Full technical notes for this release: docs.lykos.ai/stability-matrix/release-notes/2.16.5
Added
- Added Fizgig, a LoRA training studio for Flux 2 Klein 9B, Krea 2, MiniMax H3 and Qwen Image 2.1 with dataset prep, training, LoRA repair and extraction in one app (#1761) - thanks to @e-nord!
- Added AMD ROCm support for Stable Diffusion WebUI Forge - Neo on Windows and Linux. Forge Neo's developer supports NVIDIA only, so report AMD issues with it to Stability Matrix (#1770) - thanks to @NeuralFault!
Changed
- ComfyUI now installs and updates from its current GitHub home, Comfy-Org/ComfyUI
Fixed
- Fixed Wan2GP installs on Python 3.11 failing with
sageattention ... has no wheels with a matching Python version tag. New Wan2GP installs now use Python 3.11 with PyTorch 2.10 and CUDA 13 with matching Triton, SageAttention and Flash Attention, the same setup Wan2GP's own installer uses. CUDA 13 needs NVIDIA driver 580 or newer. GTX 10XX cards stay on Python 3.10 - Existing Wan2GP installs on Python 3.10 keep their PyTorch 2.7 setup when updated. If a newer model such as LTX 2.5 fails with a Triton error, add Wan2GP again as a new install, which now uses Python 3.11, and move your outputs across (#1756)
- Fixed ComfyUI updates downgrading SageAttention post6 to post4 and reinstalling compatible wheels (#1764).
- Fixed ComfyUI installation and updates failing on Python 3.13+ because the NumPy 1.x constraint has no compatible wheel (#1754)
- Fixed AI-Toolkit installation failing on Linux while unpacking Node.js with a "Dangerous link path" error (#1767)
- Fixed Windows ROCm PyTorch installs pairing mismatched
torchandtorchvisionversions, which broke packages atimport torchvision(#1769) - thanks to @NeuralFault! - ControlNet preprocessing in Inference now offers to update older tagged Inference Core Nodes installs to a version with the ControlNet import fixes and Python 3.12/3.13 support
- Fixed the package uninstall confirmation counting your whole shared model library for packages using symlink sharing. Only the package's own files are counted; shared models were never deleted
- Fixed logs ignoring
--home-dir, and opening a missing folder or file showing an error dialog instead of a short notice (#1766) - thanks to @e-nord! - Fixed shared model folder links hiding DiffusionModels in the Checkpoint Manager, Image Lab and Inference, including when DiffusionModels is itself a link (#1743, #1763) - thanks to @NeuralFault and @ionite34!
- Fixed the Model Browser's Base Model filter being lost when using Next Model / Previous Model or refreshing a model's details (#1436)
- Fixed Open on CivArchive opening a file URL instead of the model page on Linux (#1765)
- Fixed emoji showing as
��in Image Lab messages and in CivitAI and CivArchive model descriptions - Fixed release-note text being cut off beside the scrollbar in the update dialog (#1757)
- Fixed text in Image Lab messages and thinking blocks being cut off at the right edge. Message bubbles now fit their text, and inline code is easier to read
- Fixed Save & Regenerate in Image Lab deleting later images and dropping the edited message's attachments. It now shows progress and can be cancelled
- Fixed Regenerate in Image Lab removing the old response before the new one was ready. A failed or cancelled regenerate now keeps it
- Fixed Image Lab pulling you back to a conversation when its generation finished while you were in another, and showing that generation's error in the one you had open
- Fixed deleting one image from a multi-image message in Image Lab deleting your original file when Image Lab hadn't kept its own copy
- Fixed your message in Image Lab not being editable or deletable after a failed or cancelled generation
- Fixed cancelling a Gemini generation in Image Lab showing "Generation failed" instead of "Cancelled"
- Fixed Image Lab not scrolling to a new message when you were already at the bottom of the conversation
- Fixed a ComfyUI generation in Image Lab not stopping when ComfyUI disconnected, if you had opened a Gemini conversation in the meantime
- Fixed package version lists and update checks showing an old release as the newest, because startup update checks ran into GitHub's rate limit
Supporters
🌟 Visionaries
This patch is a good one for AMD owners: Forge Neo now runs on ROCm on both Windows and Linux, and Windows ROCm installs always get a torch and torchvision that actually match. It also adds Fizgig, a whole LoRA training studio, to the package list. Most of it arrived as community contributions, and our Visionaries are the reason we have time to review that work properly and say yes to it. Thank you Waterclouds, MrMxyzptlk12836, bluepopsicle, dispenser, Droolguy, TwistedDragon, snotty, cusalapapen1481, moon_milky2843, sn3232323233350, CC, hasezou11013179, and Akros. It's great to see SnooSnooEternal back with us and jweg79 moving up from the Pioneers, and hello to japublicspot78364519 and jonny90, our newest Visionaries; we're so happy you're here. 💛
🚀 Pioneers
To our Pioneers SinthCore, Jisuren, Hurbie53, Cjloha, Alligator1907, CommissarGiygas16050, Zombop, Silerae, joshsciascia72, cy_hart691842, and okachako: thank you for backing a patch full of fixes that make things work on more machines. And hi acleveranagram, our newest Pioneer; we're really glad to have you! 💛