This is a very simple ComfyUI beginner friendly Text-to-Image workflow that will work with a single Flux GGUF model with multiple LORAs (This workflow does require ComfyUI's LORA Manager plugin to function. It is a good idea to install both ComfyUI manager and LORA Manager plugin to help you easily download and manage Checkpoints, LORAs and other resources. Not only these two are helpful for this workflow but they will help you a lot in any other cases). I also included all inclusive required model download scripts (for Windows and for Linux/Mac) with this so you can just get started creating your AI Art ASAP as these scripts download all essential core SD 1.5, SDXL 1.0, Flux GGUF and HiDream GGUF files for you. Make sure you install GGUF addon for ComfyUI using ComfyUI manager and place the correct files in correct places. Also check out my other workflows for SD 1.5 + SDXL 1.0 and HiDream.
How to use this -
#1. Just select your desired Flux GGUF model first and then load
#2. one or multiple matching LORA(s) from LORA Manager to start
#3. then input your prompt (The GGUF models for Flux does not accept negative so leave that empty).
#4. select how many images you want (Change the number besides the "Run" button)
#5. select image sampling methods, cfg, steps etc. settings
#6. finally press the run button to generate. That's it..
*** This workflow will also function if you know how to properly disable the LORA Node
and bypass it
Enjoy!
( Update - V2.0 with new tweaks now saves output images with timestamp, it's much easier to follow, generates better outputs and I also checked this version with Flux Krea GGUF, besides the old download scripts file I also included inside the new workflow archive file a text file with download links to all new models like WAN 2.2, Flux Krea, HiDream E etc I have been using that were not included on the scripts or require logging into HuggingFace and acceptance of terms of use like in cases of SD 3.5 and some other models.. check that one too )
Description
FAQ
Comments (4)
Is there a workflow or approach to optimize the issue of increasing VRAM fragmentation, which leads to progressively slower performance after prolonged use of ComfyUI? Currently, I can only restart my computer to resolve it.
I still haven't found such optimizations, I do have to restart my Intel laptop for some tasks but I prefer my Mac Mini for these tasks more because here CleanMyMac can free up VRAM easily when needed plus Mac is surprisingly more optimized even though I have same amount of RAM in both Mac and my Intel Laptop (16GB) on Mac the RAM acts as Unified memory.. with that in mind I am planning to ditch my Laptop and will go for a Minisforum AI X1 mini PC ( https://store.minisforum.com/products/minisforum-ai-x1?variant=46484012957941 ) with AMD Ryzen AI 9 365 processor.. this processor kinda brings that unified memory concept in which you can allocate say 32 GB out your 64 GB RAM to graphics processing and AI tasks.. folks have told me that this even beats some older midrange GPUs in AI tasks.
sarcastictofu Great knowledge, glad to know it. My Mac has been idle, so maybe I can use it for some experiments.
Launch with "--disable-smart-memory --cache-none --reserve-vram 8.4 --lowvram" (for 16GB VRAM). The coders of ComfyUI literally haven't the first idea about memory management, how windows automatic caching of files work, or how AI models for all but LLM need to stream, not to store in VRAM. This is for Windows, of course.
Oh, I'm too used to seeing wan2.2 t2i- for Flux there may be more of a hit than you like, but system stability is much improved. The other option (or I use it as well) is to put model clear nodes at key points in the workflow. When nothing works, it isn't 'fragmentation' but memory leaks due to bad code in comfy or a particular node!