update changelog

Signed-off-by: Vladimir Mandic <[email protected]>

CHANGELOG.md

@@ -1,14 +1,23 @@
 # Change Log for SD.Next
 
+## Update for 2026-09-21
+
+- **Models**
+  - [Qwen-Image 2.1](https://huggingface.co/Qwen/Qwen-Image-2.1) unified text-to-image and image editing model  
+    7B single-stream block-causal transformer with Qwen3-VL 8B text encoder and 64-channel RGBA autoencoder  
+    native transparent output, up to 10 reference images for editing, 2K native resolution  
+- **Compute**
+  - update `diffusers==0.41.0.dev0`
+  - `rocm` windows installation move to *stable*, thanks @resonantsky  
+    at the moment, this results in installation of `torch==2.13.0` with `rocm==10.0.0`  
+- **Fixes**
+  - prompt cache: bypass when condition images are encoded
+  - group offload: places the vae on-demand behind the vae hijack
+  - vae tiling: tile size setting applies to 3d vaes, 16x vae tile geometry
+  - pixelsmith: vae tiling, thanks @li-lizhe  
+
 ## Refresh for 2026-09-19
 
-- [Qwen-Image 2.1](https://huggingface.co/Qwen/Qwen-Image-2.1) unified text-to-image and image editing model  
-  7B single-stream block-causal transformer with Qwen3-VL 8B text encoder and 64-channel RGBA autoencoder  
-  native transparent output, up to 10 reference images for editing, 2K native resolution  
-- update diffusers to 0.41.0.dev0
-- prompt cache bypass when condition images are encoded
-- group offload places the vae on-demand behind the vae hijack
-- vae tile size setting applies to 3d vaes, 16x vae tile geometry
 - preview mapping for reference networks
 - name mapping for network models and loras
 - additional DLSS reporting

TODO.md

@@ -4,7 +4,6 @@
 
 - Control tab verify overrides handling, @vladmandic
 - LTX: Implement LTX2DFRPipeline
-- ROCm: v10
 - Video: unify execution path for ui and api
 - Torch: update ipex, rocm to torch==2.14