium is faster still — close to four to five times Flux.2’s speed.