Fooocus KSampler Custom Node for ComfyUI SDXL

Recently I tried Fooocus by Lyumin Zhang (Illyasviel) which fulfills its promise to allow one to “Focus on prompting and generating” - it is certainly easy to use! But shortly after its release, someone has “ported” the code to ComfyUI as a Custom Node! So of course it’s time to test it out...

About Fooocus

Among the many enhancements, Fooocus implements a “native refiner swap inside one single k-sampler... Fooocus uses its own advanced k-diffusion sampling that ensures seamless, native, and continuous swap in a refiner setup.” Compare this to using two sampler nodes in ComfyUI.

Install ComfyUI_Fooocus_KSampler

I noticed a post by GerardP19 on Reddit, entitled “My new workflow utilizing the Foocus sampler in Comfy, with the styler and nested nodes. ” but he did not provide a link to his GitHub repo, instead requiring ComfyUI-Manager to install.

Alternatively:

• Download ComfyUI_Fooocus_KSampler.
• Unzip and place the folder in ComfyUI\custom_nodes.
• Start (or re-start) ComfyUI.
• Create a workflow to test...

Error! Cannot import ...\ComfyUI\custom_nodes\ComfyUI_Fooocus_KSampler-main module for custom nodes: cannot import name 'load_additional_models' from 'comfy.sample'

• Edit the file ComfyUI_Fooocus_KSampler-main\sampler\Fooocus\core.py
• Search for load_additional_models.
• Compare with similar function in ComfyUI get_additional_models .
• Search for all instances of load_additional_models to replace with get_additional_models:
  • replace in from comfy.sample import prepare_mask, broadcast_cond, load_additional_models, cleanup_additional_models
  • replace two instances of models = load_additional_models(positive, negative, model.model_dtype()) with models = get_additional_models(positive, negative)
• Re-start ComfyUI and re-test.

Error! RuntimeError: [enforce fail at ..\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate ###### bytes.

I encountered this previously, but I just worked around it by splitting my workflow into two discrete steps (Base and Refiner), though sometimes re-running the workflow managed to squeeze everything in memory since the nodes cache outputs. Just wrongly assumed I did not have enough VRAM and mentioned this in a previous post, but that’s not the root cause!

I now realize it’s because I have too little system RAM, only 16GB! One solution is to buy more... but for now, increasing the Windows virtual memory paging file size allows me to continue:

• From the Start Menu, type View advanced system settings.
• In the System Properties panel, navigate to the Advanced tab > Performance Settings...
• In the Performance Options panel, again navigate to the Advanced tab
• Under Virtual Memory, click Change...
• Disable Automatically manage paging file size for all drive and manually specify a Custom size: I set a range from 18000 MB to 30000 GB
• Click Set to save, then back out of the panels by clicking the OK buttons.

No-Code Workflow

At minimum:

• A CheckpointLoaderSimple node to load SDXL Base.
• A CLIPTextEncodeSDXL and a CLIPTextEncode node for the positive and negative prompts.
• A CheckpointLoaderSimple node to load SDXL Refiner.
• A CLIPTextEncodeSDXLRefiner and a CLIPTextEncode for the refiner_positive and refiner_negative prompts respectively.
• A EmptyLatentImage specifying the image size consistent with the previous CLIP nodes.
• Wire up everything required to a single KSampler With Refiner (Fooocus) node - this is so much neater!
• And finally, wire up the latent output to a VAEDecode node followed by a SameImage node, as usual.

For the curious, prompt credit goes to masslevel who shared “Some of my SDXL experiments with prompts” on Reddit. And as a teaser... coming up, SDXL ControlNet for OpenPose (v2) !