comfyui-better-flow

Nodes for controlling workflow and reducing VRAM fingerprint. All assembled for ComfyUI for the goal of making workflows easier and spare computing space.

progress

• (Model Offload/Recall)
  • fixed for GGUF flux model by unpatching the model (ModelPatcher and GGUFModelPatcher)
    • overrides the patch_on_device property

TODO: untested release

• offload/recall not tested on low_vram and high_vram flags and other models besides flux
• evtl nunchaku support?
• Cache any cleanup not working

TODO: add image use cases for each custom node

Custom Nodes

Model offload/ Model recall

When triggered during runtime, it moves models on their load_device (e.g. cuda) or offload device(e.g. cpu). For finer control of VRAM use.

Works for flux via gguf model and .safetensor model, regardless of the low_vram flag (when calling ComfyUI) or the load_on_device property (from GGUFModelPatcher) Nunchaku is unsupported/ignored since it mangages its own VRAM usage outside python.

Any to Hash

returns a md5 hash for the input object. Limitations:

• doesn't support None inputs
• Inconsistent with objects of type Tensor: Those can be located on RAM or VRAM, hence returning different values. In the case of an image, we migrate the image to CPU before calculating the Hash

Any to Hash x2

Same as any to Hash but combines two individual hashes. It re-hashes the concatenation of the md5 hash of each input

Cache any

For caching/ignoring time consuming steps in a workflow.

This node saves new input into a .pkl file, referenced by

• a key name
• a hash calculated by its input During the next execution, if a matching name+hash is found it will skip the heavy computing and return the pickled value instead.

[!Tip ]Multiple inputs keys can be used when combined with "any to hash x2" (e.g. image + text)

Checks whether the name+key exists:
- NO: 
	- requests the any_to_cache input
	- write cache file
	- returns input
- YES:
	- ignores input
	- reads from cache file
	- returns from cached file
- Finally
	- deletes other cached files in the `output/cached_outputs` folder (todo: fix it)

Known issues:

• Tensor type objects seem to produce unreliable hashes. I put a fix for images but some data types must dealt with (e.g. Conditionnings)

Experimental nodes (not even tested)

Wait

Stops the execution of the workflow until the trigger value has been executed. Due to lazy loading, this specific node is deprioritized.

Wait Multi (experimental)

Wait multiple triggers at once, (dynamic number of input?: the node dynamically adds new inputs after new input connected)

Reroute Triggerable

Reroute that can be triggered using the advanced Mode>OnTrigger UI elements. It enables a leaf node such as "image preview" not not be a leaf anymore if connected.

Install

Don't forget to remove the comfyui-offload-models (because it only works on the .safetensor models, and can be confusing).

Via ComfyUI

Search for "comfyui-better-flow" in the available node, or paste the github repo link to install from it.

Manual Install

# Go to the custom nodes
cd ./ComfyUI/custom_nodes
# Install the repo
git clone https://github.com/lokinou/comfyui-better-flow.git

versions

• 0.1.2
  • Model Offload/Recall
    • fixed for GGUF flux model by unpatching the model (ModelPatcher and GGUFModelPatcher)
    • ModelPatcher: overrides the patch_on_device property
  • Hash, Cache Any: images now produce consistent hashes
• 0.1.1
  • bug fixes
• 0.1.0
  • initial release from all node
  • migrated here (comfyui-offload-models 1.1.0)