SD 1.5 vs SD 2.1 vs SDXL: Differences and Which to Use on iPhone

SD 1.5 vs SDXL comes down to size: SDXL is the newer, larger model and generally makes better images out of the box, while SD 1.5 is smaller, faster and lighter on memory. SD 1.5 was trained at 512x512 with one CLIP ViT-L text encoder, while SDXL trains at about 1024x1024, adds a second, larger OpenCLIP ViT-bigG text encoder and uses a UNet about three times bigger (2.6B vs 860M parameters, per the SDXL paper). Stability AI reports SDXL well ahead of SD 1.5 and 2.1 in user preference, but the extra size means every SDXL image costs more time and memory. On iPhone the gap grows: in Apple’s Core ML benchmarks the iOS SDXL build runs at 768x768 and takes 31 s on an iPhone 15 Pro Max and 77 s on an iPhone 14 Pro Max, against 7.9 s for SD 2.1 Base on the same iPhone 14 Pro Max (the SDXL run there loads models one at a time to fit in memory, which adds much of that time); SD 2.1 Base shares SD 1.5’s 512x512 size and similar-size UNet. PhoneDiffusion’s own SDXL packs are fast 4 to 8 step builds, so these benchmark times are not app speeds.

SD 1.5 vs SD 2.1 vs SDXL at a glance


SD 1.5

SD 2.1 Base

SDXL Base 1.0

Native resolution

512x512

512x512

About 1024x1024 (Apple’s iOS build: 768x768)

Text encoder(s)

CLIP ViT-L/14

OpenCLIP ViT-H

CLIP ViT-L plus OpenCLIP ViT-bigG

Relative size (UNet)

860M parameters

865M parameters, same class as SD 1.5

2.6B parameters, about 3x larger

Speed on phone (Apple benchmark, 20 steps, 2023)

Not in Apple’s table; expected close to SD 2.1 Base (similar-size UNet, same 512x512)

7.9 s on iPhone 14 Pro Max, 10.8 s on iPhone 13 (reduced-memory mode)

31 s on iPhone 15 Pro Max, 77 s on iPhone 14 Pro Max (reduced-memory mode)

Fine-tune ecosystem (Hugging Face, Nov 2026)

Large: about 1,000 fine-tunes and 3,500 adapters (old runwayml and current repos combined)

Small: 58 fine-tunes and about 700 adapters

Largest: about 1,200 fine-tunes and 9,800 adapters

Best for

Fast drafts and long prompt sessions on older iPhones

Trying a different text encoder at SD 1.5 cost

Best image quality on newer iPhones, when you can wait

Sources: SD 1.5 model card, SD 2.1 Base model card (Apple’s Core ML version; Stability’s original repo is no longer public), SDXL Base 1.0 model card, SDXL paper (Table 1), Apple ml-stable-diffusion benchmarks and Hugging Face fine-tune listings for SD 1.5 (plus the old runwayml repo), SD 2.1 Base and SDXL, checked November 16, 2026.

In PhoneDiffusion, SD 1.5 and SD 2.1 Base packs generate at 512x512 and SDXL packs at 768x768. All packs are curated, so you cannot import your own checkpoints or LoRAs.

If you want to run Stable Diffusion directly on your iPhone and generate AI images offline, check out PhoneDiffusion.

The short answer

  • Choose SD 1.5 if speed, a deep, long-running style ecosystem, and quicker local iteration matter most.

  • Choose SD 2.1 Base if you want an alternative to SD 1.5 at the same size and speed, with a different text encoder and far fewer fine-tunes.

  • Choose SDXL if you are on stronger iPhone hardware and care more about output ambition than fast iteration.

  • Do not assume the best desktop answer is automatically the best iPhone answer.

Why model choice feels different on iPhone than on desktop

On desktop, users can often brute-force around a heavier model with more VRAM, bigger cooling headroom, and a stack of supporting tools. On iPhone, model family changes much more than image taste. It changes setup size, generation time, heat, device fit, and how patient the workflow feels.

Apple’s public Core ML releases make that split visible. Its iPhone-ready Stable Diffusion v1.5 and 2.1 Base builds are 512x512-class models, while the iPhone-oriented SDXL release is a separate 768x768 build with more aggressive iOS-focused compression. That alone tells you something important: on iPhone, model choice is not just a quality debate. It is a product and workflow decision.

SD 1.5: still the easiest way to keep a local workflow nimble

Apple publishes Core ML v1.5 builds for Apple Silicon, including a 6-bit palettized version, and the original v1.5 model card is still rooted in 512x512 training. That smaller starting point is one reason SD 1.5 still matters for mobile-minded users.

The bigger reason is ecosystem. Outside Apple-specific deployment, SD 1.5 has the oldest backlog of checkpoints and LoRAs, much of it on Civitai, though SDXL now has more fine-tunes and adapters listed on Hugging Face. If you care about anime styles, older fine-tunes, or fast prompt exploration, that matters more than the base model’s age. If LoRAs are new to you, start with What Is a LoRA? A Beginner’s Guide for Stable Diffusion Users.

In PhoneDiffusion you choose from curated SD 1.5 packs (such as DreamShaper 8, Realistic Vision 5.1 and Anything V5).

On iPhone, SD 1.5 makes the most sense when your goal is:

  • fast drafting and iteration

  • broader style experimentation

  • a lighter local workflow

  • less patience for long per-image wait times

The tradeoff is also clear. The official model card still flags weak legible text, weaker compositional tasks, and imperfect faces. So SD 1.5 is still a smart iPhone choice, but not because it has the highest ceiling. It is useful because it is the easiest family to keep flexible.

SD 2.1 Base: an alternative to SD 1.5

Stable Diffusion 2.1 Base rarely gets the same attention as SD 1.5 or SDXL, but it matters more on iPhone than it does in generic desktop debates.

Apple’s public Core ML benchmark table uses Stable Diffusion 2.1 Base at 512x512 as a reference model for iPhone-class hardware, and Apple also publishes ready-made Core ML 2.1 Base weights for Apple Silicon. The official 2.1 Base model card still warns about familiar limitations: poor legible text, imperfect people, and harder compositional prompts. It is not a magic answer.

SD 2.1 Base runs at the same 512x512 size and about the same speed as SD 1.5. It uses a different text encoder (OpenCLIP), so prompts written for 1.5 can behave differently, and far fewer fine-tunes exist for it. Treat it as an alternative to SD 1.5, not a middle step toward SDXL.

On iPhone, SD 2.1 Base makes the most sense when you want:

  • a practical local baseline

  • something lighter than SDXL

  • a different text encoder (OpenCLIP) than SD 1.5

If SD 1.5 is the fast-and-flexible option and SDXL is the quality-first option, SD 2.1 Base sits next to SD 1.5, not between SD 1.5 and SDXL.

SDXL: best when output ambition matters more than friction

SDXL earns its reputation. The SDXL model card describes a larger system with two text encoders, and Stability AI’s published evaluation notes show SDXL outperforming Stable Diffusion 1.5 and 2.1 in user preference.

But that higher ceiling comes with a different mobile profile.

Apple’s iPhone-oriented SDXL release is not the same thing as running desktop SDXL unchanged. The public iOS build uses a 768x768 target, 4.04-bit mixed-bit palettization, and an iOS-specific Core ML packaging path to make the model viable on Apple devices at all.

Apple’s published benchmark table makes the workflow tradeoff concrete. On iPhone 14 Pro Max, Apple measured 7.9 s for SD 2.1 Base at 512x512 and 77 s for SDXL at 768x768. The SDXL run had to load models one at a time to fit in memory. On iPhone 15 Pro Max, SDXL took 31 s. All runs used 20 steps. Those are not universal promises, but they are a useful public reference for how large the mobile gap can be across model families.

On iPhone, SDXL makes the most sense when you care about:

  • better composition and higher output ambition

  • higher-resolution local output

  • fewer but more deliberate generations

  • newer hardware that can justify the extra weight

The downside is exactly what users feel in practice: bigger setup cost, slower iteration, more heat, and less tolerance for a casual "try 40 prompts in a row" workflow.

Which model family fits which iPhone workflow?

Use SD 1.5 when you want the fastest path to local experimentation and a deep, long-running style ecosystem.

Use SD 2.1 Base when you want to try a different text encoder at about SD 1.5 cost.

Use SDXL when you care more about output ambition than iteration speed and you are comfortable with a heavier local workflow.

Many serious local users end up thinking in exactly these terms: smaller model families for faster exploration, heavier ones for more deliberate output. On iPhone, that pattern matters even more because the cost of a heavier model is impossible to ignore.

How PhoneDiffusion fits this topic

PhoneDiffusion offers SD 1.5 and SDXL packs, both base models and fine-tunes, plus an SD 2.1 Base pack. Some packs are distilled (Lightning, Hyper, LCM) and finish in 4 to 8 steps, so an SDXL Lightning pack can be much faster than a standard SDXL model at 20 steps. SDXL packs only appear on devices with enough memory.

It also offers optional cloud models, such as FLUX and Nano Banana, on paid monthly or yearly plans. Cloud is text-to-image only and sends your prompt online. There is no user account and no cloud sync of the gallery.

PhoneDiffusion does not pretend one model is universally best. The tradeoff is clear:

  • smaller and faster is a real user need

  • fast distilled packs change the speed math

  • higher-quality local generation is valuable, but it is not free

That is a better framing for technical users and prosumers than generic AI-app marketing.


If you want to run Stable Diffusion directly on your iPhone and generate AI images offline, check out PhoneDiffusion.


FAQ

Is SDXL always better than SD 1.5 on iPhone?

No. SDXL has the higher ceiling, but on iPhone the cost in setup, time, and heat is also much higher. If fast iteration matters more than maximum output ambition, SD 1.5 can still be the better fit.

Why does SD 1.5 still matter in 2026?

Because local workflows still care about speed, control, and ecosystem depth, not just headline quality. A smaller model family is often the most practical way to keep local generation usable on phone-class hardware.

Is SD 2.1 Base obsolete?

No, but it is not a middle step either. It runs at the same 512x512 size and about the same speed as SD 1.5, with a different text encoder and far fewer fine-tunes. Treat it as an alternative to SD 1.5, not a step toward SDXL.

Can one iPhone app support multiple model families well?

Yes, but users should not expect every phone to feel the same with every model family. Public Apple benchmarks and the current iPhone app landscape both make that clear.

Final takeaway

The best Stable Diffusion model on iPhone is not a universal best.

SD 1.5 remains the quickest way to keep a local workflow light. SD 2.1 Base is an alternative to SD 1.5 at about the same cost. SDXL is what you reach for when you want a higher visual ceiling and have hardware that can carry it.

That is why PhoneDiffusion offers more than one model family. The right iPhone image-generation product does not flatten model choice into a slogan. It helps users choose the local workflow that actually fits them.