StreamDiffusionTD 0.2.0🚀🖼️ [ Changelog ]

Get the TOX here >>> https://www.patreon.com/posts/110313168

StreamDiffusionTD 0.2.0 [ Installation + Setup ] >>> https://www.youtube.com/watch?v=F2JSWAKex4k

New Features

• V2V temporal consistency for smoother output
  - New V2V parameter page (alt + hover over parameters for more info on each parameter)
  - Can be enabled/disabled during streaming or before start
  - Not supported with TensorRT

• Pause / Play / Unload commands
  - Streaming can be paused or unloaded to free up resources/VRAM
  - Can be restarted (immediately if paused) without rebooting the server process by hitting Start Stream (or turning Pause to OFF, if paused)
  - 'Stop Stream' pulse still shuts down the entire server
  - 'Unload Models' keeps the server alive and listening but frees vram. (Config - model parameters + step sliders + acceleration mode, etc - can be updated when unloaded before starting stream again)

Improved Model Loading + Offline usage

• ControlNet compatibility for sd-turbo + sdxs + sdxl-turbo added
  - New 'Download ControlNet' option will automatically download controlnet for 1.5, sd-turbo, and sdxs models with tested working configurations. Choose which model type to download and it will store the ControlNet file in 'StreamDiffusion/Models/ControlNet' folder

• All models run offline (unless downloading for first time)
  - Added support for direct link to huggingface safetensors file

• Force Model Pipeline option
  - Forces the model loading process to use a specific model type
  - Helpful when model loads as the wrong type
  - See additional printouts in server CMD prompt that display the model type being loaded

• Add option to set Huggingface Cache (in Install parameter page)

• LoRA Loading now supports swapping models and weight adjustment
  - When loading the LoRA, there may be a pause in the output from streamdiffusion, but adjusting the weight once loaded should work with minimal lag
  - Cmd window will print error if LoRA tries to load and cannot be loaded, but will not interrupt the output.

Added display icons to show general settings in top left of the operator (can be toggled off in About page.)

• Shows server status, txt2img or img2img, and whether or not tensorrt, V2V, and controlnet are each active.

"Create Synced Component" (Synccreate) pulse added (Settings 2 page, under 'Op Create')

• Adds a base component with independent timeline that moves forward at the same rate as your current stream

• Adds necessary callback code for syncing with StreamDiffusionTD (onReceiveFrame must be ON)

• Exposes timeline parameters with additional Reset/Resetpulse on newly created component

TensorRT model loader

• Dropdown list of built TensorRT engines in current installation of streamdiffusion

• "Load Engine Parameters" pulse (Setengine) for loading model, step count, and TensorRT settings

• Note: Some settings (e.g., CustomLCM/SkipLCM toggles) may need manual adjustment

And More:

• Skip LCM LoRA toggle in Models page can be used to not load an LCM LoRA... Useful for models with merged LCM weights (like PhotonLCM)

• Much improved print logs in the command window + more detailed messages for common errors.

• Optimized image buffer in/out of TD for fewer frame drops

• Removed built-in parameters (option to toggle back on in About page) + renamed "Stream Settings 1" page to "Settings 1"

• Improved installation process + updated install guide + fixed install guide loading issue