First test of the Krea2 Turbo INT4 model NEW 6.9 GB NVIDIA RTX 20 and later versions — small and fast on older graphics cards.
Model merger using a short script created by ChatGPT
template to use for the merge https://huggingface.co/Hippotes/Krea-2-Experiments/blob/main/krea2_turbo_int4-convrot-simple.safetensors
parametres generation
euler simple er sde simple __8 steps 1080*1440_1024*1536_972*1296_ 896*1152
Thanks to https://civarchive.com/user/PrincesOfDarkness for the buzz
Description
FAQ
Comments (5)
You have no idea how much I like this checkpoint.
---------------------------
+
It perfectly matches the description
---------------------------
+ generation speed
Thank you very much, especially as this is a test run with Int4.
@Vince_AI I could go on at length about what’s good about it, but you only need to look at the samples.
In general nice but as a suggestion I'd request to provide additionaly NVFP4 or Q4 GGUF. INT4 lacks of details and sometimes it made a wrong impression of the checkpoint (it's the worst quant on my taste among the 4-bits, but as fast as old good SDXL). As 8gigs VRAM owner I use 4bit TE so output becomes flat (from both quantization and TE).
Donated some yellows because it was nice to find something quantized into INT4 and even working with Forge. Pity that modern quantization tools make checkpoints not working with it.
Thanks for the feedback; I’ll take your comments on board for future models. This is my first time using INT4 – I usually work with FP8 as I don’t have a top-of-the-range graphics card. Have a good day.