(Links to files are provided at the bottom of the description.)
v2 - adds Prompt cycler to autogenerate images with a text file list of prompts.
v3 -adds SeedVR2 to upscale images to 4K resolution while maintaining ALL the detail quality.
This is a condensed GGUF version of the ComfyUI template workflow for FLUX 2, designed with LOW VRAM USERS in mind.
replaces the diffusion model with GGUFs. optional GGUF text encoder as well for extremely low vram users. (for some reason the fp8 works just as well as the fp16 with WAY less resource use. BE WARNED the text encoder is HUGE, around 18gb for the fp8. for some reason it runs easily on my 8gb vram card with 16gb of regular ram, i think that is because the text encoder and gguf weights arent stored on the vram card at the same time so your either just running sampling steps or encoding it but never at the same time so it should run well on most pcs.)
IMPORTANT!!! :::
FluxGuidance node is tricky, for regular images just use 4.0 . for TEXT you may need to raise the value to 11.0 or 8.0 if artifacting is too strong.
FOR THOSE USING "cow_mistral" text encoder either install missing custom nodes or use the other text encoder file. its small and can fit on most cpus by default!
❗️IMPORTANT FOR PROMPT CYCLER IN VERSION 2❗️
Copy the address of the text file and paste it into the simple string node. REMOVE QUOTATIONS FROM BEGINNING AND END OF FILE ADDRESS OR THE CYCLER WONT RECOGNIZE THE FILE. Then feed the connector node into the text of the clip text encode node of your image generator, NOT THE NODE ITSELF. (Theres a smaller GREEN connector dot on the left side of the box of the clip text encode node, essentially your feeding the string into the box and not the clip nodes connector.)
YOUTUBE VIDEO TUTORIAL FOR PROMPT CYCLER (using Z-Image-Turbo as model, workflow is interchangable so any image model can be run with it, including FLUX 1,FLUX 2, QWEN Image Edit 2509)
LINKS:
FLUX 2 GGUF
https://huggingface.co/orabazes/FLUX.2-dev-GGUF/tree/main
Vae
https://huggingface.co/Comfy-Org/flux2-dev/tree/main/split_files/vae
VAE GGUF
https://huggingface.co/chatpig/flux2-dev-gguf/tree/main
Text Encoder (fp8 or fp16)
https://huggingface.co/Comfy-Org/flux2-dev/tree/main/split_files/text_encoders
Description
version 2 includes a prompt cycler for AUTO-GENERATING images from a text file without the need for inputting single prompts 1 by 1. extremely efficient for overnight generations.
FAQ
Comments (34)
How low on VRAM are we speaking? What would be the lowest config the 'worst' GGUF could run on? Pretty impressive the GGUF versions are already existing.
Im running the Q2 with absolutely no issues but im going to test the Q4 next. I only have 8gb vram and 16gb ram so the fact that the q2 even runs on my specs is weird but it runs very clean! I think theres some offloading or something happening onto cpu somewhere because this should crash my pc.
@realrebelai Thank you. I might give it a try then, looks promising.
@TheGlowingGuardian no problem brother im here to help, just know your walking into about 30gb of resources in total download and run
wow, this flux 2 came at the perfect time, just the yesterday i was messing around with FFGO, and now this.
Thank you for the dono bro! ❤️ glad i could get this out early to you all!
Change the Base Model, should be Flux.2 D
It wasnt available when i posted the workflow and im not sure i can actually change it now that its posted sadly 😔 i got to it a little TOO early 😅
GGUF Text Encoder (cow-mistral3-small-q2_0.gguf) is giving me this error: “Unexpected text model architecture type in GGUF file: ‘cow’”
Edit: I figured it out. There were two GGUF nodes. One (https://github.com/city96/ComfyUI-GGUF) gave this error. One (https://github.com/calcuis/gguf) worked.
The file may be broken, try the other option and see if that gguf version works. Ill look into this later for you when im off work!
You can fix that by loading the workflow from the huggingface repo of that gguf (it's on the bottom) and installing the missing nodes. The clip loader in this workflow does not support that architecture.
@Ohmlaash It obviously worked for the OP though so that may not be necessary.
@Ohmlaash probably should add that i updated the comfyui before creating this workflow, its the native worlflow FOR flux 2 so its just an update issue
@realrebelai Cow needs calcuis' gguf nodes
@qek install them or use the other sized text encoder.
Thanks for this.
Not sure what "pig_flux2_vae_fp32-f16.gguf" is but I couldn't find it. Also, why are we putting it in the text-encoder path? It seems to be a unet. ??? It does not seem to be needed. ???
It's a small mistake, it's a vae as in the name, it is not needed, only if you want to use it for that specific workflow of that repo the file is in.
@Ohmlaash No worries! It certainly works without it. Nice workflow!
you can find the vae you mentioned on pigs HF site. the link for the text encoder gguf is where you can find it
my mistake, i corrected the description and added the correct text encoder GGUF <3
WORKFLOW HAS BEEN UPDATED TO FIX VAE ISSUE! TRY AGAIN <3
FOR THOSE USING "cow_mistral" either install custom nodes missing and implement the gguf vae loader or use the other vae file.
Which node can load qwen3 gguf textencoder?
Tried type flux2 and qwen_image but both gives a math error
@realrebelai fresh installed comfyui today...
@zoom83 im not sure why your using the qwen text encoder on a Flux workflow. thats literally your issue right there bud... didnt realize it until just now but yeah your using the wrong file.
@realrebelai I think it's because the links direct him towards that.
@yorgash that would be because im a complete idiot and put that in the wrong workflow 🤣 my apologies ill correct that immediately.
@zoom83 completely my mistake, that was for another workflow. Ive corrected the issue for you. My apologies!
@realrebelai No prob, I could easily decipher qwen is not mistral after all :)
well done.
better than comfyui workflow!
Uselessly. I used Flux2-dev-q6_k.gguf 25GB, 4090 24VRAM and 64 RAM, I couldn't wait, it's an awfully long time, I interrupted the process. I'll try Q5 and Q4 again. Upd: I tried Q4, the user sampler is infinitely long, I did not wait and interrupted the process.
Yeah it takes about 20-40 minutes for 1 image on my rtx 3070 lol
it works, BUT never said it worked GREAT 😅🤣❤️
@realrebelai I get it, but why mock yourself?))
@AlG80 i think your missing the entire point of my page buddy. this is a LOW VRAM workflow. its meant to give people with terrible specs a chance to use the model. its not meant to be fast at all. none of my workflows are meant to be fast. they are just meant to WORK. most people cant run this model but with my workflow, it can generate. really not that hard to understand, your simply just upset its not generating images for you in 30 seconds. respectfully not my problem.