@KiraNugget i have additional folders, do i drop them in 1 of them or no?

@KiraNugget ok, and there's a llm prompter node that won't install or update.

is this error on my end?

@KiraNugget it works. Thanks for your help

I dropped from Q6_K to Q5_K_M 20.7GB (along with the 1GB mmproj gguf, display taking up maybe 1.5GB of VRAM, etc) and that seemingly fits well into 24GB VRAM. First run was 44 seconds, subsequent runs took 22 seconds each, so much faster when it all fits into VRAM. 24GB VRAM card users, I'd recommend using Q5_K_M 20.7GB version.

@KiraNugget Q5_K_M is giving me good descriptive paragraphs in 22 seconds with 24GB VRAM. I used Google Gemini to help me decide which to use, and I had it score Q6 vs Q5 vs the default IQ2 and it showed Q6_K is near lossless compared to the base model (97-98%), Q5_K_M (95-96%) Very High quality, virtually identical to Q6. IQ2 it says falls off (70-75%) "Falls completely off a cliff. At 2 bits, models suffer from severe token repetition, broken logic, and an inability to accurately parse complex visual details from mmproj" Gemini recommends Q4_K_M for 16GB VRAM GPUs. It's ~18GB plus 1GB for mmproj, so it will overflow VRAM into RAM and slow down, likely around 2 minutes to generate a prompt, but Q4_K_M it estimates would retain (93-94%) quality compared to the base model. If you're getting good results with IQ2 for this workflow and it's fast, that's fine. If there are any noticeable problems, try Q4_K_M for 16GB cards (expect slowdowns, but for better accuracy).

how did you check which one is best for my system? I have 8 gigs of vram

@KiraNugget i don't see any checkmarks

@Alkorr64 With 8GB VRAM, you'll be massively overflow VRAM into RAM and running much slower using any of these LLMs, which the lowest, IQ2, is 11.7GB, plus about 1GB for mmproj, plus you likely have your display hooked into your GPU taking up another 1.5 or 2GB VRAM or so.. You would need some much smaller LLM, which I'm sure there are options, but you wouldn't be getting as good results

In have it too i have manual instal, per manager it will not work :(

@KiraNugget I’ve tried it ten times, but it still doesn't work; it isn't being recognized. The manager keeps telling me to download the requirements—which I’ve already installed—but it still won't work. The video method didn't work either, unfortunately, since the new version isn't included there; it just keeps turning red and doing nothing.

@tomookazaki87856 The comfyUI manager links to a branch of the main LLM-text-processor node (labeled MOD) instead of the regular node. go to the github site (https://github.com/KingManiya/ComfyUI-LLM-text-processor) and follow the instructions to install the "main" node in your custom nodes folder. It's a ComyUI manager link error.

Please let me add one more thing which I just found:

1- Expand short concept only, all others disabled -> If I enable only 1st image, then Image Compare (RMBG) 🖼️🖼️ generates a black image -> resulting bad prompting as won't consider input image.

If I enable 2 or 3, then combines them correctly.

I'm not sure, but this may be RMBG node issue. If you can reproduce it, there are multiple image combining nodes out there, including built-in Stitch Images which may work.

Thank you again.