This article was originally published on runaihome.com
If you are picking an image generation model for a local rig in 2026, your real choice is
between three families: Stable Diffusion 1.5 (the lightweight veteran), SDXL (the
1024 native step up), and Flux (the new king of quality, with the heaviest VRAM bill).
This article walks through what each one is good at, what it costs, and which is the right
fit for your hardware.
If you only want the headline:
- 6–8 GB VRAM: SD 1.5
- 10–12 GB VRAM: SDXL
- 16 GB+ VRAM: Flux (with caveats), or SDXL with heavy LoRA / ControlNet stacks
- 24 GB+ VRAM: Flux comfortably, or experiment with whatever ships next
The rest is the why.
A 60-second history
Stable Diffusion 1.5 dropped in October 2022 as the open-source moment that put generative
AI on home GPUs. It is small (4 GB FP16 weights), trained at 512 × 512, and the entire
fine-tuning ecosystem — every LoRA, every checkpoint mix, every ControlNet — calibrated
itself around SD 1.5 first.
SDXL (Stable Diffusion XL) released in July 2023, bumping native resolution to 1024 ×
1024, doubling weight count to ~3.5 B parameters across the base + refiner pair, and
substantially improving prompt adherence and photorealism. SDXL is the workhorse of the
"good enough quality, reasonable VRAM" tier today.
Flux (Flux.1 dev / schnell / pro by Black Forest Labs, the team behind the original
Stable Diffusion) launched in August 2024. Flux is a different architecture — a flow-matching
transformer rather than a UNet — and beats both SD 1.5 and SDXL on prompt following, hands,
text rendering, and fine detail. It is also dramatically heavier: ~12 B parameters.
By 2026 each of these still has a real audience. They have not replaced each other; they
serve different sweet spots.
VRAM and speed comparison
| Model | Native resolution | Parameters | FP16 VRAM (model alone) | 4-bit VRAM | First-gen speed (16 GB GPU) |
|---|---|---|---|---|---|
| SD 1.5 | 512 × 512 | ~860 M | ~2 GB | ~1 GB | 1–3 sec |
| SDXL base | 1024 × 1024 | ~3.5 B | ~7 GB | ~3.5 GB | 4–10 sec |
| SDXL Turbo | 512 × 512 | ~3.5 B | ~7 GB | ~3.5 GB | <1 sec (1–4 steps) |
| Flux dev | 1024 × 1024 | ~12 B | ~24 GB | ~10 GB | 15–40 sec |
| Flux schnell | 1024 × 1024 | ~12 B | ~24 GB | ~10 GB | 4–10 sec (4 steps) |
These are weight-only numbers. In practice you also need a few GB for the VAE, encoders, and
working memory. SDXL on a 12 GB card is comfortable; Flux dev on a 12 GB card requires the
quantized variants (flux-dev-fp8 or GGUF builds), which run but are noticeably slower.
For background on what 4-bit and 8-bit quantization actually do, see our
quantization explainer — the core ideas apply
to image models too.
Quality: where each one shines
SD 1.5: the LoRA universe
SD 1.5's quality at the base level is below SDXL and far below Flux. But its fine-tuning
ecosystem is unmatched. Civitai alone lists tens of thousands of LoRAs, checkpoints, and
embeddings calibrated for 1.5 — covering anime styles, photoreal humans, specific artists,
specific characters, specific lighting setups. If you want a particular look, there is
probably a LoRA for it on 1.5 already.
For artistic generation, illustrations, anime, or any style that has been heavily explored by
the fine-tuning community, SD 1.5 with a curated LoRA stack still produces results that often
beat what you would get from base SDXL or even Flux at "first try." It is also fast enough
that iteration is pleasant.
SDXL: the photoreal default
SDXL out of the box produces clearly better photorealism than SD 1.5: better skin, better
hair, better hands (most of the time), much better prompt adherence. The native 1024 × 1024
resolution alone removes the upscaling step that 1.5 workflows usually need.
The SDXL ecosystem is mature. Custom checkpoints like JuggernautXL, DreamShaperXL,
and RealVisXL are excellent general-purpose photoreal bases. SDXL LoRAs and ControlNets
are widely available, and almost every workflow tutorial you find online assumes SDXL today.
For 90% of users on 12–16 GB cards, SDXL hits the sweet spot — high quality, large enough
ecosystem, manageable VRAM, fast enough iteration.
Flux: when only the best will do
Flux dev (the open-weights variant; pro is API-only) clearly leads on:
- Hands and anatomy — the long-running embarrassment of diffusion models is much better.
- Text rendering — Flux can produce legible text in images far more reliably than SD or SDXL.
- Prompt adherence — Flux follows complex compositions and multi-subject prompts more reliably.
- Fine detail at native resolution — sharper, less smudgy.
The cost is real:
- VRAM: Flux dev FP16 is 24 GB. You need a 4090 / 5090 / A6000 / 6000 Ada to run it comfortably. On 16 GB cards, fp8 quantization is required and quality takes a small hit.
- Speed: Flux dev needs 20–28 sampling steps for best quality, taking 15–40 seconds per image even on top hardware. Flux schnell (the distilled fast variant) generates in 1–4 steps but with a quality drop.
- Fine-tuning is harder. The transformer architecture trains differently than UNets; LoRA quality is improving but the ecosystem is younger.
Flux is the right choice when image quality is the goal and hardware can keep up.
Fine-tuning ecosystem comparison
| Aspect | SD 1.5 | SDXL | Flux |
|---|---|---|---|
| LoRA count on Civitai | 100,000+ | 30,000+ | growing fast (5,000+) |
| Custom checkpoints | thousands | hundreds | dozens |
| ControlNet support | universal | universal | partial (some types) |
| LoRA training VRAM (rank 32, 1024 res) | ~10 GB | ~16 GB | ~24 GB |
| IP-Adapter support | universal | universal | partial |
| Inpainting models | strong | strong | improving |
The trend: SDXL's ecosystem reached parity with SD 1.5 in 2024–2025; Flux is approaching
parity now in 2026 but still trails on niche or stylistic fine-tunes.
Which one for which workflow
Goal: maximum visual quality, you have 24 GB+ VRAM
→ Flux dev (fp16 if 24 GB, full fp16 if 32+ GB)
Goal: best photoreal photos, 12–16 GB VRAM
→ SDXL with a custom checkpoint like JuggernautXL or RealVisXL
Goal: anime, illustration, artistic styles, character LoRAs
→ SD 1.5 with the right LoRA stack
(or SDXL if you want higher resolution out of the gate)
Goal: speed-iterate during prompt design
→ SDXL Turbo or Flux schnell — both designed for 1–4 step generation
Goal: legible text in images (logos, signs, posters)
→ Flux is the only realistic option
Goal: Limited VRAM (6–8 GB)
→ SD 1.5 (or SDXL with --medvram flag in A1111 / sequential offload in ComfyUI)
For more practical setup help, see our ComfyUI on Windows guide —
all three model families load the same way.
What about SD3, SD3.5, the newer models?
Stability AI released SD3 and SD3.5 in late 2024 / early 2025. They are technically
competitive with Flux at a similar parameter scale, but adoption has been slower because of
their commercial license terms (free for non-commercial; paid for commercial use beyond a
revenue threshold). The Flux dev license is more permissive (non-commercial) and Flux schnell
is fully Apache 2.0 — both reasons the open-source community gravitated to Flux as the SD3
successor in practice.
If you are building a commercial product, the licensing differences matter. For personal use,
either family works.
Bottom line for your wallet
If you are buying a card specifically for image generation in 2026:
- Budget pick (under $400): RTX 4060 Ti 16GB or RTX 5060 Ti 16GB — the VRAM is the thing. 16 GB lets you run SDXL comfortably and Flux dev with fp8 quantization. Either is vastly better than the 8 GB options at similar prices.
- Sweet spot ($800–1200): [RTX 4080 Super](https://www.amazon.com/s?k=RT
Top comments (0)