1.  [Private Diffusion](/en)
2.  /
3.  Draw Things Alternative

# Looking for a Draw Things Alternative?

Draw Things is the best-known on-device diffusion app on iPhone and iPad, and it earned that reputation honestly: 233 models, seven quantizations of Krea 2 alone, an open-source engine a hobbyist can read line by line. If you want that much control, keep it. Private Diffusion is for the rest of the job: you want a frontier image model on your own device, and you would rather not choose a quantization before you choose a prompt.

## Why People Look for a Draw Things Alternative

Draw Things puts the decision in front of you every time: which of seven Krea 2 quantizations, whether that build even touches the Neural Engine on your device, which config to load. Private Diffusion made that decision once, for every model in the roster, and trained the weights so the quality holds. One download. No quantization menu.

## Private Diffusion vs Draw Things

A model-by-model claim needs its own citation, so here is the comparison at the level we can back today.

| Dimension | Private Diffusion | Draw Things |
| --- | --- | --- |
| Quantization | Trained (QAD) at INT4/INT6 against a full-precision teacher | Cast and palettized after training, no data or loss |
| Model choice | One curated build per model, ten total | Pick among up to seven quantizations per model, 233 in the catalog |
| Krea 2 Turbo on a 16 GB iPad | Resident in unified memory | Streams from SSD mid-inference on any device under 23 GiB |
| Neural Engine | Requested by design for encode, preview, decode, and upscale | Engages only if you pick the i8x variant on an M3 or M4 |
| Gallery lock | Optional Face ID or Touch ID on Hidden | None in the open-source app |
| Setup | One download, zero configuration | Config files, samplers, checkpoint management |

Sources and full receipts for every row above are tracked internally and re-checked quarterly. Quantified speed claims stay off this table until we publish a same-device benchmark.

## Trained Weights, Not Cast Weights

Draw Things' quantizer is data-free: it reads the full-precision tensors and casts them down, k-means palettization for its q4p through q8p formats, an iterative least-squares pass for the newer i8x format. No training, no loss function, no forward pass. That is real engineering, not a shortcut. Draw Things' own developer has said quantization-aware training is "the only way to recover performance" at extreme compression, and that he has yet to try it.

We already have. Every Private Diffusion model ships as quantization-aware distillation: we quantize the weights, then train the quantized model against its own full-precision teacher so the compressed version is built to match it, not approximate it after the fact. The target is INT4 with grouping of 32 (INT6 on Anima and Flux.2 Klein 4B). That is the mechanism behind "Quantization without compromises": we train the weights to hold the same quality Draw Things' cast weights can only approach.

## The Neural Engine, On by Design

Every Private Diffusion model requests the Neural Engine for encode, live preview, final decode, and the latent upscalers, chosen by profiling, not gated behind a setting you have to find. In Draw Things the Neural Engine only engages if you happen to pick the i8x quantization and own an M3 or M4; the default Krea 2 download is the 8-bit format, which never touches it. Both apps use the same chip. One asks you to find the right switch first.

## One Decision Instead of a Quantization Quiz

Draw Things offers seven quantizations of Krea 2 alone, and picking the wrong one means an M4 iPad never touches the Neural Engine. Private Diffusion is one download. We made that decision for every model, then trained the weights so the quality holds. Ten curated models against Draw Things' 233 is a real trade: range against one answer that already works.

## Frequently Asked Questions

Is Private Diffusion a real alternative to Draw Things?▼

For someone who wants a picture, yes. Draw Things is the more capable app: 233 models, full quantization control, an open-source engine you can extend. Private Diffusion is ten curated models, one download each, tuned so the Neural Engine and unified memory get used without you configuring anything. Concede the difference up front, then decide which one matches what you actually want to do.

Does Draw Things run Krea 2 on a 16 GB iPad?▼

Yes. Draw Things self-hosts seven Krea 2 variants and its own developer has confirmed Krea-class models run on a 16 GB iPad, just slower. Private Diffusion runs Krea 2 Turbo resident in unified memory instead of streaming roughly 15.5 GB off SSD mid-generation, because our build is trained at INT4, half the bytes of Draw Things' 8-bit default. Same hardware tier, different path to get there.

What's the difference between Private Diffusion's quantization and Draw Things'?▼

Draw Things reads the full-precision weights and casts them down: k-means palettization or iterative-scale int8, no training pass. That is a legitimate compression technique, not the naive rounding some critics call it. Private Diffusion trains the quantized weights against their own full-precision teacher first (quantization-aware distillation, INT4 with INT6 on Anima and Flux.2 Klein 4B), so the compressed model is built to hold the teacher's quality rather than approximate it after the fact.

Is Private Diffusion faster than Draw Things?▼

We believe so, on the models we ship: our trained INT4 weights move half the bytes per denoising step of Draw Things' 8-bit default Krea 2, and they stay resident in memory instead of streaming from storage mid-run. We have not published a timed benchmark on identical hardware yet, so treat 'faster' as the architecture argument it is until we do.

Does Private Diffusion require an account, like cloud image generators?▼

No. Private Diffusion runs entirely on your Apple Silicon device. No account, no upload, no server between you and the image. Free in the TestFlight beta.

## Try Private Diffusion Free in the Beta

Ten on-device image models, one download each, trained so the quality holds. Krea 2 Turbo needs a 16 GB iPad or Mac; Anima and Flux.2 Klein 4B run on any supported iPhone.

[Join the Beta](/en#beta)

[See all ten models →](/models)