Hey everyone, sorry it took so long to get this out. This update contains a bunch of things that I thought we had already released -- a couple bug fixes and some cool new features. Some very big things are on the way though, so be on the lookout for the next update.
- Add Multi Gaussian filter to Normal Map Generator (#2255)
- Add ability to limit the length of certain iterators (#2218)
- Add support for sub-pixel distance transform (#2235)
- (ncnn) Allow disabling Winograd/SGEMM (for cpu ncnn) (#2237 -- thanks @JeremyRand)
- Fixed being able to close the dependency manager while installing deps (#2227)
- Fixed RGBA preview image background (#2225)
- Fixed pip using bad caches on dep installs (#2232)
- Fixed dragging bug on image previews (#2236)
- Fixed bug from read-only image in CAS (#2239)
- Fixed CAS for grayscale images (#2241)
And thanks to these maintainers: @joeyballentine and @RunDevelopment
This is a hotfix to fix the Windows installer installing to the wrong place. If you installed v0.20.0 on Windows, I recommend checking if your system registers two separate chaiNNer installs, and uninstalling the extra one after applying this update. If your system is not registering a second chaiNNer instance, I recommend navigating to
C:\Users\[your user]\AppData\Local and seeing if a directory called
app exists, and if it contains
chaiNNer.exe. If it does, I recommend just deleting that
app directory entirely. The v0.20.1 installer will not automatically fix this for you. I am very sorry for this inconvenience.
- Fix windows installer installing to the wrong place (#2219)
- Fix PyTorch erroring when both CPU and FP16 modes were enabled (#2217)
Hey everyone! This is a pretty major update with some pretty overdue changes. The main big change is that we've upgraded to Python 3.11 (for integrated Python) as well as PyTorch 2.0 (if you're using chaiNNer for its PyTorch capabilities). Between these two changes, you should hopefully see a bit of an overall speed boost when processing things. That isn't all that got changed though, so take a look at the rest of the changelog to see everything that's new.
NOTE: After installing v0.20.0, integrated Python will be redownloaded, and you will need to re-install any dependencies you are using from the dependency manager (assuming you are not using system Python). Also, dependency-related settings have been reset in this build, so make sure to double-check that everything looks correct)
- Update Integrated Python to Python 3.11 (#2144, #2092)
- View Image nodes now have resizable image previews (#2167, #2174, #2205, #2214)
- Improved Video Frame Iterator codec and container selection (#2158 -- thanks @adegerard)
- Add ability to pass custom additional ffmpeg params to Video Frame Iterator (#2158 -- thanks @adegerard)
- MacOS builds are now code-signed (#2136, #2180 -- thanks @stonerl)
- MacOS builds are now universal (x64 and arm64) (#2162, #2166 -- thanks @stonerl)
- (PyTorch) DAT and SRFormer support (#2153, #2155, #2156)
- (PyTorch) Support for reading .ckpt files (for certain models) (#2193)
- (PyTorch) Support for DiffBR's SwinIR models (#2202)
- (PyTorch) FP16 now gets enabled by default for compatible GPUs (#2203)
- Linux build improvements (#2179 -- thanks @stonerl)
- Wayland Linux support (#2178 -- thanks @illode)
- Added BC5_SNORM format and additional format information to DDS options in Save Image (#2199)
- Added support for uncompressed DDS (#2201, #2211)
- Performance improvement for iterator progress updates that should use less CPU and lag the UI less (#2182)
- Visually show feature-disabled nodes as disabled (#2194)
- Added docs for Mip options in Save Image (#2207)
- Added anti-aliasing to Threshold node (#2209)
- Text as Image (#2125, #2213 -- thanks @adegerard)
- This node allows you to create an image from text, for use with overlaying on other images.
- Update PyTorch to 2.0.1 (#2143)
- Update ONNX-related dependencies for better 3.11 support (#2153)
- Fix .chn files being unable to be opened with chaiNNer on MacOS (#2150 -- thanks @stonerl)
- Fix BC5 DDS loading (#2198)
- Fix High Pass for grayscale images (#2206)
- Fixed logging empty lines (#2208)
and thanks to the following maintainers: @joeyballentine @rundevelopment @stonerl
This update contains a few bug fixes, a few new features, and a bunch of improvements specifically for macOS.
- Add position to Blend Images node (#2087 -- thanks @adegerard)
- Made some node outputs more consistent (#2111 -- thanks @stonerl)
- Add "Separate Alpha" checkbox to upscale nodes (#2127)
- Handle unsaved changes on chaiNNer restart (#2132 -- thanks @stonerl)
- Various macOS DMG improvements (#2122 -- thanks @stonerl)
- Set minimum macOS version to 11.0.0 (#2124 -- thanks @stonerl)
- Remove the macOS portable build as it was broken and has likely never worked (#2119 -- thanks @stonerl)
- Associate .chn files with chaiNNer on macOS (#2115 -- thanks @stonerl)
- Add "New Chain" and "Open Recent" menus to macOS dock (#2140 -- thanks @stonerl)
- Slight optimization for NCNN model loading when using CPU NCNN (#2142 -- thanks @JeremyRand)
- Fixed bug related to opening certain older chain files (#2116)
- Fixed tooltips being placed behind modal popups (#2128 -- thanks @stonerl)
- Fixed hotkeys (such as pressing F5 to run a chain) sometimes not working (#2114 -- thanks @stonerl)
- Fixed incorrect inputs on new unified Crop node (#2134 -- thanks @adegerard)
And thanks to the following maintainers: @joeyballentine @RunDevelopment
Another update which adds a few things and fixes a few bugs.
- (PyTorch) SCUNet architecture support (#2102, thanks @adegerard)
- Threshold node improvements (#2091)
- Generate Threshold (#2096)
- A node to generate the threshold value for the Threshold node.
- Fixed Save Image not creating directories correctly (#2099)
- Fixed issue with displaying image previews (#2101)
- Disable multi-instance support on MacOS as it did not work correctly (#2107, thanks @stonerl)
- Display modal overlays on top of toast messages (#2110, #2112, thanks @stonerl)
And thanks to the following maintainers: @rundevelopment
This update contains a few important bug fixes as well as some new long-requested features.
- Use Rust implementation for dither nodes (#2075)
- This implementation is between 100 and 500 times faster than the previous pure-python implementation.
- Use Rust implementation for Copy to Clipboard (#2079)
- This removes the multiple platform-specific dependencies as well as enables support for this node on silicon Macs.
- Add support for saving 16 bit and 32 bit images (#2083)
- Add setting for allowing multiple concurrent chaiNNer instances (#2089)
- Light mode improvements (#2064, thanks @stonerl)
- Add support for saving BMP in Save Image (#2077)
- "Webp (Lossless)" in Save Image is now just "Webp" with a lossless checkbox (#2080)
- Added warnings in node documentation for nodes with limited color depth (#2090)
- Quantize to Reference (#2071)
- A custom quantization implementation that tries to emulate a reference image as closely as possible. Useful for pixel art upscales.
- Percent (#2076, thanks @adegerard)
- Similar to the Number node, this node lets you input a percentage via slider to connect to other nodes.
- Fixed dropdown settings resetting their values unintentionally (#2081)
- This fixes the ONNX execution option problem people have been reporting recently. Make sure to check that you aren't using CPU by accident.
- Fixed Image File Iterator not sorting alphabetically (#2067)
- Fixed Stretch Contrast (#2082)
- Ignore file extension case in file image iterator (#2084)
Thanks to the following maintainers: @joeyballentine @RunDevelopment @theflyingzamboni
This update has a critical fix for the image file iterator as well as many MacOS improvements.
- System usage meters now use traffic-light color scheme (#2023, thanks @stonerl)
- MacOS clipboard improvements (#2031, thanks @stonerl)
- Help menu improvements (#2041, thanks @stonerl)
- Other MacOS improvements, such as a new icon (#2042, #2049, thanks @mafiosnik777)
- Per-platform dependency description improvements (#2047, thanks @stonerl)
- Add new option to enable hardware acceleration for the GUI rendering. We do not recommend enabling this unless you're using CPU upscaling as it will reduce GPU performance. (#2055, thanks @stonerl)
- Fixed glob pattern issues and related bugs with Image File Iterator (#2039)
- Fixed header layout for small screens & window sizes (#2036)
- Fixed broken input information for iterator errors (#2040)
- Fixed transparent image color contamination with Average Color Fix (#1907, thanks @Splendide-Imaginarius)
- Reduced the chances of seeing an error during an intentional backend restart (#2043, #2030)
- Fixed MacOS not quitting properly (#2051, #2049, thanks @mafiosnik777)
Hey everyone. It's been a while since the last update, and in my opinion, it's been worth the wait. We have a lot of interesting improvements and new features to share with all of you.
First, we now have documentation for all available nodes. This can be accessed both by a button in the top right button bar, as well as by right-clicking on a node in the sidebar and clicking "Open Documentation". The documentation includes a lot of information about nodes and their inputs & outputs. Not to mention, the docs are also searchable. This will hopefully be a huge help to both new and existing users.
Second, the update screen has finally received a makeover. This has been a long-overdue change. Starting with 0.19.0, you will get a popup when there is a new update, which will have an option for ignoring the update or viewing the update's changelog. If you view the changelog, a modal will open that will show you a full changelog of the update -- or multiple updates if you happen to be a few versions behind. From there you can view the release on GitHub or open chaiNNer.app to download it. If you do choose to ignore the update, don't worry, because you will see a green download indicator next to chaiNNer's version number which you can click on to view the update again.
Third, we've made some improvements to how the backend sets up chaiNNer. This is mostly to prep for future features, but it also conveniently makes it slightly easier to set up chaiNNer on a remote machine, as all the necessary dependencies will be installed via the backend, instead of the frontend. This shouldn't affect existing users much, except you may notice that on starting up chaiNNer this time you see a slightly different progress indication on the splash screen as it installs the few new necessary dependencies. We also now bundle a couple of small, cross-platform dependencies that should make first-time setup a bit faster. If anyone is having trouble with the new first-time setup process, definitely let us know ASAP.
Fourth, chaiNNer is now taking advantage of some first-party Rust code. This means we were able to speed up certain nodes by a significant amount. Expect future updates to improve the speed of other nodes as we move some intensive operations over to Rust.
And now for the changelog:
- Node Documentation (#1880, #1900, #1902, #1923, #1925, #1928, #1926, #1932, #1929, #1930, #1937, #1946, #1950, #1952, #1956, #1961, #1962, #1963, #1993, #1992, #1995, #2004, #2009, #2017)
- Update checking improvement (#1935)
- Mandatory dependencies now get installed via the backend (#1834, #1856, #1862, #1863, #1866, #1869, #1881, #1882, #1878, #1885, #1960)
- New design for sliders (#1966, #1971, #1976, #1975, #1977, #1986, #1987, #1988)
- Added Contrast-Adaptive-Sharpening to High Boost Filter (#1841)
- Added weight loading for NCNN Scale op and fix InnerProduct bug (#1864)
chainner_ext (rust code) for Fill Alpha (#1865)
- Added nearest color mode for Fill Alpha (#1883)
- Allow color inputs in merge transparency (#1888)
- Improved type error messages for colors (#1889)
- Add a wcmatch.glob field to image iterator (#1840, #1939, thanks @zeptofine)
- Hide unimportant labels (#1919)
- Add more algorithms to the Color Transfer node (#1903, thanks @adegerard)
- Update pytorch model file input label (#1927)
- Add window size detection to OmniSR (#1933)
- Splash screen is no longer always on top (#1938)
- Better computed alpha for upscaling nodes (#1942)
- Move system stats to backend (#1948)
- Changed empty-string behavior of text inputs (#1954)
- Shorter names for channel inputs/outputs in Split/Merge Transparency (#1967)
- Stack many images on Stack node (#1972)
- Consistently use sliders for blur radius-like number inputs (#1973, #1978)
- Optimized box blur implementation (#1974)
- Optimized Gaussian blur (#1981)
- Input number on enter (#1982)
- Variable precision for log-scale sliders (#1984)
- Added context menu entry to extract value into node (#1985)
- Added linked inputs group (implemented for Box Blur) (#1983)
- Use slider in Gamma node (#1989)
- Renamed "LUT" to "color palette" (#1991)
- Simplified printed structs by omitting default fields (#2003)
- Use a later version of pytorch that supports MPS for Arm MacOS (#2012)
- Update ONNX + ONNXRuntime & Improvements for ONNX on MacOS (#2010)
- Some UI fine-tuning around GPU selection on Mac (#2020, thanks @stonerl)
- Image Statistic (#1748)
- A node to get the min, max, mean, media, n-th percentile of an image.
- Get Bounding Box (#1884, #1901, thanks @zeptofine)
- Like Crop Content except it returns a bounding box insted of cropping.
- Regex Replace (#1893)
- Like Text Replace, but uses a regex search.
- Fast NL means (#1890, thanks @adegerard)
- Uses Fast NL means to denoise an image.
- Pick Color (#1913)
- Grab the color of a pixel in an image.
- Directory (#1921)
- Pick a directory and reuse it throughout the chain, like you can with numbers and strings currently.
- Switch (#1945)
- Connect multiple inputs and select which one you want to output.
- Stretch Contrast (#1959)
- Auto-levels-like functionality. Automatically improves contrast.
- Fixed 2 bugs in Dither Palette (#1844)
- Fixed Unsharp Mask modifying input image (#1851)
- Fixed bug with CLI mode timing out if it took too long to process (#1853)
- Fixed settings dropdowns where invalid values would get stuck with no way to change them (#1876)
- Fixed error reporting (#1908)
- Fixed ordered dithering (#1911)
- Don't check whether the directory exists in Save Image (#1912)
- Require RGB in Write Output Frame (#1920)
- Fixed linux & macos RAM usage reporting (#1948)
- Fixed wide dropdowns (#1958)
- Fixed round-half-up for negative integers (#1970)
- Enforce non-empty images (#1997)
- Add color limit for Palette from Image (#2005)
- Always estimate tile size as 256 on Mac for NCNN (#2006)
- Changed default tile size to 256 for ONNX (#2007)
- Backend-defined dependencies (#1834, #1856, #1862, #1863, #1866, #1869, #1881, #1882, #1878, #1885, #1960)
- Backend-defined GPU list (#1877, #1879)
- Added more contribution documentation (#1837, #1847, #1848, #1894, #2019)
- Added emoji support for node icons (#1831, thanks @jackcarey)
- Fixed backend script path detection with external Electron (#1839, thanks @JeremyRand)
--devtools argument (#1849, thanks @JeremyRand)
- Support NCNN CPU inference (#1867, thanks @JeremyRand)
--no-backend boolean arg to
- Enforce naming conventions (#1968, #1969)
- Upgrade to Navi 0.6.0 (#2001)
- Update minor/patch version of electron (#2022)
And of course thanks to the regular maintainers of the project: @joeyballentine @RunDevelopment @theflyingzamboni
I know I just did a release, but I couldn't wait to get this out. I finally fixed the broken windows installer :)
- Fixed the windows installer not opening after install/update (#1830)
- Fixed being unable to pass in electron-specific launch arguments to chaiNNer (#1830)
- (unconfirmed) Fixed MacOS and Linux portable builds to be truly portable (#1829)
Note: this is another reminder that the windows installer doesn't open after installing. It's a known bug we haven't been able to fix
Another small update with some bug fixes and a couple of new features.
- Add FFv1 support to Video Frame Iterator (#1779)
- (onnx) Add FP16 option for tensorrt (#1795)
- (pytorch) OmniSR Support (#1804, #1813, #1815)
- Pixelate (#1777, thanks @harisreedhar)
- Pixelate an image, similar to Photoshop's "Mosaic" filter
- Fixed bug with generating hashes after an upscale (#1797, thanks @Zarxrax)
- Fixed a bug that would show invalid nodes in the right-click node-selector context menu (#1821)
Also, thanks to the following maintainers: @joeyballentine @RunDevelopment
This is a small update with some new features and bug fixes.
- Small performance optimizations (#1743)
- Colors are now a data type and are selectable via color picker in many nodes (#1732)
- Added missing previewing features on Text Areas (#1766)
- Generate Hash (thanks @Zarxrax) (#1738)
- Z-Stack Images (thanks @JustNoon) (#1745)
- (webui integration) Image Upscale (thanks @mateusz) (#1692)
- (webui integration) fixed caching not respecting image changes (#1762)
- Fixed Fill Alpha causing errors in certain chains (#1768)
- Fixed iterators not properly catching certain types of errors (#1771)
Also thanks to the maintainers who contributed to this release: @joeyballentine @RunDevelopment
This is a slightly bigger update that didn't quite qualify for a major version number. It includes lots of bug fixes and a few new highly-requested features.
- Changed some node ordering and group names (#1679)
- Add weight support to CodeFormer when using Face Upscale (#1714)
- Add ETA to iterators (#1730)
- Optimizations (#1710, #1717)
- Rename Spritesheet Iterator's dimension inputs (#1716)
- Only show the "Separate Alpha" option for RGBA images (#1721)
- Better generated error reports (#1724)
- Removed the "disable hardware acceleration" option, and always disable hardware acceleration (#1722). A user noticed that having it off significantly improved upscaling times, so we figure there is no reason to use it (all of us devs have had it off). If you need an override because disabled hw acceleration is causing issues, please let us know.
- Fixed a bug with Combine RGB if no alpha channel was given (#1715)
- Fixed some bugs with output normalization (#1717)
- Fix ESRGAN-2c2 scale detection for lower scales (#1712, #1729)
- Fix "isDestroyed" startup errors (#1724)
- Fixed SD nodes not having text input handles (#1728)
- Fixed "Copy audio" when input video does not have an audio stream (#1733)
- Added support for using a later version of PyTorch when using system Python with 3.10, as previously it would just fail (#1672)
Thanks to the maintainers that contributed to this release: @joeyballentine @RunDevelopment @theflyingzamboni
This is a small bugfix release that addresses a few issues some users noticed since the last release.
- Add ISNet support to ONNX background removal (#1706)
- Fix ONNX RemBG node error (#1706)
- Fixed iterators constantly reloading starting nodes (#1702)
- Fix "dynamic scale" (they aren't really dynamic scale, but they act like it) realesrgan NCNN models (#1698)
- Fixed image dimensions in error messages (#1701)
Thanks to the contributing maintainers: @RunDevelopment and @theflyingzamboni
This version fixes a critical error that many people have experienced trying to install chaiNNer today, as well as adds a bunch of small new things and fixes some other bugs. While this is a smaller release, big things are coming soon :)
- Allow numbers as seeds (#1633)
- Reload starting nodes on file change (#1640)
- Various optimizations (#1653, #1664, #1667)
- Ctrl+Click on export button to export viewport to clipboard (#1658)
- Added model directory output for NCNN Load Model (#1663)
- Added blur noise to Create Noise (#1669)
- Added support for 2C2-ESRGAN architecture in PyTorch Load Model (#1670)
- Larger settings modal (#1675)
- Parse Number (#1621, thanks @JustNoon)
- Parses numerical text into a number
- High Pass Filter (#1630, thanks @harisreedhar)
- Performs a high pass filter on an image
- Fixed issue with backend starting up that started recently (#1695)
- Fixed overlapping handle connect radius (#1637)
- Fixed a bug connecting to SD webui (#1648)
- Fixed tiling for grayscale images (#1668)
- Fixed device detection for M2 Macbooks (#1674)
and as always thanks to the maintainers that contributed to this release: @joeyballentine @rundevelopment @adodge
This is a hotfix to fix an issue affecting all stable diffusion nodes.
- Fixed issue causing Stable Diffusion nodes to not work (#1623, thanks @adodge)
NOTE: This is another reminder that the installer still has an issue that makes chaiNNer not open automatically after install. It does still install, so just open it manually after.
This update addresses a few bugs, but mostly adds a few cool nodes.
- Consistent UI font between platforms + switch to using Open Sans (#1610, #1614)
- Seeds are now their own datatype rather than numbers (#1611)
- Text Slice (#1580)
- A node for getting a substring of text
- Text Length (#1580)
- A node for getting the length of text
- Round (#1603, thanks @adodge)
- A math node for performing rounding operations
- Derive Seed (#1611)
- Use various sources to derive a seed for use with nodes that deal with randomness
- (PyTorch) Guided Upscale (#1506, #1616)
- Based on PixTransform, guided upscaling upscales an image based on another
- For example, you can use a high quality diffuse texture to upscale a lower quality normal map of that texture
- Fixed running MAT on GPU (#1599)
- Fixed loading SRGB DDS (#1606)
- Fixed being able to drag on the + button on some nodes (#1607)
And thanks to the normal maintainers, @joeyballentine @RunDevelopment
NOTE: The installer still doesn't open after it is done installing. It does still install, so just manually open it after. We are still working on fixing that.
This is a small update which addresses some issues and adds a few features.
- Don't forget to check out the previous changelog if you havent already updated to 0.18.0.
- Better SD url detection (#1588, thanks @adodge)
- Now compatible when the SD webui is running under
--nowebui and TLS
- (PyTorch) Add MAT support to inpainting (#1596, thanks @theflyingzamboni)
- Create Noise (#1585, thanks @adodge)
- Create images made up of Value/Simplex noise
- Fixed iterator child nodes appearing underneath an iterator when the iterator was selected (#1593, thanks @joeyballentine)
- Fixed SD not working on macos (#1588, thanks @adodge)
NOTE: The windows installer still doesn't open after it is done installing. It does still install, so just manually open it after.
A huge requested feature is finally here! (well, kinda). Stable Diffusion support was something I've wanted to add for a while. However, I never thought I'd be able to match the experience from the AUTOMATIC1111 webui. Luckily for us, @adodge thought of a brilliant workaround for that: Just use the webui. Specifically, the webui's API. This way, we wouldn't have to set up any new dependencies or worry about integrating the code incorrectly -- it would just work. The only problem with this approach is it does require you to run the webui in api mode (by running
./webui.(sh|bat) --api before chaiNNer starts up. This will make a new category with the SD nodes appear. At the moment adodge is looking into an internal implementation, but for now this works pretty well.
There's plenty of other new features in this update as well, such as inpainting support (both via opencv and LaMa), rembg support, a bunch of new nodes, and more. See the full changelog below:
- Stable Diffusion support via the AUTOMATIC1111 webui API (#1500, #1540, #1553, #1559, #1562, #1557, #1575, #1586, #1583 thanks @adodge)
- Adds Stable Diffusion support to chaiNNer by running the webui in API mode
- Supports Text-To-Image, Image-To-Image, Outpaining, and clip interrogation
- Node recatigorization (#1576, #1581)
- LaMa inpainting support (#1569)
- RemBG background removal support (#1533)
- The external Stable Diffusion nodes mentioned above.
- Color Levels (#1532, thanks @harisreedhar)
- Color levels adjustment similar to Gimp and Photoshop.
- Dither & _Dither (Palette) (#1502, thanks @adodge)
- Dither an image, either by number of colors or with a specified palette.
- Palette From Image (#1502, thanks @adodge)
- Generates a palette from an image, to be used with dithering.
- Convert Normal Maps
- Allows conversion between DirectX and OpenGL normal maps, as well as to and from the octahedral normal map format used by RTX Remix
- Inpaint (#1546, thanks @harisreedhar)
- Uses a mask image to inpaint. This is not an AI inpainting node, but something more similar to content aware fill.
- Create Gradient (#1544, thanks @adodge)
- Generates a grayscale gradient with a variety of options, which can then be colorized using the LUT node.
- Signed Distance Field (#1552, thanks @adodge)
- Can be used to create low-resolution textures that can be scaled up rendered at a higher resolution with fewer artifacts.
- Dilate & Erode (#1565, thanks @harisreedhar)
- These two nodes work similarly to the dilate and erode features in Gimp and Blender.
- (PyTorch) Inpaint (#1569)
- Inpaint using the LaMa model, which performs significantly better than the opencv implementation used in the normal Inpaint node.
- (ONNX) Remove Background (#1533)
- Remove the background of an image using .onnx RemBG models.
- Added option to duplicate node with input connections (#1541)
- Slightly improved type error messages (#1542, #1579)
- Rearrange inputs on upscale nodes for consistency with other nodes (#1539, thanks @Zarxrax)
- Added "export viewport to clipboard" in the file menu (#1550)
- Added more DDS output format options (#1555)
- Added "Copy Audio" toggle for video iteration (#1563)
- Added "Use Pip Directly" toggle in the dependency manager (#1560)
- Changed the design of resize handles (#1584)
- Create node on enter in search pane, when only one node is an option (#1578)
- Fixed the bug that would cause TextAreas (like what is used in the Note node) to randomly get much larger (#1582)
And of course, thanks to the maintainers: @joeyballentine @RunDevelopment @theflyingzamboni
NOTE: As a reminder, the windows installer still does not open after installing anymore. It still does install though, so just manually open it after installation. We are actively investigating why this happened.
This update includes some new nodes, a quality-of-life improvement related to stopping execution, even more portability, and a bit more. See the full changelog below:
- Long-running upscales can now be stopped, without chaiNNer seeming to hang (#1521)
- Added DDS option for separating alpha (#1510)
- Slight styling changes for slider input numbers (#1504)
- Random Number (#1499, thanks @Zarxrax)
- Allows using a randomly generated number for number inputs
- NCNN Model File Iterator
- Similar to the PyTorch model file iterator, these allow you to iterate over folders of NCNN models.
- ONNX Model File Iterator
- Ditto above, but for ONNX
- Lens Blur (#1516, thanks @harisreedhar)
- Applies lens blur to an image
- Fixed RGBA DDS loading (#1514)
- Fixed chaiNNer closing when visiting the update link, or dismissing the box (#1518)
- Fixed some missing portability when using a portable zip (#1523, #1530)
And of course, thanks @joeyballentine and @RunDevelopment
NOTE: with this release it seems chaiNNer no long automatically opens after installing an update -- don't worry though, the installer still works.
This update adds a few somewhat important dependency updates for system compatibility reasons, as well as a few major new features, and lots of minor ones.
I'll keep this brief this time, and just have you read the changelog below:
- ONNX -> 1.13
- With this update, ARM (M1/M2) MacOS users can now use ONNX, since this update provides an arm64 MacOS version. This means you can properly convert to NCNN through ONNX.
- OpenCV -> 4.7 (automatic)
- Usually I don't mention dependencies that update automatically, but I wanted to mention this one specifically -- because now we should theoretically support MacOS versions under 10.15, specifically to 10.13 now. This has not been officially tested by us, but it should work. We probably scared away all the < 10.15 people by now, but if any of you see this, please give it a try now.
- CLI Mode (#1469, #1470, #1468, #1489, #1492, #1494)
- This was our first-ever commissioned feature, so thanks to Flaviu Radulescu for funding this feature.
- This feature now allows you to execute chains from the command-line, without needing the GUI to run. This includes being able to override inputs for easier chain reuse.
- Documentation can be found by running
chainner run --help in the command line.
- True portability for zip builds (#1484, #1487, #1489)
- ZIP builds were never truly portable as integrated python, settings, and a few other things were saved to appdata. Now, these will download & save to the root directory of the exe.
- If you don't want this functionality, simply delete the
portable file that is now present in the zip builds.
- Improved startup speed (#1468)
- chaiNNer now starts up significantly faster than it used to (assuming it has already gone through first-time setup and updates)
- Image Convolve (#1442, thanks @JumpyJacko)
- Runs arbitrary convolution with a user-defined kernel
- Metal to Specular & Specular to Metal (#1457)
- Converts metal/specular maps to specular/metal maps
- Number & Text
- These nodes allow easy re-use of numbers or text between nodes
- Apply LUT
- You can now apply a Look-Up Table (LUT) to images
- Support for mode DDS formats (#1471)
- This is specifically DXT1, DXT3, and DXT5. This is still a Windows-only feature.
- Improved JPEG saving options (#1480, #1483)
- This includes adding a chroma subsampling dropdown as well as a "progressive" option
- Allow dragging videos into chaiNNer for loading like images/models (#1488)
- Enhanced type tag size (#1404, #1478)
- Added output alpha option to Change Colorspace (#1405)
- Made directory and file name output types immediately computed (#1408)
- Added angle addition method for normal maps (#1409)
- Better error message when loading a PyTorch model fails (#1433)
- Add seed input to Add Noise node (#1444)
- Made the output type for Math node a bit better (#1452)
- Improved some type error messages (#1493)
- Improved logging (#1495)
- Fixed a small issue with number input width (#1443)
- Fixed potential errors caused by manually clearing PyTorch memory during out-of-vram events (#1456)
And as always thanks to the maintainers: @joeyballentine @RunDevelopment @theflyingzamboni
This is just a minor update which includes a couple of fixes and QOL improvements. Make sure to check last version's changelog if you haven't grabbed that version yet.
- Better execution time formatting (#1392, #1394)
- Increase default codeformer weight to 0.7 (#1381)
- Replaced some true/false dropdown inputs with checkboxes (#1377)
- Changed max of Add Normals amount to 200% (#1398)
- Made text of dropdown options larger (#1375)
- Fixed DDS rename not working across drives (#1376)
- Removed broken window menu from menu bar (#1386)
- Fixed some errors not showing in the UI (#1407)
As always thanks to @joeyballentine @RunDevelopment and @theflyingzamboni for contributing to this release.
This is a pretty big update with some noticeable changes and improvements, including a design overhaul and many quality-of-life features.
Major UI improvements (#1305, #1326, #1302, #1313, #1314, #1330, #1340)
- The first thing you might notice in this version is a bit of a redesign. Nodes have been majorly overhauled to be much more compact. While it might be a bit of a shock how different it looks at first, I guarantee you'll quickly get used to it and prefer this new look. The compactness of the nodes gives you much more space to fill out your chains with, and there is just a lot less wasted space overall. Check out this comparison of the space used by the old vs the new nodes
- With this compactness change also came a change to sliders. In my opinion, they are now a lot easier to see and interact with, and are closer to something like you'd see in blender.
- There have also been changes to pop-up boxes. Settings had a bit of an overhaul to fix some scrolling issues, and all pop-ups now have a darker color. Alerts also got icons representing the type of alert.
Node-on-Connection insertion (#1231, #1341)
- Something we all have always wanted in chaiNNer was the ability to insert a compatible node in-between two connected nodes, without having to delete the connection and make two new ones. Well, now this is a reality. All you have to do is hold
alt while dragging a node in the editor, and both the node and the connection line will visually show you that the node is able to be connected this way. It's really convenient to be able to insert nodes into existing chains this way, and can save a lot of time.
Informing you about invalid connections (#1299, #1306, #1308)
- The UI now will inform you when a connection you are trying to make is invalid. Before, we would tell you why a node was invalid after it was already connected and then made invalid. However, we had never explained why some connections were not possible to begin with. This had confused many users as they would try to do things like connect a GFPGAN model to Upscale Image and think there was something wrong rather than it being something intentional. While the UI would not let these users make that connection, it would not tell them why. This update adds an explanation tooltip that explains why a connection can't be made as you attempt to drop the connection on a handle. This should hopefully improve the user experience and lead to less confusion when this happens.
Increased View Image preview resolution (#1290, #1342, #1351)
- View Image just became a lot more useful. Now, instead of seeing only a 512x512 preview of your image, you will see a maximum of 2K resolution when zoomed in to the max. This still isn't ideal for extremely large images (larger than 2k) where you need to zoom in even more than chaiNNer allows, but a feature for that will be coming in the future. For now, this is a huge improvement and makes in-line previewing much much better. We also improved the performance of the node, so you may notice previews loading a bit faster than they used to.
Better DDS Support (Reading & writing, Windows only) (#1266, #1356)
- With this update, previously unsupported DDS textures files can now be read and written on Windows. We make use of Texconv, which is a small texture utility that we now bundle with chaiNNer. Unfortunately, Textconv is Windows-only, meaning Linux & MacOS users will not be able to take advantage of this. But we figured it was still better to have it than not.
- CodeFormer support (#1327)
- This Face Upscaling architecture was highly requested to be added, and now you can use it in chaiNNer with the Face Upscale node.
- NCNN Optimizer (#1259)
- NCNN models now get optimized on load as well as on convert, which can save help them save a little bit of time when upscaling. For most models, this will only be minute optimizations, but for batch upscaling it can save quite a lot of time overall.
- Show current overall execution progress on taskbar (#1343, #1365)
- Made amount input a slider in High Boost Filter (#1288)
- Added amount and threshold inputs for Unsharp Mask (#1293, #1294)
- Added the ability to cache ONNX TensorRT conversions (#1287)
- Added a context menu for multi-node selections (#1289)
- Added Ko-fi donation button to header (#1300)
- Improved Convert Colorspace node to be more generalized for alpha modes (#1322)
- Add .avs support to the video selection menu (#1345)
- Allow users to select alert text (#1349)
- Support node-search context menu in iterators (#1369)
- Create Color (#1285)
- Three new nodes that allow you to create single-color images in RGB, RGBA, or Grayscale.
- Surface Blur (#1292)
- This node blurs an image using a bilateral blur filter, also known as "surface blur."
- Fixed picking a custom python path (#1283, #1284, #1315)
- Fixed FaceSR (GFPGAN, RestoreFormer) and Transformer-model (SwinIR, HAT, Swin2SR) interpolation (#1281)
- Fixed the file select window prompting you to create files when opening directories that do not exist (#1282)
- Fixed the integrated FFMPEG download error to just be a warning (#1279)
- Fixed blurry checkers in image preview (#1320)
- Fixed being able to select text in file and directory inputs (#1331)
- Fixed opening a file with chaiNNer while an existing chaiNNer instance is already open (#1344)
- Fixed exact-size ONNX models not being properly run (#1336)
- Disallow negative crop values (#1350)
- While not enabled/disable behind the experimental features setting, both Apple MPS (via pytorch-nightly) and Microsoft DML (via pytorch-directml) support should theoretically work with a proper environment set up through your system python. People who have tried to set this up have had issues though, and ultimately I need more people to test this before I can officially say we support these things. However, if you want to test it yourself and know how to set these things up, feel free to give it a go and report back to us. (#1280, #1359)
As always, thanks to @joeyballentine @RunDevelopment and @theflyingzamboni
Theoretically, all model loading should now be fixed. However, if you are still running into models that previously worked but are now broken, let us know ASAP.
- Fixed another missing torch type when loading a model, which caused certain models to not load properly (#1272)
- Fixed iterator child node positions when placing presets (#1264)
- Fixed presets not being recorded in undo history properly (#1270)
Just a quick bugfix since we accidentally broke loading certain models in the last release. You don't really need this version unless you plan on using SwinIR, Swin2SR, or HAT. To see the previous changelog, go here.
- Fixed model loading for certain models being blocked by the new restricted loader (#1263)
Thanks @theflyingzamboni for fixing that
- Viewport Export (#1243) (thanks @Sryvkver)
- Export the current chain as a PNG using the bottom-most button in the button panel located at the bottom left of the chain editor.
- Ctrl+D now duplicates selected nodes (#1252)
- Added a context menu for connections (#1255)
- Blur and sharpen nodes have been reorganized and renamed (#1254)
- Iterators now have a "resize" indicator (#1257)
- Add Noise (#1251)
- Add different types of noise to an image
- High Boost Filter (#1254)
- A more traditional type of sharpening
- Fixed face upscaling causing broken video outputs (#1246)
- Fixed unsaved changes dialog box having extraneous &s on Linux (#1249)
- Fixed categories not hiding when a search would return no results for that category (#1253)
- Patched a security vulnerability
And thanks to the usual contributors, @joeyballentine @RunDevelopment @theflyingzamboni
This update contains a decent amount of quality-of-life changes. We have some new systems in place for making the experiences with inputs better (which will be brought to more nodes in the future), as well as some general UX improvements.
We also now have an opt-in experimental features setting that we are starting to use in order to get feedback on unfinished features, but without interrupting the regular experience of chaiNNer.
- Variable-length inputs (#1176, #1184)
- Now certain nodes can have more than just 4 inputs, with a button that lets you add more.
- This is limited to just Text Append and Text Pattern at the moment, but will be brought to more nodes in the future.
- Conditional inputs (#1200)
- Now certain nodes can show or hide some inputs based on things like items selected in a dropdown.
- This is currently only present in the Save Image node, but will probably be used with more nodes in the future.
- Experimental Features Toggle (#1142)
- We now have a setting (in the advanced tab) for turning on "experimental features". This will opt you into using anything we deem "experimental" that is in chaiNNer but not quite ready for prime-time. I will list anything new that is experimental under a new Experimental Features heading in the changelog.
- Drag N' Drop support for more filetypes (#1215)
- This mainly includes all model file types (.pth, .param/.bin, .onnx)
- Added "WEBP (Lossless)" option to Save Image (#1186)
- We had previously accidentally made it impossible to save lossless WEBP, and now you can
- Better "Save Changes" popup (#1191)
- The "Save Changes" dialog box that appears when you try to quit with unsaved changes is now a more standard format, with "Save", "Don't Save", and "Cancel" options.
- Show the current position of alerts in the alert box queue (#1193)
- Added menu entry for manually reloading/refreshing starting nodes and their previews (#1194)
- Added "Copy to clipboard" button for all alerts (#1203)
- Added type tags to strings and directories (#1216, #1226, #1244)
- Changed placeholder text on some inputs (#1223)
- Added custom installation animation on Windows (#1230)
- Moved update check to after the main window opens (#1232)
- Normal Map Generator (#1181, #1208)
- Generate Normal Maps from images, with a variety of customization options similar to Nvidia Texture Tool
- Presets (#1142, #1240, #1241)
- We now have some built-in bundles of nodes that we figure could be useful for setting up a variety of chains. However, they are not what we would consider feature complete yet (hence why they are behind the experimental features toggle). We would like to hear feedback from users as to what they think would be helpful (for example, clusters of nodes they tend to use all the time) and we can change/add presets accordingly. There is also still a lot of UI to be done around these, but they are at least usable at the moment.
- Fixed being unable to read MKV video duration (#1189)
- Fixed saving as GIF in Save Image (#1201)
- Fixed pressing arrow keys inside text and number inputs moving nodes around (#1205, #1212)
- Fixed FP mode not being loaded correctly for NCNN (#1210)
- Disable shortcuts while alerts are open (#1214)
- Fixed clear not affecting text inputs (#1217)
- Fixed iterator helper node type persistence (#1222)
- Reset invalid dropdown values (#1227)
- Fixed supported video format list in file open window (#1242)
Thanks as always to the main contributors: @joeyballentine @RunDevelopment @theflyingzamboni
- HAT Support (#1169, #1172, #1173)
- Adds the HAT architecture, which is another improvement on SwinIR
- Offset image preview nodes created by the shortcut button based on output index (#1178)
- Prevent FaceSR models from being converted to ONNX & NCNN due to being incompatible (#1136, #1174)
- Calculate number of frames estimate when nb_frames data is unavailable on video (#1136)
- Fixed issue where entire video would attempt to be processed when an error occurs (#1136)
- Fixed Swin2SR "compression" model loading (#1171)
- Fixed improper fp16 ONNX to NCNN conversion (#1162)
- Fixed slider value resetting visually but not actually on double-click (#1179)
- Fixed Split Transparency erroring when inputting grayscale images (#1182)
As always, thanks to everyone who contributed to this release: @joeyballentine @RunDEvelopment @theflyingzamboni
This is a small update that fixes a few bugs and adds a few things as well.
- Swin2SR Model Support (#1158)
- The successor to SwinIR now works in chaiNNer!
- Image Pairs Iterator Node (#1157)
- A highly requested feature: being able to iterate two paired directories, for things like already upscaled image comparisons. Limited to only two directories at the moment.
- Added handle to Write Output Frame directory input (#1163)
- Improved Change Colorspace's inputs (#1139)
- Fixed rounding inconsistency between backend and type hints (#1161)
- Added proper error message for libx264 failures in the Video Frame Iterator and made it stop immediately after an error (#1163)
Thanks to everyone involved with this release: @joeyballentine @RunDevelopment @theflyingzamboni
After a long wait, 0.14.0 is finally here. This update brings a couple of pretty important changes.
One of the biggest changes is improved video support. Many of you have found that the current video support isn't great, as it has limited encoding options which tend to not work depending on the machine. This is due to how OpenCV's video support works (or I guess, doesn't work), which I won't go into detail about here. As of this update, the video support has been revamped to use a fully-featured FFMPEG build. This means that now we are able to have many more encoding options compared to before, so it gives you a lot more flexibility. It also means that it's pretty much guaranteed to work on any system now. This is a huge improvement that is long overdue.
Another long overdue change is a revamp of how auto-tiling works. For a long time, chaiNNer hid tiling options from users, and just automatically handled it behind the scenes. If you don't know, tiling is something that many upscaling applications implement to get around VRAM limitations when an upscale would use more VRAM than your GPU physically has. We eventually decided to expose manual tiling options to the users, but this still used our internal "split factor" system, leading to the option being called "number of tiles". This was temporary but stuck around longer than we thought it would, and it rightfully confused users.
Thanks to @RunDevelopment's tiling changes, we now have a similar but much better implementation that uses actual tile sizes. This is much more in line with other upscaling applications and is a bit easier to comprehend and plan around. We also have more options in the dropdown, as well as new VRAM estimation for NCNN. You'll see this as "Auto (estimate)" and "Maximum". Picking "Maximum" will function similarly to how it did previously, while "Auto" will function by pre-determining a tile size to use based on available VRAM. PyTorch has been using this method for a while, but now NCNN has it as well. And of course, now we have better options for picking actual tile sizes, meaning you can use the same tile size on multiple images and not have to worry about changing it all the time (if you need to set it manually, that is). We still generally recommend you use the Auto or Maximum modes though.
There's also a lot more in this update, so check out the full changelog below.
New Features / Major Changes
- Proper "Tile Size" setting on upscale nodes (#1056, #1060, #1095, #1112, #1115)
- This replaces the "Number of Tiles" dropdown
- VRAM estimation for NCNN (#1068)
- Similar to how PyTorch has been for a while, Auto mode now pre-estimates the required tile size based on available VRAM instead of auto-splitting.
- This should hopefully cut down on the amount of vkQueueSubmit errors, since now we theoretically should prevent out-of-memory errors from happening.
- Better video support in Video Frame Iterator (#1103, #1108, #1110, #1113, #1123)
- We are now using a proper FFMPEG build to handle video iteration. This means we have unlocked far more encoding settings to make available for you to use, plus added stability.
- Paste image from clipboard directly into chaiNNer (#1102)
- Compression option for saving an image as JPEG or WEBP (#1126)
- Added NCNN model preview and type information to the Load Model node, similar to PyTorch (#1124, #1133, #1147)
- Added Lightness slider to Hue & Saturation node (#1080)
- Changed how brightness in Brightness & Contrast node works (#1081)
- Improved the internal Nvidia GPU check (#1083)
- Changed how Change Colorspace defines "from" and "to" values (#1089, #1138)
- More blend modes (#1072)
- Added HSL and CMYK to Change Colorspace (#1068)
- Double-click a slider to reset its value (#1097)
- Renamed "Amount" to "Radius" for all blurs (#1098)
- Slightly improved logging to reduce both spam and user confusion (#1116)
- Adjusted PyTorch's VRAM estimation values tio allow more VRAM use at once (#1120)
- Improved Text Append default values and made separator optional (#1125)
- Improved some error messages (#1121, #1134, #1137, #1141)
- Allow middle-click panning, even over nodes (#1063)
- Text Padding
- Add padding to any text or number, mainly for formatting numbers to a specific number of characters (such as converting "2" into "000002")
- Improved output type for crop content node (#1066)
- Fixed light theme issue on node timers (#1119)
- Fixed bugs that allowed improper cross-iterator connections (#1128)
- Fixed copying grayscale images (#1132)
- Prevented some crashing when certain node import errors occur (#1135)
- Fixed image iterator improperly sorting lower/uppercase (#1144, #1145)
- Made Sharpen consistent with blur nodes (#1150)
PS, we also now have nightly builds, which you can find here. Feel free to try these at any time to test out the newest features ahead of release.
As always, thanks to the main contributors: @joeyballentine, @RunDevelopment, @theflyingzamboni
Sorry for the delay in updates. The initial plan was to release a smaller update soon after 0.13.0 got released, but we ended up delaying it due to some issues with some upcoming things we merged. Those features will be out eventually, but for now I figured I'd release this small patch update with a few changes/bugfixes.
- Node categories will now display an indicator if any nodes in that category are missing, and will let you know there are extra dependencies you need to download. We had something similar before, but this is much more robust. For example, it should be much more obvious now when you are missing the newer torchvision & facexlib dependencies. (#1047)
- The utility category is now above PyTorch/NCNN/ONNX, at the very bottom of the "built-in" nodes (that do not require extra dependencies), so don't be alarmed if you don't see it in its usual spot. (#1051)
- Added video name and directory to Video Frame Iterator (#1050)
- Connections can now be removed with a double click (#1055)
- Fixes an issue where some nodes would throw an error if used directly after upscaling a grayscale image (#1070)