SD.Next Release 12-11-2025 #4464
Locked
vladmandic
announced in
Announcements
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
SD.Next release 2025-12-11
What's new?
New native kanvas module for image manipulation that fully replaces img2img, inpaint and outpaint controls, massive update to Captioning/VQA models and features
New generation of Flux.2 large image model, new Z-Image model that is creating a lot of buzz, new Kandinsky 5 Lite image model and new Photoroom PRX model
And first cloud models with Google Nano Banana 2.5 Flash and 3.0 Pro and Google Veo 3.1 video model
Also new are HunyuanVideo 1.5 and Kandinsky 5 Pro video models
Plus a lot of internal improvements and fixes
ReadMe | ChangeLog | Docs | WiKi | Discord | Sponsor
Details for 2025-12-11
FLUX.2-Dev is a brand new model from BFL and uses large 32B DiT together with Mistral 24B as text encoder
model is available for text, image and edit tasks and can optionally use control input as second input image
this is a very large model at ~100GB, so use of prequantized model at ~32GB is strongly advised
using prequant version and default offloading, model runs on GPUs with ~20GB
note: model is gated
Z-Image is a powerful and highly efficient image generation model with 6B parameters and using Qwen-3 as text encoder
unlike most of new models that are far larger, Z-Image architecture allows it to run with good performance even on mid-range hardware
note: initial release is Turbo variant only with Base and Edit variants to follow
it comes in text-to-image and image-edit variants
first cloud-based model directly supported in SD.Next UI
note: need to set
GOOGLE_API_KEYenvironment variable with your key to use this modelPRX (Photoroom Experimental) is a small 1.3B parameter t2i model trained entirely from scratch, it uses T5-Gemma text-encoder
HunyuanVideo 1.5 improves upon previous 1.0 version with better quality and higher resolution outputs, it uses Qwen2.5-VL text-encoder
distilled variants provide faster generation with slightly reduced quality
larger 19B (and more powerful version) of previously released Lite 2B models
note: need to set
GOOGLE_API_KEYenvironment variable with your key to use this modelkanvas is a full replacement for img2img, inpaint and outpaint controls
see docs for details
experimental: report any feedback in master issue
massive update to both features and supported models, thanks @CalamitousFelicitousness
models:
mooondream-2featuresmoondream-3-previewqwen3-vlwith thinkinggemma-3-vlfinetunesXiaomiMiMoui:
e.g. actualy mark detected regions/points
features:
with option to output answer-only or reasoning-process
/controlendpoint is now fully compatible with scripts/controladditional params to to control xyz gridsee
cli/api-xyz.pyfor simple example/detailersnew endpoint to list available detailers, both built-in and any custom downloaded/face-restorersexpanded to list model folderstorch==2.9.1for cuda, ipex, openvino, rocm backends1.0.2Beta Was this translation helpful? Give feedback.
All reactions