Half the people who email me about Flux.2 already own a 24GB card and are running Flux.1 Dev just fine. They don’t need a new GPU — they need someone to talk them out of one. The other half own a 12GB card, want ControlNet, and are quietly wondering whether Flux.2 is finally the reason to upgrade.
This piece is for both groups: what actually shifted on the hardware side between Flux.1 Dev and Flux.2, and where the upgrade math breaks.
Quick answer
- On a 24GB card (4090 / 3090 / 7900 XTX) running Flux.1 fine: you are already on the only tier that runs Flux.2 properly. Nothing to buy.
- On a 12GB card (4070 / 3060) frustrated by ControlNet OOMs: upgrade to 24GB. A 16GB card fixes Flux.1 but leaves Flux.2’s 4-bit build with no headroom.
- On an 8-10GB card: stay on Flux.1 Schnell or SDXL. Both Flux.2 and full Flux.1 Dev will thrash your system RAM.
- Building fresh in mid-2026: 24GB. A used RTX 3090 at ~$820 runs Flux.2 where a newer 16GB RTX 5080 at ~$1,400 does not.
NVIDIA GeForce RTX 3090
24GB GDDR6X24GB is the tier where Flux.2 4-bit leaves room for a ControlNet, and Flux.1 Dev runs at FP16 outright.
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 already run Flux.1 Dev locally and you’re staring at the Flux.2 release notes wondering whether the 32B parameter count is a real problem or just a headline. Or you’re shopping for a first serious image-gen card and you can’t decide whether to size your build for Flux.1 (cheaper, easier) or Flux.2 (better outputs, hungrier).
Either way, the decision is mostly about VRAM, ControlNet, and how many seconds you’re willing to wait per generation. Compute matters less than most benchmark posts pretend.
What actually changed in Flux.2
Flux.2 Dev is a 32B parameter model — roughly 2.7x the size of Flux.1 Dev’s 12B — and the weights bear that out exactly. Black Forest Labs ships the Flux.2 transformer as seven BF16 shards totalling ~64GB, against 23.8GB for Flux.1 Dev’s single weights file. Those are file listings, not estimates.
FP8 halves each: ~32GB for Flux.2, ~12GB for Flux.1. Four-bit halves them again: ~16GB and ~6GB. The text encoder sits on top in both cases, and Flux.2’s is a large language model in its own right, which is why Black Forest Labs’ only consumer example keeps it off the GPU entirely.
That is the whole story of this upgrade question, and it is harsher than the usual framing. Flux.1 Dev is comfortable on a 16GB card at FP8. Flux.2 is not comfortable anywhere below 24GB: FP8 alone is more than any consumer card has, so you are on the 4-bit build, and 16GB of weights in a 16GB card leaves nothing for anything else. The floor moves from 16GB to 24GB, not from “16GB” to “16GB with FP8 support”.
VRAM head-to-head
Numbers below are for a 1024x1024 generation, stock ComfyUI pipeline, no offloading, no LoRA stack. Add roughly 1-2GB for a full ControlNet + IP-Adapter combo on either model.
| Model | Params | FP16 VRAM | FP8 VRAM | Q4 VRAM | Realistic floor |
|---|---|---|---|---|---|
| Flux.1 Dev | 12B | 23.8 GB | ~12 GB | ~6 GB | 16GB card at FP8 |
| Flux.1 Schnell | 12B | ~23.8 GB | ~12 GB | ~6 GB | 16GB card at FP8 |
| Flux.2 Dev | 32B | ~64 GB | ~32 GB | ~16 GB | 24GB card at 4-bit |
The gap that matters is the ControlNet stack. Flux.1 Dev at FP16 is already 23.8GB on its own, so two ControlNets put it past any consumer card — which is why most Flux.1 users run FP8 (~12GB) and why 16GB has been the practical Flux.1 tier. Flux.2 has no equivalent comfortable setting: its FP8 is ~32GB, so the 4-bit build is the only option, and that is 16GB before ControlNet. This is why RTX 3090 and 4090 have been the “safe” cards for two years. Flux.2 FP8 with the same ControlNet stack hits ~19-20GB, so 24GB is still the comfortable ceiling. If you’re already on a 4090 and only using Flux.1 without ControlNet, you’ve been leaving VRAM on the table — Flux.2 uses it.
For the exact activation and KV cache math behind these numbers, the how much VRAM for Flux guide breaks it down. The overhead ratios haven’t changed between Flux.1 and Flux.2 — just the base weight footprint.
Speed head-to-head
Generation times below are modelled from memory bandwidth and per-vendor efficiency, the way every throughput figure on this site is derived — see our methodology. They assume 30 steps at 1024x1024 on a stock ComfyUI pipeline, and they are estimates rather than stopwatch readings.
| GPU | Flux.1 Dev FP8 | Flux.2 4-bit | Delta |
|---|---|---|---|
| RTX 5090 | ~8-10 s | ~12-14 s | +40% |
| RTX 4090 | ~12-14 s | ~14-16 s | +15% |
| RTX 5080 | ~10-12 s | ~14-16 s | +30% |
| RTX 5070 Ti | ~13-15 s | ~18-22 s | +40% |
| RTX 4070 Ti Super | ~14-16 s | ~22-26 s | +55% |
| RTX 4060 Ti 16GB | ~28-32 s | ~55-70 s (Q4) | +100%+ |
The surprising line is the 4090. Quantised down to 4-bit, Flux.2 only runs ~15% slower than Flux.1 on the same card — nowhere near the 2.7x parameter gap would suggest. That’s the actual argument for Blackwell / Ada Lovelace over older cards: on a 3090, Flux.2 FP8 pays a larger penalty because tensor core FP8 is emulated, and the gen-time delta climbs closer to 50%.
Check NVIDIA GeForce RTX 4090 on Amazon→Buy on Shopee SG→Quality delta — is it visible?
This is the part nobody wants to say out loud: on 80% of prompts, Flux.2 output is not obviously better than Flux.1 Dev. Where it clearly wins is text rendering (Flux.1 fumbles longer strings), multi-subject scenes with spatial relationships (“the cat behind the vase to the left of the lamp”), and photographic realism at portrait distance.
Where the gap is small or nonexistent: stylized art, landscapes, single-subject portraits, most LoRA-driven workflows. If those are your daily bread, the 32B model is burning VRAM to solve problems you don’t have.
Should you upgrade?
Working decision tree, in order:
- Do you have a 24GB card and Flux.1 does what you need? Hold. There is nothing here worth $1,500.
- Do you have a 12GB card and hit OOM on ControlNet with Flux.1? Upgrade to 24GB, not 16GB. A 16GB card fixes the Flux.1 problem and immediately recreates it on Flux.2.
- Do you have a 16GB card (4060 Ti / 4070 Ti Super / 5070 Ti)? Flux.1 is comfortable there; Flux.2 will load at 4-bit and leave you nothing. If Flux.2 matters to you, the upgrade is to 24GB. If it does not, upgrade your workflow instead — add sage-attention, xformers, torch.compile.
- Do you have an 8-10GB card? Skip both Flux.2 and Flux.1 Dev. Flux.1 Schnell at FP8 is the ceiling that runs comfortably.
- Buying fresh with $1,400-1,200? A used RTX 3090 at ~$820 plus the change, or stretch to a 4090. At this budget the 16GB RTX 5080 is faster but cannot give Flux.2 room to work.
- Buying fresh with $4,900+? RTX 5090 if you specifically want Flux.2 FP16 or plan to move to video models next.
The deeper card-by-card breakdown for Flux.2 lives in the best GPU for Flux.2 guide, and the corresponding Flux.1 buyer’s map is in the best GPU for Flux guide. Both stay current with the used-market pricing that actually matters for this decision.
When to hold on Flux.1
Contrarian bit — if you’re already happy with Flux.1 on a 24GB card, hold. I mean it. The upgrade math doesn’t work.
Flux.1 has two years of LoRA back-catalog on Civitai. Flux.2 has months. The ControlNet ecosystem is mature on Flux.1, sparse on Flux.2. The community fine-tunes (Pixelwave, Aurora, various photorealism forks) are Flux.1-native, and porting is slow.
If you generate 100 images a day for a specific style you’ve dialed in with three Flux.1 LoRAs, moving to Flux.2 costs you all of that and buys you incremental prompt adherence you probably don’t need. Upgrade when your workflow demands it, not when the release notes look shiny.
Common mistakes
- Assuming 16GB fits Flux.2 comfortably. It fits the model. It does not fit the model + ControlNet + IP-Adapter + LoRA stack. That’s 24GB territory.
- Buying a 16GB card for Flux.2 because it is newer. An RTX 5080 is quicker than a used 3090 on every workload that fits in 16GB. Flux.2 with any ControlNet is not one of them.
- Trying to make Flux.2 fit a 12GB card. The 4-bit build alone is ~16GB, so it does not fit at all. At 12GB, Flux.1 Dev at FP8 is the better model — FP8 is the milder compression of the two.
- Ignoring workflow overhead in ComfyUI. A ComfyUI graph with three models loaded pushes both Flux.1 and Flux.2 setups past their nominal VRAM budgets — see the ComfyUI GPU guide for real memory floors.
For the broader head-to-head against Stable Diffusion 3.5 on the same hardware, the Flux.2 vs SD 3.5 comparison covers when a lighter model buys you more than a heavier one.
Final verdict
| Your current setup | What to do | Card |
|---|---|---|
| RTX 4090 / 3090, Flux.1 fine | Hold | none |
| RTX 4090, want Flux.2 + ControlNet | Keep — it works | RTX 4090 |
| 12GB card, ControlNet OOMs | Upgrade | 24GB — used RTX 3090 or RTX 4090 |
| 16GB card (4060 Ti / 4070 Ti S) | Optimize workflow | none |
| 8-10GB card | Stay on Flux.1 Schnell | none |
| Fresh build, ~$1,400 | Buy for Flux.2 4-bit | Used RTX 3090, or save toward a 4090 |
| Fresh build, ~$4,900+ | Buy for FP16 headroom | RTX 5090 |
NVIDIA GeForce RTX 3090
24GB GDDR6X24GB used for ~$820 — the cheapest card that runs Flux.2 4-bit with room for a ControlNet, and Flux.1 Dev at full precision.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
The one-line takeaway: Flux.2 is a real generational step, but it’s a hardware upgrade only if your current card is already fighting Flux.1. If Flux.1 runs clean at 24GB, keep it, save the money, and revisit when Flux.3 forces the issue.
Frequently asked questions
Do I need to upgrade from Flux.1 to Flux.2?
Only if you’re on 12GB and hitting ControlNet OOMs, or if you need Flux.2’s text rendering and multi-subject strengths. On a 24GB card running Flux.1 without issues, holding is the right move.
Will Flux.2 run on a 12GB GPU?
No. Flux.2’s 4-bit build is about 16GB on its own, so a 12GB card cannot hold it. Flux.1 Dev at FP8 is the model to run at that tier. Note that FP16 is not an option at 12GB for either model — Flux.1 Dev alone is about 23.8GB at full precision.
Is Flux.2 slower than Flux.1 on the same GPU?
Roughly 15 to 40 percent slower depending on the card. On Blackwell and Ada Lovelace the FP8 optimization keeps the gap small; on Ampere it grows because FP8 is emulated.
Can a 24GB RTX 4090 run Flux.2 with ControlNet?
Yes, on the 4-bit build with the text encoder offloaded — about 16GB of weights plus a ControlNet and IP-Adapter lands near 19 to 20GB, which a 24GB card holds. FP8 would be ~32GB and does not fit.
Should I buy an RTX 5080 or a used RTX 3090 for Flux.2?
The 3090, for Flux.2 specifically. Its 24GB is what lets the 4-bit build load with room for a ControlNet; the 5080’s 16GB does not, and no amount of Blackwell speed fixes a capacity limit. Buy the 5080 if Flux.1 and SDXL are your real workload after used-market variance.