Draw Things @drawthingsapp Twitter profile

Pinned Tweet

Draw Things

3 months

1.20240306.2 just released in iOS / macOS AppStore (). In this release, we added Würstchen v3.0 (Stable Cascade) support, added necessary sampler changes to work with Playground v2.5, and updated to moondream2 as image captioner. 1/6

5

29

97

Last Seen Profiles

@cale_brand86222

@CricketScotland

@eternal_garbage

@KOKE0u0

@Maria_Garland_

@_On1k

@jainoynoy2

@Asanda_mkhize_

@mothdayo

@jandakembangstw

@Nanamagari_Plus

@The_Old_Sarge

@NewChocoFactory

@Davinci23638919

@AHSLadyTrojanBB

@Davinci23638919

@SirGary_y

@ondaceromelilla

@kiwi_fartist

@jandakembangstw

@SheepEsports_FR

@softwarecooliee

@differland_

@Rodge_This

@Jacob_Rochester

@stw_pdg

@markclaflin

@daocollab

@Tigress_X

@BenGultig

@Muppetiers

@miguelcobochef

@sooob_bean

@_cirunei

@Wolfonabridge

Draw Things

@drawthingsapp

10 months

You can download in AppStore now for SDXL! 1.20230726.1 just released with support for @StabilityAI 's latest image generation model. It runs on all devices, but runs well on iPad Pro with 8GiB RAM or MacBooks. Stay tuned for quantized model that runs well on all devices! (1/9)

14

23

136

Draw Things

@drawthingsapp

6 months

Here is how to use SDXL Turbo within Draw Things today: 1. Download the model from HuggingFace; 2. Import through our import flow; 3. Edit "custom.json" file (under ~/Library/Containers/Draw Things/Documents/Models) with the following change; 4. Enjoy HQ 4 step generation!

Stability AI

@StabilityAI

6 months

Introducing SDXL Turbo: A real-time text-to-image generation model. SDXL Turbo achieves state-of-the-art performance with a new distillation technology, enabling single-step image generation with unprecedented quality, reducing the required step count from 50 to just one. The

131

596

3K

15

14

134

Draw Things

@drawthingsapp

7 months

153 hours ago, @SimianLuo released LCM LoRAs and LCM SDXL. We are late to the party, 1.20231115.0 released in iOS / macOS store today () added LCM-family models / LoRAs support, which marks a significant speed up when running image generations locally:

6

24

111

Draw Things

@drawthingsapp

8 months

1.20231004.1 was released in macOS / iOS AppStore (). This version brings on-device model fine-tuning (in the form of LoRA training) to people with iPhone 15 Pro / Pro Max, iPad and Mac! (1/6)

8

14

95

Draw Things

@drawthingsapp

4 months

1.20240205.0 was released in iOS, macOS AppStore two days ago. It packs so many new features that I am procrastinating to introduce them! Let's unpack: (1/6)

6

5

69

Draw Things

@drawthingsapp

1 year

Took longer than expected. But we are getting there!

9

11

67

Draw Things

@drawthingsapp

2 months

1.20240330.0 was just released in iOS / macOS AppStore (). In this release, we introduced "Community Models", which is a list of community maintained models / LoRAs that you can directly download into your devices.

4

21

68

Draw Things

@drawthingsapp

6 months

1.20231210.2 released in AppStore (). We heard the community feedback and made the model selectors more compact. We also added Stable Video Diffusion model support in this release.

7

12

59

Draw Things

@drawthingsapp

1 year

1.20221208.1 is out! The biggest feature in this release is to bring your own custom model to the app! Yes, that's right. You can use variety of DreamBooth model training providers, either paid or free @huggingface Space, and bring the trained model into the app. Here is how:

3

5

58

Draw Things

@drawthingsapp

5 months

New year! 1.20240110.1 () is an overhaul of our interface on macOS and iPad now available in AppStores.

8

4

58

Draw Things

@drawthingsapp

1 year

Took a while. But hey, it is free!

5

8

55

Draw Things

@drawthingsapp

2 months

1.20240414.1 released in AppStore a few hours ago (). In the past a few versions, we enhanced the app with the ability to generate transparent images, ultra-high resolution images (8192x8192) and more. Let's dive in! 1/6

1

17

56

Draw Things

@drawthingsapp

7 months

1.20231018.0 just dropped in iOS / macOS AppStore! This version fixed LoRA training issues that "overcook" the loss. It should make your training process much smoother. We also added a "small" language model (BLIP2 + OPT 2.7B) to help caption training images easier. (1/2)

9

4

50

Draw Things

@drawthingsapp

10 months

To make this work, we integrated MFA GEMM kernel from @philipturnerar of project. The result is instant: it improves the overall performance on SD v1 / v2 / XL by about 15%. And this is just a beginning! (4/9)

GitHub - philipturner/metal-flash-attention: Faster alternative to Metal Performance Shaders

Faster alternative to Metal Performance Shaders. Contribute to philipturner/metal-flash-attention development by creating an account on GitHub.

github.com

1

48

Draw Things

@drawthingsapp

2 months

1.20240321.2 was released in AppStore () a few moments ago. In this and preceding versions, we improved usability on a wide-range of devices. Let's dive in! 1/6

1

8

48

Draw Things

@drawthingsapp

6 months

And you can import this model directly into @drawthingsapp today.

Suhail

@Suhail

6 months

1/ Introducing Playground v2: A new commercially open model from our team that we trained from scratch. Most notably our model was preferred 2.5x more than the current leading open model (SDXL). We're excited to give back to the community as we're just getting started.

81

213

1K

3

1

47

Draw Things

@drawthingsapp

1 year

1.20221222.0 is out! In this version, we added the ability to paste images into the app, some fixed presets to simplify common tasks, and depth2img support! Using iPhone-captured depth map, img2img generation can preserve it's structure, and it is awesome!

7

45

Draw Things

@drawthingsapp

11 months

8

10

46

Draw Things

@drawthingsapp

1 year

1.2023065.1 is here! (). This version introduced Infinite Canvas. A helper to solve many UI problems we have when working on iterative image generation tasks.

14

4

45

Draw Things

@drawthingsapp

9 months

1.20230913.1 was released in iOS / macOS AppStore (). This version enables just-in-time weight decompression such that our 8-bit models will enjoy lower RAM usage. This is particularly important for large models on iPhone / 8GiB MacBooks (1/5)

5

42

Draw Things

@drawthingsapp

5 months

1.20240114.1 was released in iOS / macOS AppStore (). This release brought the same UI update we had on macOS / iPad to iPhone users. Hopefully, it is more organized, and easier to navigate.

3

9

41

Draw Things

@drawthingsapp

1 year

1.20230422.0 was released to iOS / macOS AppStore last week. Last week's release added ControlNet 1.1 family, including LineArt, Scribble, Canny, Pose, Depth, Inpainting, Shuffle and Instruct Pix2Pix. We also added the long waited "Face Restoration" as post-processing step.

6

9

38

Draw Things

@drawthingsapp

4 months

1.20240212.2 was released in iOS / macOS AppStore a few hours ago. In this version, we brought you a new image captioning model, ControlNet LoRA support, higher resolution for depth map generation and more! Let's get started: 1/6

3

7

40

Draw Things

@drawthingsapp

1 year

Textual inversion is here! 1.20230130.0 just published on iOS and macOS AppStore a few minutes ago and the headline feature is the textual inversion support you asked for:

12

5

39

Draw Things

@drawthingsapp

2 years

We are alive!

9

13

39

Draw Things

@drawthingsapp

10 months

1.20230807.0 is available in iOS / macOS AppStore (). This version fully integrated Metal FlashAttention. We saw 43% to 120% faster image generation times, while keeping peak RAM usage lower than before (you can generate higher resolution images).

4

7

37

Draw Things

@drawthingsapp

4 months

1.20240118.2 was released in iOS / macOS AppStore (). This release fixed a few bugs related to paint color selection, model deletion and LoRA usage with 8-bit models. A new inpaint model / LoRA for SDXL is added, which IMHO is the best trained one for SDXL.

2

8

34

Draw Things

@drawthingsapp

1 year

1.20230428.1 was released in iOS / macOS AppStore ().

4

6

30

Draw Things

@drawthingsapp

9 months

1.20230826.0 is released in iOS / macOS AppStore (). We are now on the streak to release with even more pro features! This release introduced "Scripting" capability, allowing you to write JavaScript to manipulate the app: (1/4)

5

4

31

Draw Things

@drawthingsapp

1 year

1.20230318.0 now is dropped in iOS / macOS AppStore (). This version added long-waited LoRA models support. You can use multiple LoRAs, importing new LoRA models. You can even enable CoreML with LoRA to speed up the generation.

5

4

31

Draw Things

@drawthingsapp

7 months

1.20231025.0 was dropped in AppStore (). It includes IP-Adapter introduced by that adds some very interesting capabilities on top of image generation models. (1/8)

2

5

30

Draw Things

@drawthingsapp

1 year

1.20230114.2 is out! This is an "in-between" version with some bugfixes and the optional CoreML based rendering. You don't need to download new models to enable CoreML and any SD v1.x custom models can benefit from it. But there are limitations so it is not the default yet:

5

2

24

Draw Things

@drawthingsapp

4 months

Moondream1 is a tiny VLM trained by @vikhyatk . It is pretty suitable to run on variety of devices. We brought it inside our app to support better image captioning. It generates pretty detailed stories about the image. It is a lot of fun! 2/6

1

25

Draw Things

@drawthingsapp

10 months

It is versatile enough out of the box to be the next base model if people steer it well! (/END)

0

2

22

Draw Things

@drawthingsapp

7 months

1.20231031.0 is released in macOS / iOS AppStore. This version beefed up our T.I. Embedding (Textual Inversion) support: you can import SDXL T.I. embedding, train T.I. embedding along with LoRA, and export your T.I. embeddings. (1/4)

2

4

22

Draw Things

@drawthingsapp

10 months

There are some other engineering tricks we implemented to make SDXL work as well as it should on Apple devices. Another shout-out goes to to @madebyollin for his SDXL VAE FP16 fix: (6/9)

madebyollin/sdxl-vae-fp16-fix · Hugging Face

huggingface.co

1

0

21

Draw Things

@drawthingsapp

1 year

On security: there are discussions online about how "insecure" these checkpoint files are. No need to worry! For this feature, we implemented PickleVM in Swift, that means no Python code executed during model conversion. Read more about how this is done:

GitHub - liuliu/swift-fickling

Contribute to liuliu/swift-fickling development by creating an account on GitHub.

github.com

0

21

Draw Things

@drawthingsapp

1 year

1.20230308.1 is dropped in iOS / macOS AppStore (or download at ). This and previously 1.20230305.0 added "Scribble" and "Pose" support to Control section.

8

5

19

Draw Things

@drawthingsapp

2 months

As part of this release, we also made part of Draw Things app open-source: It is a rough start, but hopefully overtime it is much easier to run this repo for everyone. Hope you enjoy and looking forward to Pull Requests!

GitHub - drawthingsai/draw-things-community: The community repository for the Draw Things app.

The community repository for the Draw Things app. Contribute to drawthingsai/draw-things-community development by creating an account on GitHub.

github.com

1

3

21

Draw Things

@drawthingsapp

1 year

People asked, how can I apply a certain style to an image with your app? Inpainting model can actually do that magically! (And will be even better for this in the future). In 1.20221231.2, we did a little tweak to allow just this!

1

3

21

Draw Things

@drawthingsapp

7 months

On iPhone 15 Pro, a typical generation (512x512) of LCM LoRAs with SD v1 models takes less than 10s. A 1024x1024 high quality generation with SDXL LCM LoRAs take less than 40s. Your familiar tools such as ControlNet, IP-Adapter all works. (2/6)

2

3

20

Draw Things

@drawthingsapp

1 year

1.20230225.0 is released in iOS / macOS AppStore (or download at ). This release brings preview of ControlNet support.

3

5

20

Draw Things

@drawthingsapp

1 year

1.20230406.0 was released to iOS / macOS AppStore last week. It supports multi-ControlNet, added T2I adapters, and added UniPC sampler. Let's take a look at these:

5

20

Draw Things

@drawthingsapp

10 months

As always, the source code to run SDXL is actually up for a while in the swift-diffusion repo: . You might ask, what's special of SDXL? Here is our opinion: with better layout of attention blocks, it is more compute-friendly and have more potential. (7/9)

GitHub - liuliu/swift-diffusion

Contribute to liuliu/swift-diffusion development by creating an account on GitHub.

github.com

1

20

Draw Things

@drawthingsapp

1 year

1.20221220.1 is out! Besides better support for model imports and more models, the headline feature in this release is the "upscaler" support. An upscaler helps you to increase the finished image resolution dramatically with minimal artifacts or artistic altering.

2

3

20

Draw Things

@drawthingsapp

7 months

If you experience crashes after upgrading to iOS 17.1, please cleanup the temporary directory to alleviate this. We will do a fix later next week to automate this process after OS upgrade. It seems MPS opcode is not stable from OS version to version.

2

1

19

Draw Things

@drawthingsapp

1 year

Model mixing with inpainting model is very versatile! Remember the "style prompt + image" trick to do style transfer I shared a few days ago? It can be supercharged with model mixing! Here is how:

1

19

Draw Things

@drawthingsapp

6 months

How much RAM you use when running Stable Video Diffusion?

1

0

18

Draw Things

@drawthingsapp

10 months

One question people have: how fast is SDXL on device? On M1 Pro, you can expect around 150s end-to-end for 30-step generation on 1024x1024. It is about 90s on M1 Max for the same generation. It is roughly the same speed at the same resolution comparing to SD v1. (2/9)

1

17

Draw Things

@drawthingsapp

1 year

Friday quick photo tip: removing subject with ControlNet inpainting.

2

4

17

Draw Things

@drawthingsapp

2 years

1.20221130.0 is out. We supported iPad and #stablediffusion2 in the version 3 days ago. So this supposed to be a smaller version. But we cannot resist ... First, now images saved to Photos will have prompt as part of the metadata. (1/4)

3

2

16

Draw Things

@drawthingsapp

1 year

1.20230328.0 is released in iOS / macOS AppStore (). Headline: there is now experimental support for Intel Macs. While slow, now we can claim availability on all Apple platforms. Second, model managent interface:

4

6

17

Draw Things

@drawthingsapp

7 months

Hope you enjoy these faster generations! There are a lot more work to do! ✌️(6/6)

1

0

17

Draw Things

@drawthingsapp

1 year

1.20221229.1 is out! This bug-fix version fixed the biggest complaint we got: it doesn't work on my [iPhone X | Xr | iPad 9th Gen]! Now it works on everything since iPhone X, and up to 512x768 for 4GiB devices ("Device Capability Warning" dialog is removed once and for all).

2

1

15

Draw Things

@drawthingsapp

1 year

1.20230108.0 on iOS (1.20230108.2 soon on macOS or ) is out! Besides fixing some problems for iOS 15.x users, this version introduced two new features: model mixing (or model merging) and project management.

4

0

16

Draw Things

@drawthingsapp

10 months

On iPad Pro M2, it reduces SD v1 at 512x512 per iteration UNet evaluation from 0.9s to 0.75s. On iPhone 14 Pro, it reduces from 1.88s to 1.6s. For SDXL at 1024x1024 on iPad Pro M2, it reduces from ~8s to ~6s and generally have better thermal curve. (5/9)

1

0

16

Draw Things

@drawthingsapp

1 year

One last thing: paste image works really well with Photos Cutout introduced in iOS 16! Select the object in Photos, copy, long press Camera button to paste in @drawthingsapp , now you can fill in the background with a prompt! (END)

2

1

16

Draw Things

@drawthingsapp

7 months

This took a while, but I think we are here with 1.20231115.0.

Emad

@EMostaque

1 year

Delighted to have native support for the AI neural engines for Stable Diffusion from @Apple , one of the 1st optimised models. 8s on MacBook Air M2, will be < 1s with distilled #StableDiffusion2 AI for all. Can't wait to see what everyone creates. 🚀✨

16

56

358

2

0

16

Draw Things

@drawthingsapp

2 years

Something something 2.0 …

3

16

Draw Things

@drawthingsapp

5 months

We organized the UI such that it is more scalable for things yet to come. It should be easier to navigate around to familiarize with all different functionalities. We also added a project navigator to help you get familiar with project-based workflow.

4

0

16

Draw Things

@drawthingsapp

1 year

ControlNet Tile is available now too! It can fill much more detail when zoom-in, run high-denosing-strength img2img with it. See for yourself!

2

3

16

Draw Things

@drawthingsapp

4 months

There are also some more UI refreshments, hope you liked it! /END

1

0

15

Draw Things

@drawthingsapp

1 year

If all you have is a hammer ...

1

3

15

Draw Things

@drawthingsapp

1 year

"make it snow" (thx Tim Brooks & A leksander Holynski!)

3

14

Draw Things

@drawthingsapp

1 year

1.20230212.0 was release in iOS and macOS AppStore this morning (direct download link: ). In this release, we fixed some interoperability issues and expanded CoreML support.

5

2

14

Draw Things

@drawthingsapp

2 years

Cooking

4

1

15

Draw Things

@drawthingsapp

4 months

Oh! This one also updated with Stable Video Diffusion XT v1.1 model!

4

0

15

Draw Things

@drawthingsapp

10 months

The two-stage approach gives software more freedom to choose where to fine-tune what, and the larger specialized text encoder has better text understanding without losing visual touch (a common error by just use off-the-shelf text encoder). (8/9)

1

15

Draw Things

@drawthingsapp

11 months

1.20230630.0 is released ()! This version introduced a 8-bit model mode. You can now also convert your models into 8-bit to save space. It is around 800MiB but performs close to the original. Let's see if you can guess it right:

1

2

15

Draw Things

@drawthingsapp

8 months

You can read more at . On-device fine-tuning is a great way to bring your own photos and styles to the model. You can fine-tune with as little as 10 minutes to an hour to see the impact on the model. Both SD and SDXL models can be fine-tune d! (2/6)

Draw Things Democratizes Local Large Model Fine-tuning on iPhone, iPad and Mac

Large language models and image generation models currently require hundreds of thousands to tens of millions of dollars for hardware…

engineering.drawthings.ai

1

2

14

Draw Things

@drawthingsapp

1 year

1.20230203.0 just dropped in iOS and macOS (macOS direct download: ). It is a bug-fix version for various textual inversion importing problems you have. It is also a version that starts to support deletion from image history. Just long press!

2

0

13

Draw Things

@drawthingsapp

3 months

The moondream2 captioner ( @vikhyatk ) was released on Monday. It hallucinates way less than moondream1 models. We also did some performance work to have text generation performance on lower-end devices on-par with llama.cpp. 5/6

1

14

Draw Things

@drawthingsapp

9 months

One last thing: we added multiple image deletion option in history view. You can simply do "Select Multiple ..." And delete selections all at once. /END

2

1

14

Draw Things

@drawthingsapp

1 year

How to fix a bad generation?

2

3

14

Draw Things

@drawthingsapp

1 year

Last but not the least! A convenient menu item added that allow you paste Generate Data directly into the text prompt. It will also configure various settings, and do automatic line break for long prompt, which matches A1111 behavior better. A handy little helper!

1

14

Draw Things

@drawthingsapp

1 year

1.20230206.1 is dropped in iOS / macOS AppStore (download: ). The headline is the new "Seed Mode":

2

3

14

Draw Things

@drawthingsapp

6 months

It doesn't have frame control, extending / inserting new frames, or any fancy video editing tools yet, but it does generate pretty clips! More over, you can animate only parts of the image like this:

1

14

Draw Things

@drawthingsapp

1 year

Oh, and of course it can do QR code. Because why not:

1

0

13

Draw Things

@drawthingsapp

1 year

Hello from iPhone Xr

1

0

13

Draw Things

@drawthingsapp

4 months

You can use LCM SDXL Refiner LoRA in conjunction with LCM SDXL Base LoRA like this in the app. 20s for high quality 1024x1024 on an iPhone is not bad at all! There are more to come! Stay tuned! /END

5

0

13

Draw Things

@drawthingsapp

1 year

To enable CoreML with LoRA, you need to select both "Use CoreML If Possible" and "Use CoreML for LoRA". Because each change for LoRA settings (different weights, different combinations) will generate a new CoreML model, it is not on by default. /END

3

1

12

Draw Things

@drawthingsapp

8 months

In this release, we also added prompt style template first seen from Fooocus project (). It is an easy way to specify style with SDXL models. (3/6)

1

12

Draw Things

@drawthingsapp

1 year

On our @discord server we have a friendly weekly competition! The top 4 contributions for the theme "Alien Landscapes", all done with @drawthingsapp , were these! Join our discord to participate, exchange ideas, and have your questions answered:

1

3

11

Draw Things

@drawthingsapp

10 months

Metal FlashAttention is an open-source project authored by @philipturnerar . It is available at .

GitHub - philipturner/metal-flash-attention: Faster alternative to Metal Performance Shaders

Faster alternative to Metal Performance Shaders. Contribute to philipturner/metal-flash-attention development by creating an account on GitHub.

github.com

2

1

12

Draw Things

@drawthingsapp

3 months

Our support for Würstchen ranges from 6GiB RAM iPhones all the way to largest Mac systems. You can read more about @dome_271 @pabloppp 's awesome model here: . It is OK with text and have slightly better prompt following comparing to SDXL. 2/6

1

12

Draw Things

@drawthingsapp

6 months

There are more to come! And hopefully this answered question about whether we are interested in videos or not once and for all. Enjoy!

0

12

Draw Things

@drawthingsapp

1 year

The ControlNet support as it is still limited. In future updates, we will add more ControlNet related features to explore this area in full! Meanwhile, please report back if you encountered any incompatibility issues with ControlNet 🙏

0

1

12

Draw Things

@drawthingsapp

3 months

We added Playground v2.5 to the model download list (). It is the same SDXL model but with slightly different schedules. It is too much hassle to import yourselves, so we did it for you. You can read how to import manually in our Discord server. 4/6

2

0

11

Draw Things

@drawthingsapp

2 years

What if the image generated is cropped like this?

1

12

Draw Things

@drawthingsapp

4 months

We bumped up resolution for our foreground / background segmentation model to 1024x1024, it now generates even more accurate masks around main protagonists! 5/6

2

0

12

Draw Things

@drawthingsapp

1 year

More over, latent diffusion methods have inherent problem with small complex subjects, such as faces or hands. But if you zoom-in, it will have much higher quality! This helps us to fix bad generations.

2

0

11

Draw Things

@drawthingsapp

1 year

Another tip: do you know that you can import textual inversion directly from ? Simply tap the play button and select Draw Things, it will download and name keyword correctly all automatically!

3

2

11

Draw Things

@drawthingsapp

2 years

The Inpainting model is amazing! and you can select from the @drawthingsapp . A thread:

Runway

@runwayml

2 years

We’re excited to release public checkpoints for Stable Diffusion Inpainting, which powers our Erase-and-Replace Tool. Code at our Stable Diffusion repo: Weights at the @huggingface model hub:

6

74

459

2

0

11

Draw Things

@drawthingsapp

4 months

We trained our own LCM LoRA for SDXL Refiner! This LCM ( @SimianLuo ) is distilled on SDXL Refiner and trained at last 300 denoising steps to be good at refining details at 1 to 2 steps. See the difference yourself! (5/6)

1

0

11

Draw Things

@drawthingsapp

10 months

SDXL also works well at lower resolution such as 768x768, for which you would expect comparable speed up (M1 Pro, 30-step, end-to-end, 80s). (3/9)

1

0

11

Draw Things

@drawthingsapp

7 months

In this release, we also added SSD-1B models from @_segmind . These models are about 80% to 100% faster on Apple systems with reduced quality. (4/6)

1

0

11

Draw Things

@drawthingsapp

1 year

First, you can either download the checkpoint file from Safari, or use AirDrop:

2

10

Draw Things

@drawthingsapp

1 year

Easy - use this unreleased feature ...

0

1

10

Draw Things

@drawthingsapp

9 months

We also polished up our Seed options. For new users, it will be new seed on each generation, which helps them to get a variety of images. For advanced users, we added "NVDIA GPU Compatible" seed mode, which will match initial noise from NVIDIA graphics card. (4/5)

1

10