PLEASE READ THE VERSION NOTES
What this is?
A turbo checkpoint tuned for the perfect balance between three things: NSFW capability, prompt adherence, and a "smartphone-like" look. The goal is a balanced crisp, casual image. This is the goal, and this checkpoint is the starting point; expect many updates with high frequency. baked on top of the official turbo model.
About FinePorn
This is an approachable but functional starting point. To hit the target, I baked in LoRAs with balanced weights + INT8 and CONVROT to achieve faster speed and good quality, with the plan to scale up future versions by adding new LoRAs, rebalancing weights and hunting for the optimal settings.
Baked LoRAs Changelog (v2):
Realism Engine — reduced weight
Realistic Snapshot — reduced weight
Krea 2 NSFW V4 — removed
SNOFS V1.1 — added
RawGirlSpicy V2 — added
PureLens V1 — added
PhonePhoto — added
All baked LoRAs will credit their creators.
⚙️ Recommended Settings
V1
res_2s + beta -> More Quality
er_sde + beta -> More Speed
CFG: 1
Steps: 8–10
V2
SamplerER_SDE + beta -> Fast and Stable
CFG: 1
Steps: 8–12
📐 Resolutions
I found that scaling resolutions by x1.25 and x1.5 gives better results:
- 832x1216 standard → 1040x1520 optimal 1248x1824 recommended
- 896x1152 standard → 1120x1440 optimal 1344x1728 recommended
- 1024x1024 standard → 1280x1280 optimal 1536x1536 recommended
💡 Notes
Good results are achievable without a face detailer or extra passes. With the right config, some patience, and a bit of seed luck, you can get great things.
I highly recommend start all your prompts using a phrase like this or similar.
this is an amateur photo taken from smartphone, casual photo⚠️ Important disclaimer
This is NOT trained by me and is NOT my own work, my job here is using my experience to bake balanced weights of my personal taste to achieve desired images using my prompt style. My only role is to share with you what, to my eye, is the best style and quality achievable with this model. All baked LoRAs credit their creators.
Description
Baked LoRAs (v1):
- Realism Engine — credits: @razzz
- Realistic Snapshot — credits: @MonkeyForever
- Krea 2 NSFW V4 — credits: @19doorsside884
FAQ
Comments (45)
Only done a few renders so far but this is looking very nice! Question: I noticed in your examples the distinct LoRAs are listed even though they're embedded in the checkpoint (if I understand correctly). For your samples were those actually added as LoRAs, or are you tagging them here only for traceability / visibility?
Thanks! yes you see the loras because i'm using the same workflow to keep testing differente/new loras and balancing weights.
The three mentioned loras are already baked into the model so you don't need to use them, but they are baked with specific weights i chose, so i will be constantly updating this model to newer versions with newer and best combinations i find in the way
A soon as i posted this checkpoint, the realism engine v3 came up, so i'll update this with that latest lora and a few new small ones on the following days
@simonishere Based on my testing, those external LoRAs still seem to be doing most (if not all) of the work. You definitely did try some renders of just your checkpoint with the LoRAs disabled? I'll post a couple examples here shortly. One will be with the checkpoint by itself, and the other with those LoRAs enabled externally.
@simonishere Link here to comparison. Maybe I'm doing something wrong. I also tried disabling the filter bypass in case that was affecting chkpt efficacy: https://civitai.red/posts/29661943
@JJ_Goldstriker This model is intended to use without LoRas and in square or vertical res, the loras are already baked in the model so you don't have to enable them, you can simply disable the loras loader, it works only with checkpoint
@JJ_Goldstriker In the description i provided the workflow clean and fixed, maybe you picked one of the images, wich was a mess, all the details are in the description
@JJ_Goldstriker also try updating comfyui, the int8 native support is recent
@simonishere Yes, my friend, I understand the point is external LoRAs shouldn't need to be loaded -- My point is that's not what I'm seeing in independent testing. Yes, running latest CUI. All good, I may be doing something wrong, no problem -- If not, I figure we'll see others commenting similarly soon, unless they have the external LoRAs enabled as in your examples. Cheers and thanks for your contribution to the community.
@JJ_Goldstriker i'll be paying attention to this, in a few days i'm uploading v2, let's see if something changes for you :(
thanks ! thanks for your efforts: PS Only releasing in format of INT8+Convrot_ means you are shutting out all mac users from using your great model as far as I am aware.. - FP8 at least there is a mac patch.
tbh i didn't think about that, thanks, i'll see it on v2
@simonishere I can confirm that INT8 models run much slower on my 3060Ti, latest ComfyUI on Windows-standalone. Someone had explained RTX30x and even 40x don't have true hardware f or the INT8 so it runs emulated, not sure if true or if ComfyUI still needs to be worked on for proper support, I really don't think it's my setup, unless I actually have to get Sageattention going, which means I have to downgrade my Pytorch or build Sageattention from source.
@AI_DK I have 3090 and 64 RAM, fp8 runs at 1 step/2s and int8 runs at 1 step/1s so it's a 100% speedup, int8 is SPECIFICALLY designed for amphere RTX (series 3000), works on other series but brights on series 3000, i don't know why u get it slower, you are the first case i see having that trouble, should be a config problem
@AI_DK https://share.google/aimode/BTFrry709OJDdTlp4 please let's avoid spreading misinformation
Work of art checkpoint- 1m in rtx 3060ti 8gb.
i'm glad to hear that!
@simonishere you're a genius. 10/10 Custom character lora without int8 Works flawlessly.
Try fp8 and tell me how fast it runs compared to an FP8. for me, int8 is slower.
And what do you mean, 1m? Hope you don't mean a minute. No LoRA FP8 on my 3060Ti 8GB(is there 16gb?), it's 30seconds..? INT8 could get close to 1min with . Witha bunch of LoRA on FP8, it could get to 48seconds(per prompt change, after LoRA is "loaded")
this v1 int8 (as said) , it can work with CPU only ?
as far as i know, it works but expect long generation times
Hey Im curious! I use the Raw model + Turbo Lora method since it gives better base seed variance. Because of that I always struggle to try out these Turbo style checkpoints because they dont fit into my workflow of using the Krea2 Raw model.
Maybe this is a dumb question, but is there anyway to train and release a Krea2 Raw checkpoint version of the model that you made here?
I would also love to hear others input who also have this same problem.
Hi bud, techically i CAN just download the raw model and make exactly the same bake i did with this turbo model, so the answer is yes, but i'm not sure about the part of the model behaving like i would expect, because i bake specific lora weight on the turbo model wich fits my 'quality checks', i have to test it, maybe in a further version i'll also include the raw one :) thanks for your comment
@simonishere Of course! Can I pass along something else also? I did some detailed research over the days since Krea2 came out and found what I think is the best workflow for realism. Right now for Krea2 'Power Users' the best way to get maximum quality is to run Krea2 Raw model through a Sampler then finish it using the Raw model + Turbo Lora.
I ended up finding an even more specific one that uses the Clownshark Sampler.
I found it off of a Reddit post originally. I ended up tweaking it a bit for Quality of life ease, but the base workflow here is made for maximizing realism. It seems to improve the facial expressions, realism, and just overall seed variance.
Overall takes ~40 seconds on my 4090 w/ 2 extra loras added and the base raw Krea2 model
@mrgenman6969 ty i'll check it
I think this is the best Krea2 NSFW model released so far. 😃
Nah haha, but thanks for the enthusiasm ❤️
Which node should be used to load the model?
The official difussion loader, comfy already has native support, so load it as every model
Fantastic for amateur, candid, snapshots.
Works great for complicated multi subject prompt adherence. A fine model, indeed.
really happy to hear that, i'll keep this in my to avoid breaking it in the next version
Thanks for your input. Sorry for the silly question, as I'm a newbie, but this works with a character lore I made for Krea 2. How do I configure it so my character lore matches your NSFW lore while maintaining consistency? If you have a workflow, that would be helpful.
Hi friend, i haven't tested this using character LoRas so it's up to you, you can download my workflow attached on the optional files, and add your character LoRa, but if you want to maintain more consistency, help the lora using precise description of your character, so the model will be guided better.
Also i noted Krea works better with order in te prompts and specific blocks, so i recommend you to start mentioning the camera angle+distance (ex: medium shot, viewed from the side in a three quarter angle...), then describing the character of you lora, then write the pose/action and details. leave the background or setting for the end, it is easy to build, be specific, and avoid ambiguous phrases, all of it will help the model to keep your character consistency
@simonishere very helpful, thanks
works in wan2gp but NOT with any loras. just gets static pixels garbage. any ideas?
How can I get rid of the cum and water spots that keep appearing in generated NSFW images? They show up very frequently, and even when I upscale the image, they remain and produce weird-looking results.
I'm pretty sure this issue is coming from one of the baked LoRAs because I get the same result even when using the original model with the bypassed LoRAs. I've also tried adjusting the LoRA weights and changing the prompts extensively, but the water spots still appear in many generations.
i'm sure you are using a word related to liquid/fluids, try to rebuild your prompt rom scratch because it's weird, are you using er_sde beta? i think there are some words related to nsfw that spawns liquid, check it
@simonishere I haven't used any of those words. I also tried negative prompts like "no wet" and "no water spots," but they didn't help.
I'm using ER_SDE Beta with your model and Euler Simple with the original model. What I've noticed is that this issue appears much more often with certain types of generations, especially scenes involving penetration or oral
@Not_SALAM interesting, i'm releasing v2 today, i hope it gets solved, i did many changes adding new loras, removing some and changing weights, i'm also going to update my workflow with another sampler node (also native i guess) wich is experimental but i think is giving me better results
@Not_SALAM I've had this kind of issue with krea 2 since I started using it. Negative prompts might be better if you say "wet" "water spots" because they're negatives - saying no wet makes it a double negative. Not sure, but worth considering.
@mrpras I used those trigger words in the positive prompt because, as far as I know, Krea 2 doesn't really understand negative prompts. Even though there's a negative prompt node in the workflow, adding those words there doesn't seem to have any effect.
This is exactly how model releases should be documented. I wish everyone did it the way you did
I am gonna be honest , I did not try your model yet ,, but since you said you are going to release the V2 today , i will wait for it and try it ,, and give you my honest opinion .
Thank you for sharing your work with us . Have a nice day
thanks friend! just updated to v2 :)