# Krea2 TextFusion Refusal-Reduction LoRA
This is a dedicated rank-64 LoRA trained with one narrow objective: reduce Krea2’s learned refusal and restriction behavior while preserving as much of the base model’s existing visual knowledge and prompt behavior as possible.
This is not a concept, character, style, anatomy, or aesthetic LoRA. It was not trained to add new visual concepts, reproduce a training-image set, or impose a preferred look. Its training objective was exclusively to push the model away from refusal behavior so that concepts already represented within the base model are less likely to be suppressed during text conditioning.
## What It Changes
Krea2 receives multiple hidden-state taps from its Qwen-VL text encoder and processes them through TextFusion before sending the resulting conditioning into the image transformer. This LoRA applies learned low-rank residuals only to the attention and internal MLP projections within:
- txtfusion.layerwise_blocks.0
- txtfusion.layerwise_blocks.1
- txtfusion.refiner_blocks.0
- txtfusion.refiner_blocks.1
The release version deliberately contains no adapter for:
- The TextFusion 1 × 12 tap projector
- The external/general txtmlp
- The image-transformer blocks or any other image-generation layers
## How It Works
The base checkpoint is never overwritten. For every targeted linear layer, LoRA adds a learned rank-64 residual to the original weight during inference:
W_effective = W_base + strength × ΔW
This changes how existing Qwen-VL text features are routed, gated, transported, and refined through TextFusion before they reach Krea2’s untouched image transformer. In other words, the LoRA is intended to improve access to visual knowledge already present in the base model, not inject a new concept or replace the model’s learned visual representations.
This release targets the specific TextFusion route isolated through layer-by-layer ablation instead of broadly amplifying activations or directly altering the projector.
## Usage
Use strength of "1.00"
Responsibility and Output Disclaimer
This LoRA modifies the model’s text-conditioning behavior but does not determine, authorize, endorse, supervise, or control the images produced by individual users, prompts, workflows, checkpoints, samplers, or third-party software.
Generated outputs remain dependent on the base model, user-provided conditioning, inference configuration, and the surrounding generation pipeline. The author assumes no responsibility or liability for what users choose to generate, how generated material is used, or whether any output complies with applicable laws, platform rules, licensing terms, or third-party rights.
Users are solely responsible for their prompts, generated outputs, distribution decisions, and use of this LoRA.
Description
v.1.0.0 Krea 2 Turbo
FAQ
Comments (45)
can you explain how yours is different than several other loras that do the same thing? is your better and degrades the image quality less?
The main cause for degradation was using the projection layer as that alters the model’s understanding and the learned projection. This lora does not target that at all.
Compliments. It does a very good job. In fact this is the best I have tested and I tested all methods released until now. Thanks!
plz share comparation
@LuringSuccubus Just try it.
one lora that i really want to someone to do is for text attention, even in 2k struggles with text, seems like the krea2 like to focus on quality and other stronger stuff and reduce the importance of text
It is very good at text from my first test. I posted a example, but it doesn't show up bellow for some reason.
@boxthrowaway2026 my credit card was blocked so i can't train on runpod anymore lmao, + local training by ostris is being anoying, because i can't train on my rtx 4060 ti.
Will this affect NSFW too or is it just for blood/violence or worse? I never quit understood these kinda of loras tbh.
Incredible.
Not only does it de-censor the model completely and make it follow the prompt far better, it also exaggerates the effect of the SeedVarianceEnhancer node in Comfy, making the model to render more variations of a given prompt. It also seems to work fine in tandem with NegPIP. Superb.
PS1. Celebs still working fine.
Amazing work, thanks for training this. I hated how normalized using the shitty bypass loras were. They destroyed character likeness and coherency for more complex prompts. No matter how surgical you try to get arbitrary weight fuckery is a dead end if you want quality. This doesn't work 100% of the time for everything, but when it works the results are better than any bypass crap made so far.
You're the real deal. Great work and way better than all the other LoRAs!
I'm not sure how to use this. Does this affect the prompt enhancer, or something even deeper than that? I'm guessing it's not the prompt enhancer, since the lora doesn't come into play before that point in the official workflow.
This changes everything! Thank you so much!
Is there any way this could be applied on the Flux.2 9B model? That would be awesome!
it even enhances my lora. some prompts now work. For example "gun pointing at the back of person's head" was avoided all the time and now works. Or, she has a pussy in her face was not properlx possible... this is a new Krea2 era for me.
No, completely different models without mentioning that Krea2 has a lot anatomical details baked in that Flux2 9B doesn't.
No, flux2klein was trained on a curated dataset; meaning it does not have a safety filter but has more of a clean curated non nsfw dataset
@StanleyPain thanks!
@Capitan01R understandable
Best one by far. No more bad images from decensor and it doesn't mess up character loras. Awesome job!
amazing job. its by far the best one i tried, thank you so much for sharing it with us <3
Do you think it is advised to be used in combination with Krea2T-Enhancer node? And if so where in the pipeline the node should be placed?
Diffusion Model -> Refusal Reduction LoRa -> ComfyUI-Krea2T-Enhancer -> KSampler
or
Diffusion Model -> ComfyUI-Krea2T-Enhancer -> Refusal Reduction LoRa -> KSampler
Do not use the "ComfyUI-Krea2T-Enhancer" node because it reduces quality or, in some cases, results in a black images in some prompts. This happened to me even though I used the default settings in this node. Use only this Lora; I have tested it and it works perfectly, and the model follows my prompts precisely.
I understand that this Laura is doing exactly what the author's Laura is doing - https://civitai.red/models/2742336/nsfw-krea2-low-vram?modelVersionId=3089248 ?
Not even close lol
@RandoMando Great work! What makes this Low VRAM version stand out is the author's clear technical explanation. They explicitly stated that it does not affect the DiT/UNET generation layers, detailing exactly what was stripped. This is a huge improvement over the 'Refusal-Reduction' LoRA, where the author didn't provide a clear or competent explanation of what their LoRA actually does to the model's architecture. This level of transparency is exactly what the community needs.
@soyv4 it's great work no doubt but training the dit layers alongside of the textfusion then stripping off the textfusion layers as standalone is not the same as training solely the textfusion path which I mentioned here, as the model will still gain style/concept bias from the already bundled DiT/Unet blocks that was paired during the training. What I am offering with this LoRA is purly what the model already had without introducing bias. I hope that can make sense :)
@Capitan01R Understood, thank you for clarifying that! It makes perfect sense now.
What strength should the LoRA be at, and also what position should it be on the rgthree Power Lora Loader? And what strength should the krea2_raw_to_turbo_r256 be? I'm using both on the Krea 2 Raw INT8 Convrot model.
If this is anything like the many safety bypass loras floating around, I tested it and it is by a wide margin the best of the lot. It has absolutely no effect on image quality aka does not degrade image quality at all.
I used a weight between 0.1 and 1. Anything higher and it deforms the image. Currently using it at 1 and satisfied with the outputs.
Amazing. It looks like you fixed Krea. THANK YOU!
This is crazy good. This is the best way I've seen so far to decensor without lobotomizing the model or introducing style & content drift. Thanks!
it just works, love it
having no luck, are you using at strength 1? with which turbo model?
@rb_29 yeah, at 1 with the official turbo model (https://civitai.red/models/2732656/krea-2-turbo?modelVersionId=3072332)
Does this make the Krea 2 Conditioning node redundant or should we still be using it in parallel with this?
yes, you should not use any conditioning node with this LoRA, matter of fact I don't recommend using any conditioning node since it alters the Model's projection and provides amplified results that basically brute forces the model which causes hallucinations and also it becomes seed dependent
@Capitan01R What about Krea2 Text Enhancement you have on git?
@eejyvrmiqtqekxnzvp501 same for that node, it was my initial experimental method but now this LoRA replacing that
It appears that a penis mosaic is randomly applied during vaginal insertion (with a 50% probability); I have posted an image showing the mosaic. (Testing confirms this applies across the board—realistic, anime line art, and 2.5D styles all feature the mosaic.)
this is due to the model's internal training as this is not a safety layer but rather the dataset Krea2 was trained on so you can get such photos, the model was trained on billions of photos so some may appear in such way if the prompt was not specific
Thanx! This is a must have :D
I thought I was going crazy, why was the reddit post on this removed? (Fuck reddit all day)