Scheduler Toolbox - ComfyUI DIY Schedulers
Added 2024-10-05 15:01:58 +0000 UTCVideo - https://youtu.be/Sc6HbNjUlgI
Overview of steps
1. Pick a scheduler and steps (always 8 for an initial hyper lora image)
2. Connect it (optionally via split sigmas) to the debug node to view it's values
3. Add a CustomScheduler node & set the steps
4. Copy the values from your chosen scheduler into the custom one
5. Tweak!
Notes
Some values may work better than others, depending on the situation. What works well for a photographic style image may not be the best for a painting or cartoon image. Absolute value's aren't really possible to provide as it's more about the difference between the values at each step along the way. These notes relate to the Flux model, other models will vary.
Generally speaking, anything less than 0.9 for the first step is too little noise and will produce a bad image. 0.00 for the final step is a full denoise.
As for the numbers in between, you don't always have to go down. Sometimes increasing the value for the next step may change the image in a way you prefer.
Workflow Information
The attached workflow is the one used in the video, and you too can use it to build your own custom scheduler. The custom scheduler nodes with preset values included were tested with ddpm++ 2m, but work without change for some other samplers too. Some have "bad" values on purpose to show what happens when changing that value.
Includes the "debug output (text)" nodes to view scheduler values and the example float value node too, so hopefully manager will auto-install those custom packs.
Example screenshot of a completed workflow run:

Custom Nodes
ComfyUI Manager typically installs missing nodes well, but if it fails then you can use search instead. Copy & paste any of these names as required:
Impact Pack
Derfuu_ComfyUI_ModdedNodes
WAS Node Suite
Extra Models for ComfyUI
ComfyUI-GGUF
ComfyUI_SUNoise
Comfyroll Studio
rgthree's ComfyUI Nodes
ComfyUI Extra Samplers
Use Everywhere (UE Nodes)
KJNodes for ComfyUI
ComfyUI CustomScheduler
ComfyUI-ppm
Models
Shakker Labs = https://huggingface.co/Shakker-Labs
Hyper-SD = https://huggingface.co/ByteDance/Hyper-SD
AuraSR = https://huggingface.co/fal/AuraSR-v2
Upscalers = https://openmodeldb.info/
The two model nodes are optional, just an easy way to switch between either for those with lower VRAM cards. Lower VRAM card users may prefer GGUF. You can also delete the loader (and switch) you don't use.
Comments
Just need a decent flux ipadapter face plus!
Nerdy Rodent
2024-10-06 20:47:51 +0000 UTCThis is getting complicated. How about an updated reposer.😁
Clayfacer
2024-10-06 13:41:32 +0000 UTC