Draw Things: AI Generation Взлом 1.20230801.0 + Чит Коды

* Fixed issues with importing SDXL model. Now these model can be imported / converted to 8-bit to save disk too.

* Fixed issues with import SDXL LoRAs. Note that there are two types of LoRA: one for SDXL Base and one for SDXL Refiner. Both are supported but they are not compatible with each other.

* Updated vae used for SDXL. You should tap SDXL model again to download the updated vae.

This is a hotfix version against 1.20230726.1.

Fixed issue with old CLIP encoder tokenization (it should pad unknownToken rather than 0) when applied to SDXL. This should improve quality of SDXL model.

Fixed issue with MFA GEMM gating, now on macOS 12 / iOS 15, it is disabled by default.

Added SDXL 1.0 model for image generation (best used on iPad with 8GiB or larger RAM).

Support refiner, which is a quick way to incorporate models specialized into generating details into the generation process.

Fixed some robustness issues with VAE (automatically switch to FP32 when FP16 fails).

Faster image generation with custom GEMM (10% to 20% faster than baseline).

* Introduce 8-bit models. These are the models quantized to 8-bit, but should have very similar output as the original. These models are around 800MiB in size, about 2x smaller than original. You can create 8-bit model off your existing model to save space.

* Fixed crashes related to saving infinite canvas. Now it should work for much larger canvas (we tried 20,000 in width).

* When load photos, it tries the best to load at its native resolution so you can zoom in, fix, and save again, while retaining the original resolution.

* Remember your last selection on save to photos, files or copy to clipboard.
* Ask for output resolution if it deemed too large.
* Fixed a crash related to infinite canvas.
* Added "new seed on generation" option for seed.

This is a bugfix release.

* Fixed drop & drop target with infinite canvas.
* Download VAE files if civitai import model doesn't have their own.
* Fixed LoRA import error for text-encoder only LoRA models.
* You can load mask on empty canvas again.
* Seed button now support long press with a little bit more options (old seed history, enter seed etc).
* You can mask / paint with an rectangle, rather than only a brush previously.
* Fix pose loading issue when you move to an empty position, zoom in / out will show old pose again.

Introducing Infinite Canvas! Infinite Canvas allows you to generate images on a infinite size canvas. It allows you to zoom in, inpainting details, fix faces, or zoom out, and generative fill the environment. All your favorite ControlNet, T2I adapters, img2img, LoRA, Textual Inversion, face restoration, upscaler, can be applied seamlessly on the canvas. It truly streamlines the workflow to create any image from your imagination.

Updated new user experience! As a first-time user, you are now presented with 3 options: select a model to download from our official server; import a locally downloaded PyTorch model; or enter a Civitai model URL to download and import model from their server.

Added ControlNet Tile model. ControlNet Tile is best companion to upscale with Infinite Canvas. It helps to retain some semblance even at high denoising strength while adding a lot details.

Added "Paste Civitai Generation Data" option in the context menu of the prompt text box. It parses the model / sampler / steps and other configuration options, while also insert automatic line breaks for long prompt. This feature helps to reproduce sample images from the site.

This is a bugfix release.
* Fixed crash with edge cases involving line breaks and textual inversion.
* Fixed unreasonable results if you have extreme imbalance of negative / positive prompts lengths (for example, 1,000 tokens for negative prompt and 90 tokens for positive prompt).
* Added the ability to pull down / up on live preview to resize it.

* Added live preview so that you can see how the image was generating while it is generating.
* Added language selector in Machine Settings so it is easier to switch to a different language.
* Support newline. You can break your prompt into different lines, and that will help the computer to know they don't have relationship with each other and emphasize more on each line (long press on the prompt edit view now will show "Insert Line Break").
* If you are downloading models from civitai.com, we will load all the configurations for you (V-Prediction, Attention at Higher Accuracy).
* Speed up "Deep Clean and Vacuum" by skipping images that is already compressed.

* Samplers now run on double precision, it might change previous generated images in subtle ways, but should have slightly better quality especially at higher steps.
* Fixed crashes with Kandinsky when Preload enabled.
* Added "Global Average Pooling" option for Shuffle ControlNet, this enables more variety of the model.
* Added "Image Prior Steps" and "Negative Prompt for Image Prior" options for Kandinsky model. Higher step counts and enable the latter option have positive effect on the quality of your generation.
* Fixed model download "Unexpected Error" issue.
* Added option to "Deep Clean & Vacuum" the project file when exporting.

* Support model "Multi-Language HD (Kandinsky v2.1)". This is a model that is not Stable Diffusion in origin, and can support multiple languages as prompt. That means you can write prompt in Spanish, German, French, Russian, Portuguese, and it will generate images following these texts. You can also use Chinese / Japanese / Korean (but need to set "CLIP Weight" to 0) as well.
* Part of this release, we support Spanish, Chinese (Simplified), Chinese (Traditional) as the interface language. We will expand the UI language support with our community.
* Fixed multi-ControlNet usage. Now orders of ControlNet doesn't matter any more.
* Fixed ControlNet weights where more than 100% previously will be clipped down to 100%.
* For model / textual inversion / LoRA importing, now if you paste civitai related links (page link, or the download link), we will automatically recognize these links and fill in relevant fields.

* Added freeform resolution support, but you have to experiment yourself to know which resolution can work on your device!
* Added Screen Saver option (in Settings -> Machine Settings). Enable this option will dim the screen during the image generation.
* Added MLSD ControlNet support. This will extract the lines from image, which is useful for architectural type of image generation.
* Optimized the storage so the project file growth will be a bit tamed (about 50% or more smaller).

* Added Face Restoration with RestoreFormer. It works best with photorealistic images and may degrade your anime / artistic images. You can use "Presets" to trigger Face Restoration only after image generated.
* Added Shuffle & Instruct Pix2Pix ControlNet from ControlNet 1.1 family.
* Added 4x UltraSharp upscaler.
* Fixed canny ControlNet doesn't use custom layer properly.

* Added Mask Blur. This help inpainting to blend better with the rest of images.
* Fixed LoRA usability issues on 3GiB / 4GiB devices.
* Added ControlNet v1.1, and new Control modes: SoftEdge, NormalBae, LineArt, LineArt Anime, Segmentation, Inpaint. You can use these by adding guiding image through "Custom ..." in guiding layers. Inpaint model should be just used as the normal inpainting main model with erasers.
* Custom guiding layer is also respected for Canny and Pose controls.

Fixed crash when using DPM++ SDE Karras sampler with T2I adapter.

Fixed T2I adapter canny / scribble doesn't work as expected.

Added UniPC sampler.

* Added T2I adapter support. T2I adapter is a faster alternative to ControlNet, which provides similar controls such as pose, canny, depth and scribble.
* Added color palette model from T2I adapter, which can guide color usage for the image generation process.
* Support multi-ControlNet / T2I adapter. Now you can combine controls, such as using both depth / canny / pose together to guide image generation.
* Can copy pose to clipboard. This helps to move the edited pose from one image to another.
* Bug fix: now when importing LoRA, it won't be affected by external folder selection when importing model.
* Minor bug fix: now when importing a model, you can select custom VAE from files too.
* Minor UI improvement: now it will show red warning background if your selected LoRA / Control is not compatible with the selected model.

* Support export models such that you can share your mixes on other websites for people to download / enjoy!
* Can select external folder to store models, such that you can move some of your models to iCloud or other SSD storage devices.
* When importing LoRA / models / Textual Inversion, now you can directly select the download files without moving them around.

* Support OpenPose formatted JSON for Pose.
* Fixed LoRA importing bug which doesn't exactly show 100% style you are looking for.
* Allow you to add additional limbs with paintbrush in Pose editing interface.
* Added "Model list" page where you can download models or delete existing models.

Added LoRA model support.

What is LoRA models? "Low-ordre Rank Adaptation" is a technique to fine-tune Stable Diffusion base model with different styles, subjects at a small reasonable additional size. You can find many LoRA models online nowadays.

With this version, you can use LoRA models, use them together, or import custom LoRA models into the app.

We support both LoRA models for Stable Diffusion 1.x and 2.x. You can enable CoreML support with LoRA models as well.

Also updated the UI a bit so the text box is a little bit larger on iPhones.

We fixed many usability issues / bugs in this version!

* Added ControlNet v2.x model support.
* Allow load / paste JSON pose file.
* Added "No Prompt" mode for ControlNet (stronger guidance from control layer alone).
* Make touch for pose / scribble more sensitive.
* Making sure inpainting model is compatible with ControlNet.
* Reset pose / scribble canvas when open new project.
* Fixed model importing and textual importing issues.

Added Pose editor to use in conjunction with ControlNet. This feature automatically detect pose from the image, you can also pinch to zoom the pose, move joints, or move the whole pose around. You can also load pose from different image (the full image, not the skeleton), use that as a starting point. For joint you don't want to use, you can use eraser to remove.

When pasting images with transparency into scribble canvas, it now treat the transparency as white regardless the original color.

Add support for Scribble (ControlNet)! Now there are several view layers you can select from: Depth & Scribble. In Scribble layer, you can either guide an existing image to a different tilt, or scribble something completely new and generate new image based on that.

Added helper settings text "Text to Image" / "Image to Image" to guide first-time user how these worked in combination to "Strength".

Raised max steps to 150. Reminder that you can always tap the number to increment it by 1.

This is a small bugfix version.

Fixed ControlNet depth map range issue. The network accepts in range 0 to 1, different from depth2img (in a range -1 to 1).

Added 3 more control knobs for ControlNet, which helps to customize the generation process.

Preview ControlNet support!

In this release, we added two ControlNets support: Canny Edge Map and Depth Map. You can expect more future ControlNet support very soon.

Canny Edge Map: This is very helpful when doing img2img task. It will keep shape intact while changing other parts of the image. If you have this on while Strength at 100%, the existing image will still influence the subsequent image generation (will try to keep up with the shape of the existing image), be aware of that!

Depth Map: This brings Depth Map support (loading separate depth map through long-press / right-click Camera button, or just load pictures taken with Portrait mode) to v1.x models. Loading a depth map, using any v1.x model and they will be guided by the depth map.

Note that at this point, ControlNet will only support v1.x models. Image generation with ControlNet will be slower due to the extra computation required.

Added tooltips around common features to make it easier to understand what each features are.

Added CLIP Skip option in Settings.

Fixed a image history overriding bug.