v4.1.0 (Stable Cascade prompt weighting, MacOS arm64 support)

v4.1.0 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

4.1.0 Features

• Stable Cascade can now use long prompts with weighting via the compel and sd-embed prompt weighters, simply specify one of these two values to --prompt-weighter to enable the feature when using Stable Cascade.

• Built and published wheels for MacOS on Apple Silicon, dgenerate can now be used with --device mps on MacOS (arm64). See install instructions here: MacOS Install (Apple Silicon Only). This is very experimental.

• Console UI hotkeys dynamically adjust for useability when the UI is running on MacOS, I am sure that the hotkeys are very abnormal when compared to typical MacOS convention for text editors, but they all function and do not overlap with important system hotkeys. Insert (for multiline input) changes to Command+i, and Ctrl+Space (for run) changes to Command+r. All other hotkeys remain the same, including copy, paste, and cut.

• Added template function frange which is just range but for float ranges. This is useful for stepping through guidance scale values or LoRA scale values for example. It uses identical logic to range, but the step value may be a float, e.g. frange(0.1, 1.0, 0.2) -> [0.1, 0.3, 0.5, 0.7, 0.9]. The default step value is: 0.1

• torch 2.4.1

v4.0.3 (Flux, IP Adapters, T2I Adapters, Input Batching)

v4.0.3 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

4.0.3 Patch Fix

runwayml has removed their stable diffusion 1.5 repository from Hugging Face, breaking single file loads in the diffusers library and anything that uses it.

This release upgrades diffusers to 0.30.2 where the issue has been worked around by instead referencing Lykon/dreamshaper-8

All references to the runwayml model on Hugging Face have been removed from this project, including in documentation.

4.0.0 Features

The command line interface remains backwards compatible for all supported model types, the major version change is mostly due to breaking library API changes and the complete removal of Flax/Jax support.

All support for Flax/Jax has been removed because there is only minimal support or adoption of it anywhere currently, and seemingly no work being done on it in the diffusers library. Also this alleviates quite a bit of dependency hell and possible issues caused by having torch and Flax/Jax in the same package as dependencies.

• Support for Flux in txt2img mode, add associated options: --flux-second-prompts, --flux-max-sequence-length. Single file loads from a .safetensor file is partially supported for Flux, you must specify the VAE and Text Encoders manually.

• Add Flux support to sd-embed prompt weighter internal plugin, prompt weighting syntax can be used with Flux when --prompt-weighter sd-embed is specified.

• New Console UI recipes for Flux that are optimized by default to run (slowly) on even old hardware or GPUs with less than 12gb of VRAM.

• Stable Diffusion 3 now supports per LoRA scale values.

• The option --lora-fuse-scale has been added to control the LoRA scale once all LoRA models have been fused together at their individual scale values. This controls at what scale the merged LoRA weights are fused into the main model, these weights possibly being merged from multiple LoRA models at different individual scales.

• Support loading a transformer model for new diffusion architectures based on transformer architecture (SD3 and Flux) via the argument --transformer, this new option is nearly identical to --unet/--unet2. Transformers can be quantized via the quantize URI argument using optimum.quanto. The quantize argument accepts a string value which is the datatype for quantization, for example qint8, qfloat8, etc. Acceptable values are listed in the --help text of the --transformer argument.

• Text encoders specified with --text-encoders/--text-encoders2 can now be quantized via the quantize URI argument using optimum.quanto. Acceptable values are listed in the --help text of the --text-encoders argument.

• Support loading multiple IP Adapter models with --ip-adapters.

• Support loading an image encoder with --image-encoder for use with IP Adapters and Stable Cascade.

• New --image-seeds syntax for specifying IP Adapter images, see: Specifying IP Adapters

• Support loading multiple T2I Adapter models with --t2i-adapters, T2I Adapters function identically to Control Nets in terms of image input specification with --image-seeds. They are essentially another form of control guidance model, and support Open Pose / Depth inputs etc. similar to Control Nets. T2I Adapters are a bit less resource intensive, but do not currently support img2img or inpainting, only txt2img.

• Support image input batching for img2img and inpainting mask images, allowing multiple input images to be processed on the GPU at once. New --image-seeds syntax feature added to support this, see: Batching Input Images and Inpaint Masks

• --seed-image-processors and --mask-image-processors now support the + syntax previously supported by --control-image-processors in order to allow for selective processing of input images when multiple img2img or inpaint masks are specified. This is covered in the Image Processors section: Multiple control net images, and input image batching

• Added the resize image processor which implements basic image resizing using the usual resampling algorithms in Pillow. This can be useful for resizing input images to Stable Cascade when multiple are provided, as there is not a mechanism to do so from --image-seeds

• Stable Cascade can now use the new input batching syntax of --image-seeds to process multiple input images as a style reference, Stable Cascade does not use multiple images for batching like other pipelines, but as image prompts.

• Improved --help output formatting and readability for dgenerate and dgenerate --sub-command image-process

• torch 2.4.0 and diffusers 0.30.0

v4.0.2 (Flux, IP Adapters, T2I Adapters, Input Batching)

v4.0.2 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

4.0.2 Patch Fix

Add the arguments --flux-second-prompts, --flux-max-sequence-length, and --lora-fuse-scale to the argument
reconstruction used by --output-configs and --output-metadata. These arguments were not being replicated in config output.

4.0.0 Features

The command line interface remains backwards compatible for all supported model types, the major version change is mostly due to breaking library API changes and the complete removal of Flax/Jax support.

All support for Flax/Jax has been removed because there is only minimal support or adoption of it anywhere currently, and seemingly no work being done on it in the diffusers library. Also this alleviates quite a bit of dependency hell and possible issues caused by having torch and Flax/Jax in the same package as dependencies.

• Support for Flux in txt2img mode, add associated options: --flux-second-prompts, --flux-max-sequence-length. Single file loads from a .safetensor file is partially supported for Flux, you must specify the VAE and Text Encoders manually.

• Add Flux support to sd-embed prompt weighter internal plugin, prompt weighting syntax can be used with Flux when --prompt-weighter sd-embed is specified.

• New Console UI recipes for Flux that are optimized by default to run (slowly) on even old hardware or GPUs with less than 12gb of VRAM.

• Stable Diffusion 3 now supports per LoRA scale values.

• The option --lora-fuse-scale has been added to control the LoRA scale once all LoRA models have been fused together at their individual scale values. This controls at what scale the merged LoRA weights are fused into the main model, these weights possibly being merged from multiple LoRA models at different individual scales.

• Support loading a transformer model for new diffusion architectures based on transformer architecture (SD3 and Flux) via the argument --transformer, this new option is nearly identical to --unet/--unet2. Transformers can be quantized via the quantize URI argument using optimum.quanto. The quantize argument accepts a string value which is the datatype for quantization, for example qint8, qfloat8, etc. Acceptable values are listed in the --help text of the --transformer argument.

• Text encoders specified with --text-encoders/--text-encoders2 can now be quantized via the quantize URI argument using optimum.quanto. Acceptable values are listed in the --help text of the --text-encoders argument.

• Support loading multiple IP Adapter models with --ip-adapters.

• Support loading an image encoder with --image-encoder for use with IP Adapters and Stable Cascade.

• New --image-seeds syntax for specifying IP Adapter images, see: Specifying IP Adapters

• Support loading multiple T2I Adapter models with --t2i-adapters, T2I Adapters function identically to Control Nets in terms of image input specification with --image-seeds. They are essentially another form of control guidance model, and support Open Pose / Depth inputs etc. similar to Control Nets. T2I Adapters are a bit less resource intensive, but do not currently support img2img or inpainting, only txt2img.

• Support image input batching for img2img and inpainting mask images, allowing multiple input images to be processed on the GPU at once. New --image-seeds syntax feature added to support this, see: Batching Input Images and Inpaint Masks

• --seed-image-processors and --mask-image-processors now support the + syntax previously supported by --control-image-processors in order to allow for selective processing of input images when multiple img2img or inpaint masks are specified. This is covered in the Image Processors section: Multiple control net images, and input image batching

• Added the resize image processor which implements basic image resizing using the usual resampling algorithms in Pillow. This can be useful for resizing input images to Stable Cascade when multiple are provided, as there is not a mechanism to do so from --image-seeds

• Stable Cascade can now use the new input batching syntax of --image-seeds to process multiple input images as a style reference, Stable Cascade does not use multiple images for batching like other pipelines, but as image prompts.

• Improved --help output formatting and readability for dgenerate and dgenerate --sub-command image-process

• torch 2.4.0 and diffusers 0.30.0

v4.0.1 (Flux, IP Adapters, T2I Adapters, Input Batching)

v4.0.1 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

4.0.1 Patch Fix

Add missing VAE tiling checkbox for Flux recipes in Console UI recipe form.

4.0.0 Features

The command line interface remains backwards compatible for all supported model types, the major version change is mostly due to breaking library API changes and the complete removal of Flax/Jax support.

All support for Flax/Jax has been removed because there is only minimal support or adoption of it anywhere currently, and seemingly no work being done on it in the diffusers library. Also this alleviates quite a bit of dependency hell and possible issues caused by having torch and Flax/Jax in the same package as dependencies.

• Support for Flux in txt2img mode, add associated options: --flux-second-prompts, --flux-max-sequence-length. Single file loads from a .safetensor file is partially supported for Flux, you must specify the VAE and Text Encoders manually.

• Add Flux support to sd-embed prompt weighter internal plugin, prompt weighting syntax can be used with Flux when --prompt-weighter sd-embed is specified.

• New Console UI recipes for Flux that are optimized by default to run (slowly) on even old hardware or GPUs with less than 12gb of VRAM.

• Stable Diffusion 3 now supports per LoRA scale values.

• The option --lora-fuse-scale has been added to control the LoRA scale once all LoRA models have been fused together at their individual scale values. This controls at what scale the merged LoRA weights are fused into the main model, these weights possibly being merged from multiple LoRA models at different individual scales.

• Support loading a transformer model for new diffusion architectures based on transformer architecture (SD3 and Flux) via the argument --transformer, this new option is nearly identical to --unet/--unet2. Transformers can be quantized via the quantize URI argument using optimum.quanto. The quantize argument accepts a string value which is the datatype for quantization, for example qint8, qfloat8, etc. Acceptable values are listed in the --help text of the --transformer argument.

• Text encoders specified with --text-encoders/--text-encoders2 can now be quantized via the quantize URI argument using optimum.quanto. Acceptable values are listed in the --help text of the --text-encoders argument.

• Support loading multiple IP Adapter models with --ip-adapters.

• Support loading an image encoder with --image-encoder for use with IP Adapters and Stable Cascade.

• New --image-seeds syntax for specifying IP Adapter images, see: Specifying IP Adapters

• Support loading multiple T2I Adapter models with --t2i-adapters, T2I Adapters function identically to Control Nets in terms of image input specification with --image-seeds. They are essentially another form of control guidance model, and support Open Pose / Depth inputs etc. similar to Control Nets. T2I Adapters are a bit less resource intensive, but do not currently support img2img or inpainting, only txt2img.

• Support image input batching for img2img and inpainting mask images, allowing multiple input images to be processed on the GPU at once. New --image-seeds syntax feature added to support this, see: Batching Input Images and Inpaint Masks

• --seed-image-processors and --mask-image-processors now support the + syntax previously supported by --control-image-processors in order to allow for selective processing of input images when multiple img2img or inpaint masks are specified. This is covered in the Image Processors section: Multiple control net images, and input image batching

• Added the resize image processor which implements basic image resizing using the usual resampling algorithms in Pillow. This can be useful for resizing input images to Stable Cascade when multiple are provided, as there is not a mechanism to do so from --image-seeds

• Stable Cascade can now use the new input batching syntax of --image-seeds to process multiple input images as a style reference, Stable Cascade does not use multiple images for batching like other pipelines, but as image prompts.

• Improved --help output formatting and readability for dgenerate and dgenerate --sub-command image-process

• torch 2.4.0 and diffusers 0.30.0

v3.10.4 Console Stability Fixes

v3.10.4 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

v3.10.4 Patch Fix

Backport console shell process monitor from t2iadapter (4.0.0) branch.

This solves a console GUI hang when \exit is used or when the shell process exits unexpectedly on its own.

v3.10.1 Features & Fixes

See: v3.10.1 Release Notes

v3.10.3 release (Console UI broken stdin pipe fix)

v3.10.3 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

v3.10.3 Patch Fix

The stdin buffer of the interpreter process in the Console UI could be easily flooded by running many commands one after another which do not execute quickly, for example, queuing up multiple diffusion or image processing commands in terminal mode.

This could have also happened with large commands, such as when pulling up a long script from the consoles history and accidentally running it twice in short succession :)

On windows the stdin buffer of a process is fairly small, and filling it up results in a broken pipe followed by a deadlock in the Console UI.

This is resolved by using a write thread, a non-blocking fifo queue, and a pre-created pipe.

The console will now warn you if the command / write fifo queue is full, and will not accept new input if the pipe being read by the interpreter process cannot accept any more data.

The maximum size of this fifo queue is 50 entries, you should not see this warning unless you are really trying to :)

v3.10.1 Features & Fixes

See: v3.10.1 Release Notes

v3.10.2 release (Download directive caching fix)

see here for latest release

v3.10.2 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

v3.10.2 Patch Fix

The \download directive and download() template function now use the webcache database to store information about the paths of files downloaded to arbitrary locations on disk (not inside the cache folder). This removes the need to make a request to the URL to figure out the filename that is going to be served if the file has already been downloaded before to an arbitrary directory. At least up until the pointer file has expired in the cache.

This uses a pointer file in the webcache that undergoes normal cache expiry using the settings mentioned here in the readme: File Cache Control

This greatly improves the speed of the directive when the file it is trying to download already exists in the output location that you have defined, and is much nicer on servers than requesting the filename every time.

v3.10.1 Features & Fixes

See: v3.10.1 Release Notes

v3.10.1 release (Generic NCNN Upscaler)

see here for latest release

v3.10.1 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

v3.10.1 Features & Fixes

1.) Generic NCNN upscaler

ncnn has been added as a package extra. When ncnn is installed, the new image processor upscaler-ncnn is available for generic upscaling using NCNN, and should work with models converted from ONNX format. This is included by default in the Windows installer / portable install environment that is attached to each release.

This upscaler supports tiling just as the normal upscaler image processor does, and essentially the same options in terms of tiling with slightly different defaults.

It does not use the device argument, but instead a combination of use-gpu=True and gpu-index=N for enabling Vulkan accelerated GPU use on a specific GPU.

By default this processor runs on the CPU.

This is because the Vulkan allocator conflicts heavily with the torch CUDA allocator used for diffusion and other image processors when they are placed on the on the same GPU, and having both allocators on the same GPU can cause hard system lockups.

You can safely use this upscaler at the same time as torch based models by running it on another GPU that torch is not going to be using.

Once you have used this processor, be aware that the process will always exit with a non-zero return code, this is due to being unable to clean up the GPU context and certain ncnn objects properly through ncnn python bindings before the process shuts down. It will technically create an access violation / segfault inside ncnn, I am not sure what bad behaviors this will cause on Linux, but on Windows the process exits with no side effects or hang ups other than a non-zero return code.

See: dgenerate --image-processor-help upscaler-ncnn

And also: Upscaling With NCNN Upscaler Models in the readme.

2.) Memory Management

Image processors now have size estimates which are used as a heuristic for clearing out CPU side memory belonging to the diffusion model cache, prior to them being loaded into memory. This should help prevent avoidable out of memory conditions due to an image processor model loading when the diffusion model cache is using most of the systems memory.

This size estimate is also used as a heuristic for freeing up VRAM, particularly the last called diffusion pipeline if it currently is still in VRAM.

If an image processor still runs out of memory, due to its actual execution allocating large amounts of VRAM, it will attempt to free memory and then try again, if an OOM occurs on the second try then the OOM is raised.

Diffusion invocations will now attempt to clear memory and try again in the same fashion for CUDA out of memory errors, but not for CPU side out of memory errors, which are already more easily prevented by the heuristics that are already in place.

The main current enemy of this application running for long periods of time is VRAM fragmentation, which is not avoidable with the default CUDA allocator.

The example runner script in the examples folder has been rewritten to isolate each top level folder in the examples directory to a subprocess when not running with the --subprocess-only flag.

The only way to clear out the memory fragmentation after running so many models of different sizes is to end the process, so each directory is isolated to a sub process to take advantage of dgenerates caching behaviors for the directory, but to avoid excessive memory fragmentation by isolating a medium sized chunk of examples to a process.

There is also now an option --torch-debug in the run.py script which if enabled will try to dump information about objects stuck in VRAM after an OOM condition, and generate a Graphviz graph of possible reference cycles. Currently I cannot find any evidence of anything sticking around after dgenerate tries to clean up VRAM.

dgenerate now sets a PYTORCH_CUDA_ALLOC_CONF value max_split_size_mb of 512 before importing torch.

It also sets PYTORCH_CUDA_LAUNCH_BLOCKING to 0 by default.

These can be overridden in your environment.

3. Fetch CivitAI model links with --sub-command civitai-links

CivitAI has made a change to their website UI (*had some sort of outage) which renders right click copying of direct API links to models no longer possible.

I have written a dgenerate sub-command that can fetch API hard links to CivitAI models on a model page and display them to you next to their model titles.

The links that this command generates can be given directly to dgenerate, or used with the \download directive in order to download the model from CivitAI.

You can use dgenerate --sub-command civitai-links https://civitai.com/models/4384/dreamshaper for example to list all available model links for that model using the CivitAI API.

You can use the --token argument of the sub-command to append an API token to the generated links, which is sometimes needed for downloading specific models.

You can also use this sub-command as the directive \civitai_links in a config / shell mode or the Console UI.

See: dgenerate --sub-command civitai-links --help, or \civitai_links --help from a config / shell mode or the Console UI.

4. Config / Shell - Environmental Variable Manipulation

You can now use the directives \env and \unset_env to manipulate environmental variables.

# using with no args prints the entire environment

\env

# you can set multiple environmental variables at once

\env MY_ENV_VAR=1 MY_ENV_VAR2=2


# undefine them in the same manner

\unset_env MY_ENV_VAR MY_ENV_VAR2

See: dgenerate --directives-help env unset_env

5.) Config / Shell - Indirect Assignment

The config / shell language that is built into dgenerate now supports indirect assignment.

You can use a basic template expansion or environmental variable expansion to select the name of a template variable.

This now works for \set, \sete, \setp, and \env.

It also works for \unset and \unset_env

All other directives which accepted a variable name already supported this.

\set var_name foo

\set {{ var_name }} bar

# prints bar

\print {{ foo }}


\env VAR_NAME=BAZ

\env $VAR_NAME=qux

# prints qux

\print $BAZ

6.) Config / Shell - Feature Flags and Platform Detection

The config template functions have_feature(feature_name) and platform() have been added.

# have_feature returns bool

# Do we have Flax/Jax?

\print {{ have_feature('flax') }}

# Do we have NCNN?

\print {{ have_feature('ncnn') }} 


# platform() returns platform.system() string from pythons platform module

# prints: Windows, Linux, or Darwin.  etc...

\print {{ platform() }}

7.) Exception handing fixes in dgenerate.invoker

The methods in this library module were only capable of throwing dgenerate.DgenerateUsageError when they should have been throwing more fine grained error types when requested to do so with throw=True.

8.) Config / Shell - Parsing fixes

Streaming heredoc templates discarded newlines from the end of the jinja stream chunks, resulting in hard to notice issues with jinja control structures used as top level templates, mostly when the result of the heredoc template was being interpreted by the shell.

9.) Image processor library API improvements

Image processors will now throw when you pass a PIL image that possesses a mode value that the processor can not understand.

Currently, all image processors only understand RGB images.

10.) Console UI updates

Removed antiquated recipes related to image upscaling in favor of Generic Image Process and Generic Image Process (to directory)

From the generic image process recipes you can just select the upscaler or upscaler-ncnn processor from a drop down and fill out its parameters to preform upscaling.

---

All image processors now expose parameters provided by their base class in the UI, such as device, output-file, output-overwrite, and model-offload.

This allows the ability to select a debug image output location with a file select dialog. This is useful if you are trying to use an image processor as a pre-processor for diffusion and need to see the image that is being passed to diffusion for debugging purposes.

The device argument is hidden in the UI where not applicable, such as the Generic Image Process recipes where the UI selects the device for the whole command instead of via an image processor URI argument.

The device URI argument for image processors is available when selecting pre / post processors for AI image generation from the UI as well as when using the Insert Image Processor URI edit feature.

---

You can now specify the frame-start and frame-end ...

v3.9.3 release (Prompt Weighting!)

v3.9.3 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

3.9.3 Patch Fix

Fix unquoted $HF_TOKEN value in Stable Diffusion 3 and DeepFloyd console UI recipe templates, causing issues
with the \setp directive used in the template. Needed to have quotes around $HF_TOKEN in the template due
to changes in how and when the template renderer decides to automatically shell quote a form value.

Fix syntax highlighter failing to immediately re-highlight after Undo/Redo, Cut/Paste, Delete operations
preformed from the UI menus without using hotkeys. Syntax highlighting would not resume until the next
keypress, temporarily displaying unhighlighted text.

Documentation typo fix for letterbox image processor.

3.9.0 Features & Fixes

1.) Prompt weighting

Stable Diffusion Web UI / CivitAI / ComfyUI / InvokeAI style prompt weighting syntax is now implemented
through the argument --prompt-weighter and a new plugin class PromptWeighter.

This argument allows selecting a specific prompt weighter plugin to preform prompt parsing, token weighting, and prompt embed generation.

Prompt weighters circumvent restrictions on prompt length, which is typically 77 tokens.

Currently there are two prompt weighters implemented, compel and sd-embed.

Their documentation can be seen with: dgenerate --prompt-weighter-help compel sd-embed

The compel prompt weighter plugin uses the compel library to support InvokeAI style prompt weighting syntax, and also supports automatically translating your prompt from Stable Diffusion Web UI syntax into compel syntax via the plugin argument syntax=sdwui.

The compel prompt weighter supports Stable Diffusion 1/2, and Stable Diffusion XL model types:

• --model-type torch
• --model-type torch-pix2pix
• --model-type torch-upscaler-x4
• --model-type torch-sdxl
• --model-type torch-sdxl-pix2pix

The sd-embed prompt weighter uses the sd_embed library to support Stable Diffusion Web UI / CivitAI / ComfyUI prompt weighting syntax. This prompt weighter is currently using code from a fork I have VRAM optimized, found here

The sd-embed prompt weighter supports Stable Diffusion 1/2, Stable Diffusion XL, and Stable Diffusion 3 model types:

• --model-type torch
• --model-type torch-pix2pix
• --model-type torch-upscaler-x4
• --model-type torch-sdxl
• --model-type torch-sdxl-pix2pix
• --model-type torch-sd3

For more information about prompt weighting with dgenerate, see: Prompt Weighting and Enhancement in the readme.

2.) Prompt length warnings

Warn the user about prompt truncation when not using a --prompt-weighter implementation, this will warn you when your prompt is too long for the models tokenizer max sequence length. If you get this warning you should select a --prompt-weighter implementation, or reduce the length of your prompt.

3.) New image processors & image processor fixes

Implemented new image processors:

• anyline (MistoLine Control Any Line preprocessor, see: https://huggingface.co/TheMistoAI/MistoLine)
• lineart-standard (Basic non model based variant of lineart / lineart-anime)
• teed (TEED tiny efficient edge detector)
• zoe (ZoeDepth depth detector)

Fixed an issue with the detect-resolution argument of various image processors not returning the image to its original size
after detection was preformed.

4.) Improved extensibility

All plugin types implemented by dgenerate can now be loaded through --plugin-modules, and \import_plugins inside config scripts.

Previously only image processors, config directives, and config template functions supported external extensibility, in a hard coded manner internally. Now sub-commands and prompt weighters can be written as external plugins as well.

An architecture change has been made so that going forward any plugin loader class can automatically discover classes through --plugin-modules and \import_plugins without any extra work having to be done internally to support it.

New plugin code examples for sub-commands and prompt weighters can be found in the writing_plugins examples folder

5.) Improved argument validation

Better argument validation when using plugin help arguments such as --image-processor-help etc.. this corrects the problem
of the CLI allowing you to mistype something such as dgenerate --plugin-paths plugin.py --image-processor-help my-processor, resulting in your plugin not being listed because the correct command is infact:

dgenerate --plugin-modules plugin.py --image-processor-help my-processor

Previously the only validated argument was the plugin help argument, with all unknown arguments being ignored. Now when one of these arguments is encountered, the other arguments are validated, meaning it will raise an error on --plugin-paths for example as that is not a valid argument name.

6.) New template functions for image dimension manipulation

Added config template functions align_size, pow2_size, size_is_aligned, and size_is_pow2 for checking and adjusting the alignment of image dimension values inside of config scripts.

See: dgenerate --functions-help align_size pow2_size size_is_aligned size_is_pow2 for their documentation.

Or: \functions_help align_size pow2_size size_is_aligned size_is_pow2 if you are working with a config file or the Console UI.

7.) Console UI, insert recipe form overhaul

The insert recipe form now scrolls to support more form fields, and now supports selecting image processor implementations by name where applicable.

Image processor selection includes dynamic form controls for specifying image processor argument values, and a help button that you can click to show the documentation for the selected image processor.

The insert recipe form can now be used to select dgenerate argument values for --post-processors, --seed-image-processors, --control-image-processors, and --mask-image-processors using this new feature.

Two new recipes named Generic Image Process and Generic Image Process (to directory) have been added which make use of this feature to build an \image_process directive that can be inserted into your config.

There have been many other changes to the insert recipe form, including additional fields / access to more options, toggles, drastically improved input validation, and the ability to select a file OR directory for applicable model arguments.

8.) Console UI, new insert / build URI options

Two new URI insertion features have been added to the Console UI.

Using Edit -> Insert Image Seed URI will allow you to use a form with file select dialogs to build an --image-seeds URI, you can specify each component of the URI (seed image, inpaint mask, control image) and a URI with proper syntax will be formatted for you and inserted into the config. You may also specify the per image-seed resize & and aspect correct resize argument values (the default value for aspect correct resizing is True)

Using Edit -> Insert Image Processor URI will allow you to insert an image processor URI in a similar fashion to inserting a Karras Scheduler URI (a previously existing feature).

A form is generated based on the selected image processors arguments.

Inserting with this form will insert an Image Processor URI formatted with all the provided arguments.

9.) Console UI, version info / latest version check

The Console UI now shows the current software version next to the Help -> Homepage menu item.

If you are connected to the internet upon launch and github is reachable, a check for the latest release will occur,
if there is a newer release, a link to it will be added to the bottom section of the Help menu.

It will be in the format Newer Release Available! (vX.X.X), and clicking it will open your web browser to the page
for the latest release on github.

This way you can check if there is a software update available from the UI.

10.) Image generation improvements.

Stable Cascade can now produce images with output dimensions aligned to 128 instead of only power of 2 images.

Stable Diffusion 3 can now preform VAE tiling --vae-tiling on images with output dimensions aligned to 16 instead of only power of 2 images.

11.) Bugfix model loading from directories

Fix bug in determining diffusers model loading method, which caused loading diffusers models out of directories to stop working.

12.) Dependency updates

Diffusers: 0.29.2

controlnet_aux: 0.0.9

No longer vendoring most of controlnet_aux, only vendoring the PidiNet detector for a bugfix which...

v3.9.2 release (Prompt Weighting!)

newer release available here

v3.9.2 release with Windows installer.

Due to the size of the packaged python environment, the installer is within a multi-part zip file.

The multipart zip can be extracted using 7-Zip: https://www.7-zip.org/

Download both dgenerate_installer.zip.001 and dgenerate_installer.zip.002 to a folder.

Unzip dgenerate_installer.zip.001 to a directory (Right click, 7-Zip -> Extract to "dgenerate_installer") and then run dgenerate_installer\dgenerate.msi to install.

dgenerate will be installed under C:\Program Files\dgenerate by default with an isolated python environment provided.

The install directory will be added to PATH, and dgenerate will be available from the command line.

Portable Install

A portable install is provided via dgenerate_portable.zip.001 and dgenerate_portable.zip.002, these contain
nothing but the dgenerate executable and a frozen python environment which can be placed anywhere.

3.9.2 Patch Fix

Handle out of memory conditions for pipelines / image processors / and prompt weighters more thoroughly.

Preform memory cleanup as best as possible so that dgenerate can continue to run in REPL mode without
VRAM memory leaks after an OOM condition occurs.

And a few documentation additions and corrections.

3.9.0 Features & Fixes

1.) Prompt weighting

Stable Diffusion Web UI / CivitAI / ComfyUI / InvokeAI style prompt weighting syntax is now implemented
through the argument --prompt-weighter and a new plugin class PromptWeighter.

This argument allows selecting a specific prompt weighter plugin to preform prompt parsing, token weighting, and prompt embed generation.

Prompt weighters circumvent restrictions on prompt length, which is typically 77 tokens.

Currently there are two prompt weighters implemented, compel and sd-embed.

Their documentation can be seen with: dgenerate --prompt-weighter-help compel sd-embed

The compel prompt weighter plugin uses the compel library to support InvokeAI style prompt weighting syntax, and also supports automatically translating your prompt from Stable Diffusion Web UI syntax into compel syntax via the plugin argument syntax=sdwui.

The compel prompt weighter supports Stable Diffusion 1/2, and Stable Diffusion XL model types:

• --model-type torch
• --model-type torch-pix2pix
• --model-type torch-upscaler-x4
• --model-type torch-sdxl
• --model-type torch-sdxl-pix2pix

The sd-embed prompt weighter uses the sd_embed library to support Stable Diffusion Web UI / CivitAI / ComfyUI prompt weighting syntax. This prompt weighter is currently using code from a fork I have VRAM optimized, found here

The sd-embed prompt weighter supports Stable Diffusion 1/2, Stable Diffusion XL, and Stable Diffusion 3 model types:

• --model-type torch
• --model-type torch-pix2pix
• --model-type torch-upscaler-x4
• --model-type torch-sdxl
• --model-type torch-sdxl-pix2pix
• --model-type torch-sd3

For more information about prompt weighting with dgenerate, see: Prompt Weighting and Enhancement in the readme.

2.) Prompt length warnings

Warn the user about prompt truncation when not using a --prompt-weighter implementation, this will warn you when your prompt is too long for the models tokenizer max sequence length. If you get this warning you should select a --prompt-weighter implementation, or reduce the length of your prompt.

3.) New image processors & image processor fixes

Implemented new image processors:

• anyline (MistoLine Control Any Line preprocessor, see: https://huggingface.co/TheMistoAI/MistoLine)
• lineart-standard (Basic non model based variant of lineart / lineart-anime)
• teed (TEED tiny efficient edge detector)
• zoe (ZoeDepth depth detector)

Fixed an issue with the detect-resolution argument of various image processors not returning the image to its original size
after detection was preformed.

4.) Improved extensibility

All plugin types implemented by dgenerate can now be loaded through --plugin-modules, and \import_plugins inside config scripts.

Previously only image processors, config directives, and config template functions supported external extensibility, in a hard coded manner internally. Now sub-commands and prompt weighters can be written as external plugins as well.

An architecture change has been made so that going forward any plugin loader class can automatically discover classes through --plugin-modules and \import_plugins without any extra work having to be done internally to support it.

New plugin code examples for sub-commands and prompt weighters can be found in the writing_plugins examples folder

5.) Improved argument validation

Better argument validation when using plugin help arguments such as --image-processor-help etc.. this corrects the problem
of the CLI allowing you to mistype something such as dgenerate --plugin-paths plugin.py --image-processor-help my-processor, resulting in your plugin not being listed because the correct command is infact:

dgenerate --plugin-modules plugin.py --image-processor-help my-processor

Previously the only validated argument was the plugin help argument, with all unknown arguments being ignored. Now when one of these arguments is encountered, the other arguments are validated, meaning it will raise an error on --plugin-paths for example as that is not a valid argument name.

6.) New template functions for image dimension manipulation

Added config template functions align_size, pow2_size, size_is_aligned, and size_is_pow2 for checking and adjusting the alignment of image dimension values inside of config scripts.

See: dgenerate --functions-help align_size pow2_size size_is_aligned size_is_pow2 for their documentation.

Or: \functions_help align_size pow2_size size_is_aligned size_is_pow2 if you are working with a config file or the Console UI.

7.) Console UI, insert recipe form overhaul

The insert recipe form now scrolls to support more form fields, and now supports selecting image processor implementations by name where applicable.

Image processor selection includes dynamic form controls for specifying image processor argument values, and a help button that you can click to show the documentation for the selected image processor.

The insert recipe form can now be used to select dgenerate argument values for --post-processors, --seed-image-processors, --control-image-processors, and --mask-image-processors using this new feature.

Two new recipes named Generic Image Process and Generic Image Process (to directory) have been added which make use of this feature to build an \image_process directive that can be inserted into your config.

There have been many other changes to the insert recipe form, including additional fields / access to more options, toggles, drastically improved input validation, and the ability to select a file OR directory for applicable model arguments.

8.) Console UI, new insert / build URI options

Two new URI insertion features have been added to the Console UI.

Using Edit -> Insert Image Seed URI will allow you to use a form with file select dialogs to build an --image-seeds URI, you can specify each component of the URI (seed image, inpaint mask, control image) and a URI with proper syntax will be formatted for you and inserted into the config. You may also specify the per image-seed resize & and aspect correct resize argument values (the default value for aspect correct resizing is True)

Using Edit -> Insert Image Processor URI will allow you to insert an image processor URI in a similar fashion to inserting a Karras Scheduler URI (a previously existing feature).

A form is generated based on the selected image processors arguments.

Inserting with this form will insert an Image Processor URI formatted with all the provided arguments.

9.) Console UI, version info / latest version check

The Console UI now shows the current software version next to the Help -> Homepage menu item.

If you are connected to the internet upon launch and github is reachable, a check for the latest release will occur,
if there is a newer release, a link to it will be added to the bottom section of the Help menu.

It will be in the format Newer Release Available! (vX.X.X), and clicking it will open your web browser to the page
for the latest release on github.

This way you can check if there is a software update available from the UI.

10.) Image generation improvements.

Stable Cascade can now produce images with output dimensions aligned to 128 instead of only power of 2 images.

Stable Diffusion 3 can now preform VAE tiling --vae-tiling on images with output dimensions aligned to 16 instead of only power of 2 images.

11.) Bugfix model loading from directories

Fix bug in determining diffusers model loading method, which caused loading diffusers models out of directories to stop working.

12.) Dependency updates

Diffusers: 0.29.2

controlnet_aux: 0.0.9

No longer vendoring most of controlnet_aux, only vendoring the PidiNet detector for a bugfix which allows it to run on
other GPUs besides GPU 0. See: huggingface/controlnet_aux#109

Known Issues

Diffusers

When auto downloading a ControlNet model using a https://huggingface.co/ model slug, such as `dif...