Krea2 FinalCut (V1 Redux)
This model is using the V1 base with large amounts of PONY training it has some of the refinements of V3
FP16 training
Comfy Quant format is group size 64 - FP16 for sensitive blocks - using this format had a 20% speed increase over BF16 or FP32
Capable of 1536x2440 native
8-12 steps with highest quality at 12 steps sgm_uniform/er_sde
This model has been trained with WD14 tags and "1girl" to allow for safe pony style prompting on Krea2
Description
FAQ
Comments (7)
When I want to test a NSFW prompt, I reach for the model that best combines speed and fidelity. There are several other models that are just as fast, but among them, this one generally gives the best results or is second-best among the group, so this is generally the one I try first. (I specifically mean the ComfyQuants 8-bit variant.)
Thanks
@Felldude Sometimes I will spend all day rotating through the models, giving each one 4 attempts which get slower and slower. It does happen a fair bit that some of the slowest models justify it with increased quality, and I end up using one of those for my final work. But it also happens a fair bit that I slog through all of them only to find out this one is still the best.
@lesjo Yeah the 1.5 cfg can double render time and really slows down 2440px renders - sometimes forget to put it back to 1.0 or have like 1.05 and wonder why everything is so slow
@Felldude And to think I keep the CFG at 2 to 3... I'm living in slow mode anyhow (12 GB RTX 3060) but I'd much prefer if my negative prompts actually do something. That was always a limiting factor with FLUX.1 (CFG had to be exactly 1.) Raising the CFG with FLUX.2 is possible but negative prompts still seem to be completely ignored. I'm glad the move to Krea gave me back my negative prompt, so I won't be running at CFG 1.
It's not just the speed of each pass though (there are many models which have the exact same level of performance), it's that this one does a good job at 10 steps, which is only true of maybe half the other "fast" models, and of those others, only NoctUncensored comes close to the "hit rate" of this model.
@lesjo CFG at perfectly 1 will be far faster with this model if you have not tested it - assuming your using comfy and not a GUI that automatically zeros negative
@Felldude Yes, I'm using ComfyUI, and I can confirm what you are saying. At CFG 3 I get about 8 steps per minute (or about 7.5 seconds per step) at 1376x1024. At CFG 1 I get almost double that rate (15 steps per minute, or 4 seconds per step). In many cases, CFG of 1 is going to look the same, too, but CFG seems to matter more if using a character LoRA (or probably any LoRAs). In further testing, it seems like a CFG of 1 causes the model to ignore the character LoRA about half the time—which means there is no net gain in the speed of usable output, and twice as many images to wade through. If you've already nailed down what you're trying to create, then CFG of 1 may be a false bargain.