Since people can't read. Strength can be used at 1.0 all the way up to 40000.0
If you use poorly quantized model then expect black image as a result if you raise strength too much.
I thought that CivitAI was gonna include the strength you set when uploading in the model page but apparently not.
This is original modifier LoRA for bypassing the detrimental refusal filter present in Krea 2.
It is very small because it has been tested thoroughly and restricted to target the vectors that are most detrimental.
Strength usage is liberal as it is almost impossible to completely overdo it. i.e. you can use anything from 1.0 to 40000.0 if you so wish to do so.
The intention of this bypass is to make the model respond better to style and emotion prompting as it tends to suffer in those areas.
Description
This modifies 3 vectors of the affected layer.
FAQ
Comments (50)
The Krea 2 model apparently doesn't understand the word “naked.” Whether I enter 1 or 7, it's all the same. You just have to use euphemisms.
It does understand nude. You have to be a bit more specific, though, like "Her upper body is nude. Her lower body is nude."
You need to use the word "nude", then it'll work fine, no euphemisms needed.
Doe this replace the Conditioning Krea2 Rebalance node?
https://github.com/nova452/ComfyUI-ConditioningKrea2Rebalance
yes. the conditioning node destroys quality and shouldnt be used
no. performs worse in my tests
@hopinghortons139 blabbing "performs worse in my tests" without presenting any tests...
This worked better than the conditioning node in my opinion. You can test this lora too "https://huggingface.co/Beinsezii/Krea-2-Turbo-Projector-Scale-LoRA-Diffusers/blob/main/pytorch_lora_weights.safetensors"
that node is just a day1 quick patch, its useless
is this usable for krea 2 turbo?
yes, think of the "turbo" as "inference", and "raw" as "base". Although it doesn't do a whole lot, it might make your outputs more crisp. At least it did for mine
do 13
Unfortunately, your LoRA performs quite poorly compared to other options published on Reddit; for instance, the z0jglf LoRA completely resolves the censorship issue. So, in my case, your LoRA turned out to be completely ineffective and useless.
agreed
that doesn't appear to exist, at least searches using "z0jglf" and other "krea" "lora" searches didn't pull up anything obvious.
edit: Found the reference: https://www.reddit.com/r/StableDiffusion/comments/1udhaio/this_custom_node_removes_the_builtin_krea_2/otd05jw/
Edit: At 1.0 strength, the effects the loras add is pretty minimal -- however, I would say I find the output of OP's lora to be the most crisp. It also didn't add any unprompted effects to skin, which the reddit lora did.
brother they do the same thing lmao
Did you read description?
@makiaeveli strange lora 0.1 not 1
"z0jglf LoRa"
damnit, we're getting Krea2 Lore now?
I really don't understand such an explanation.
When you mention a Strength range of 1 - 40000, people will find that the maximum value allowed by the LORA loader in the comfyui is only 100. Then I began to doubt whether I had made a mistake somewhere.
People use all kinds of custom nodes, so while the default native solution lora loader don't have over 100, other custom nodes might.
turbo
it started working at 13 :)
https://huggingface.co/lilcheaty/Krea2-INT8-ConvRot
max I got was 4.61622 for my Krea2-Turbo-int8-ConvRot.safetensors before it started giving me a colorful screen of doom.
46000%
what node are you using to load the model?
mine goes up to 100 and it doesn't change anything, the filter is still there
@ZelashZelash ComfyUI-INT8-Fast
https://github.com/BobJohnson24/ComfyUI-INT8-Fast/tree/main
@Docuei sorry i meant to load the lora model.
@ZelashZelash It comes with it.
@Docuei the int8 grouped lora node? but it goes up to 10 strenght
okay nevermind, i was using another base mode, i was using Krea2_Turbo_convrot_int8mixed.safetensors
the lora doesn't work with that one, but with the one you posted it works perfectly at 4,86 strenght. thanks!
@ZelashZelash good to hear!
I tried Krea2-Raw-int8-ConvRot with turbo lora, it's giving me washed out image..
@shab987 hmm I'm using it right now on Raw just fine.
"49": { "inputs": { "seed": 362892245, "seed": 840763653, "seed": 868295058, "seed": 928685681, "998x626": 292104679, "seed": 832550256, "seed": 113377, "seed": 34567, "steps": 6, "cfg": 1.0, "comment": "Reddit", "scheduler": "ddim_uniform", "sampler_name": "dpmpp_sde", "comment": "bad for everything else", "sampler_name": "res_2s", "scheduler": "bong_tangent", "comment": "c.", "scheduler": "simple", "sampler_name": "er_sde", "comment": "style, somewhat bad at action.", "scheduler": "simple", "sampler_name": "euler", "denoise": 1, "model": [ "177", 0 ], "positive": [ "3", 0 ], "negative": [ "4", 0 ], "latent_image": [ "5", 0 ] }, "class_type": "KSampler", "_meta": { "title": "KSampler" } },
"166": { "inputs": { "unet_name": "Krea2-Turbo-int8-ConvRot.safetensors", "unet_name": "Krea2-Raw-int8-ConvRot.safetensors", "weight_dtype": "default", "model_type": "krea2", "on_the_fly_quantization": false, "enable_convrot": true, "lora_mode": "None" }, "class_type": "OTUNetLoaderW8A8", "_meta": { "title": "Load Diffusion Model INT8 (W8A8)" } },
"177": { "inputs": { "lora_1": "krea2/krea2filterbypass3.safetensors", "strength_1": 3.5, "lora_2": "krea2/krea2_turbo_lora_rank_64_bf16.safetensors", "strength_2": 1, "lora_3": "None", "strength_3": 1, "lora_4": "None", "strength_4": 1, "lora_5": "None", "strength_5": 1, "lora_6": "None", "strength_6": 1, "lora_7": "None", "strength_7": 1, "lora_8": "None", "strength_8": 1, "lora_9": "None", "strength_9": 1, "lora_10": "None", "strength_10": 1, "model": [ "166", 0 ] }, "class_type": "INT8GroupedLora", "_meta": { "title": "INT8 Grouped LoRA" } }
I get this error when trying to run that checkpoint you linked; Unknown quantization format for layer blocks.0.attn.wq. not sure what's wrong as comfyui is fully up to date. 0.26.0, and other krea2 models have worked fine. Is there a something I need to download? I downloaded the 'https://github.com/BobJohnson24/ComfyUI-INT8-Fast/tree/main' and added it to custom nodes folder... still get the same error message
@Docuei I used prelora, thats why its giving me weird output. Anyway,it's not faster with loras attached, im back to bf16.
@Light7799 make an issue on their page, they might be able to fix it.
@shab987 does BF16 have faster startup speeds per image? Kind of getting tired of waiting 3-4 seconds before generation starts.
@Docuei everything is faster on bf16 for me. I just did I test, unloaded all model. bf16 with 3loras 12steps is 35s with each step 2.5s once its normalized. int8 with 3loras 12steps is 44s with each step takes 3.3s+ once its normalized.
I'm using 3080 10gb+96gb+pcie5.0ssd.
INT8 is much faster at 1.3s/t ONLY without loras for me.
fp8_scaled is like ~5% faster than bf16 so im not going to use it.
I'm not encountering lagging 3~4s for starting to generate, first step is usually ~1s slower than other step, which is the same with int8/fp8/bf16.
@shab987 Thanks, I'm gonna make the switch!!!
@ZelashZelash im also stuck at 10 strength? using INT8 Grouped LoRA
@shab987
F, got an OOM. Forced to stay on INT8 -_-
genuinely just UNLOCKS the model, can generate literally whatever you can think of. I am using the rgthree lora stack loader with weights between 3-5 (sometimes higher needed for some really complex or degen prompts lol)
yeah 1 seems to work for basic stuff but I noticed other stuff wouldn't fully work does the image degrade at all when using it at higher strength ?
@Hzoid not to my knowledge, or at least enough for me to notice, I do feel prompt adherence is ever so slightly better at lower values so I aim to use the lowest possible value for the scene (simple nudity can be done at as little as 1 strength, intercourse requires closer to 5 to be properly good)
does it work for gguf models?? i tried this https://civitai.red/models/2727829/krea-2-enable-nsfw-prompt-adherence-krea2-nsfw?modelVersionId=3066310 but it gives black image for krea2 turbo q4_k_m gguf
the dev say this "If you use poorly quantized model then expect black image as a result if you raise strength too much"
@weskeralbert80s598 I used strength 1.0 still black image
@TaiLong the problem is the quantization, this lora work with bf16 and fp8 versions
Can you please do something for gguf krea2 models??
Incredible! Thank you
Hello, thanks for amazing lora and NSFW unlock. I have a question, does anyone have issues with it working in the latest update from comfy(nightly)?
Before it worked perfectly for me, but now i can't get it working anymore.
It used to work consistent, now it very much struggles.
I use the Krea2 Turbo FP8 from comfy, not some quant from unknown source. I use the basic standard load lora node and tried various strengths as well, even going up to maximum (100).
I want to know if something changed in the nightly release or if it just a "me" issue.