add ltx framewise decode

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

CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## Update for 2025-01-13
 
+Two weeks since last release, time for update!  
+*What's New?"  
+- Large [Wiki](https://github.com/vladmandic/automatic/wiki)/[Docs](https://vladmandic.github.io/sdnext-docs/) updates  
+- New models: **Allegro Video**, new pipelines: **PixelSmith**, updates: **Hunyuan-Video**, **LTX-Video**  
+- New schedulers (TDD)  
+- Improvements to **Detailer**, **XYZ grid**, **Sysinfo**, **Logging**  
+- And a tons of hotfixes...  
+
+### Details for 2025-01-13
+
+- [Wiki/Docs](https://vladmandic.github.io/sdnext-docs/):
+  - updated: Detailer, Install, Update, Debug, Control-HowTo, ZLUDA  
 - [Allegro Video](https://huggingface.co/rhymes-ai/Allegro)  
   - optimizations: full offload and quantization support  
   - *reference values*: width 1280 height 720 frames 88 steps 100 guidance 7.5  
@@ -12,12 +24,8 @@
   - select from *scripts -> pixelsmith*  
 - [Hunyuan Video](https://github.com/Tencent/HunyuanVideo) LoRA support
   - example: <https://huggingface.co/Cseti/HunyuanVideo-LoRA-Arcane_Jinx-v1>
-- ZLUDA v3.8.7
-  - new runtime compiler implementation: complex types, JIT are now available  
-  - fast fourier transformation is implemented  
-  - experimental BLASLt support via nightly build
-    - set `ZLUDA_NIGHTLY=1` to install nightly ZLUDA: newer torch such as 2.4.x (default) and 2.5.x are now available  
-    - requirements: unofficial hipBLASLt
+- [LTX Video](https://github.com/Lightricks/LTX-Video) framewise decoding  
+  - enabled by default, allows generating longer videos with reduced memory requirements  
 - **Logging**:
   - reverted enable debug by default  
   - updated [debug wiki](https://github.com/vladmandic/automatic/wiki/debug)  
@@ -43,11 +51,15 @@
   - since different TAESD versions produce different results and latest is not necessarily greatest  
     you can choose TAESD version in settings -> live preview  
     also added is support for another finetuned version of TAESD [Hybrid TinyVAE](https://huggingface.co/cqyan/hybrid-sd-tinyvae-xl)  
+- **ZLUDA** v3.8.7  
+  - new runtime compiler implementation: complex types, JIT are now available  
+  - fast fourier transformation is implemented  
+  - experimental BLASLt support via nightly build  
+    - set `ZLUDA_NIGHTLY=1` to install nightly ZLUDA: newer torch such as 2.4.x (default) and 2.5.x are now available  
+    - requirements: unofficial hipBLASLt  
 - **Other**
   - **XYZ Grid**: add prompt search&replace options: *primary, refine, detailer, all*
   - **SysInfo**: update to collected data and benchmarks  
-- [Wiki/Docs](https://vladmandic.github.io/sdnext-docs/):
-  - updated: Detailer, Install, Update, Debug, Control-HowTo, ZLUDA  
 - **Fixes**:
   - explict clear caches on model load  
   - lock adetailer commit: `#a89c01d`  

installer.py

@@ -492,7 +492,7 @@ def check_diffusers():
     t_start = time.time()
     if args.skip_all or args.skip_git:
         return
-    sha = '03bcf5aefef13a064c34b605e489c0730052cca8' # diffusers commit hash
+    sha = 'c944f0651f679728d4ec7b6488120ac49c2f1315' # diffusers commit hash
     pkg = pkg_resources.working_set.by_key.get('diffusers', None)
     minor = int(pkg.version.split('.')[1] if pkg is not None else 0)
     cur = opts.get('diffusers_version', '') if minor > 0 else ''

scripts/ltxvideo.py

@@ -160,6 +160,7 @@ def run(self, p: processing.StableDiffusionProcessing, model, model_custom, deco
         shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model)
         shared.sd_model.vae.enable_slicing()
         shared.sd_model.vae.enable_tiling()
+        shared.sd_model.vae.use_framewise_decoding = True
         devices.torch_gc(force=True)
 
         shared.sd_model.transformer.cnt = 0