We have engineered checkpoints that can output almost any compositon. Let's take a step back.
This model merges NSFW Aesthetics V1 lora into Fineporn V4 and resets txtfusion layers to default krea2 turbo which has a drastic effect (it basically nullifies any decensoring), making it less capable of hardcore porn/amateur style photos and more oriented towards hyperrealistic professional photography and a cleaner slate you can add your own uncensor patches to. I use it with NEG for forcing greater realism.
CFG 1
8 steps
Sampler: SDE_ER
Scheduler: Beta or Bong_Tangent (maybe)
Prompt Advice: Use "Pornograhic", "Explicit", "Uncensored", in the begining of the prompt for nsfw contexts. If you need anything more explicit, use for example Donut Krea2 Fusion Control which supports most uncensoring methods (recommend the donut-made presets), or LoRAs. All preview images was made with no uncensoring or loras applied.
My kofi (any amount helps with AI bills): https://ko-fi.com/donutsdelivery
Discord available for posting resources and help: https://discord.gg/MWqMaxfkRs
Description
initial release.
The preview images are stolen prompt from other popular images to compare results, they were not upscaled.
FAQ
Comments (21)
Would you be so kind as to upload a full bf16? The quant that calls itself fp16 but is mostly 8-bit comes out pretty blurry for me.
needs Lora to be extracted from this TURBO model.
you can just take the recipe in the checkpoint metadata and apply it to your own model. extracting loras like this can degrade more than saving as a checkpoint.
BF16 version please
Great you seeing you making a Krea2 model! Thank you
Any plans for Q4/NVFP4/W4A8 format?
You need all of those? XD
@DonutsDelivery NVFP4 pretty please with sugar on top.
@DonutsDelivery any you can provide, slash stands for "or" :)
i tried 2 custom nodes that can quantize bf16 to NVFP4 and both of them output a broken model.
@gurilagardnr hell no nvfp4 is 🗑️🚮
@DonutsDelivery did you tried with ggufy?
@whateverr you cant just say that and not say a better alternative xD
@DonutsDelivery almost any other quant is better than nvfp4 in terms of retaining the original quality while giving a significant speed-up. int8 or in8 convrot are the best options for this
glad you came to your senses and made a model lol
I still advise using the latest workflow instead. Loading the recipe yourself feels slightly higher quality, but this is a good demo.
Thanks, liked your SDXL's looking forward. I've been using res_2s bong_tangent, 5 steps (2s = 2 steps so like 10 "steps" using 5 steps), on most krea2 models. res_3s_ode provides better usually, but then you're doing 3 "steps" per step. These new samplers and schedulers are odd, but useful! "steps" and actual steps generally take ~the same amount of time, so there's no real time savings, but the outcome is still better than 10 steps in most other samplers.
I think the important distinction is that those samplers working better at 5 steps doesn’t necessarily mean the model benefits from 5 steps. It can just mean their sigma/noise progression is poorly matched to the turbo model at 8 steps, and changing the step count happens to compensate for that mismatch.
If the model was distilled for an 8-step trajectory, I’d rather use a scheduler like simple or beta that behaves properly over those 8 steps than change the step count to make a mismatched scheduler behave better. You may get an improvement over running that same sampler at 8 steps, or over Euler/simple at an arbitrary 5 steps, but you’re still making a compromise compared with sampling along the trajectory the model was actually trained for.