# 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 (89)
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
@Capitan01R 🙏 OK~thanks
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)
was wondering the same, I'm glad he hasn't completely vanished
Post was at stable_diffussion subreddit? It should be at unstable_diffusion (nsfw)
@flo11ok874 there needs to be a better discussion subreddit that allows NSFW. Unstable diffusion and all the other NSFW ai subs are all just people sharing or trying to sell their porn collection with minimal discussion on technique, prompts or tools.
@DrawingBard The is a DegenDiffusion subreddit since a couple of days ago
@flo11ok874 There is nothing about this that makes it NSFW by default. It's IDEAL for prompt adherence regardless of what you're doing.
This is great, game changer. Thanks!
Approved
Could someone give an example prompt where this LoRA shines over the others like filterbypass3? I'm not challenging the praise here, clearly others are seeing something I'm not, I just haven't been able to see it in my own testing and would like to. Is it like violence/gore-specific or something? Also it seems all the examples aren't even things that would be censored anyway, so I'm doubly confused.
The author is actually just not allowing any NSFW content to be shown on the generated examples viewed on his page that’s all. But hands down this is the BEST one out there trust me lol
You don't need to show **** content rolleyes.
I guess people are so excited because the lora also unlocks nsfw capabilities. However, it also improves other sfw related stuff. For example, I had a prompt describing a vampire with bloody vampire fangs. The base model gave the vampire normal teeths. The bypass lora gave him fangs. This lora also added the blood. So it seems to be more comprehensive than the bypass lora. Also, the author is right that the bypass lora is a relatively strong impact on the model. Only few parameters changed, yes, but this changes the vector representation of your text tokens very strongly. So this lora might be less impactful, but I haven't tested that.
@gueswork Yes, it's true. This thing really works well and doesn't affect the style.
@_kaidu_ Have you tried this bypass lora https://civitai.red/models/2746817/krea2-filter-bypass-fedor ? I've found it recently but had been using RR since it came out, works really nice. The fedor one has detailed explanation which I like, but it's not clear if the two loras works on different levels and if there's a difference in that.
@green_anger The "detailed explanation" is fully AI generated and partly bullshit. This alone is reason enough to avoid this lora. Maybe it works, but my feeling is that its fully vibe coded and the author has no clue what he is doing. I'm a computer scientist myself and I also use vibe coding from time to time, so I'm not against vibe coding in general. But I also noticed how llm models, even the smartest ones like Claude, are hallucinating a lot if you do not give them clear commands. So I would be careful using loras or comfyui plugins by users who have no clue what they are doing and just run an llm.
@_kaidu_ In my test it wasn't better than RR, and since 99% of stuff is in .safetensors format it's supposed to be safe to use regardless of content quality and author's intent. Plugins is another case.
With any AI generated output it's very hard to know if it's fake or real without vast domain knowledge, I have close to zero in how these loras operate, so for me it looks legit. Since my testing didn't show any improvement and you say it's somewhat bs, I'll take your word on it and skip it.
Personally, I use Claude for full stack development (and used other models before that), but I cannot call it vibe coding since I never take what it produces as is and like in 90% write all from scratch, only using provide info/code as a reference and always double check. And yes, all LLMs give gibberish very often, it's surprising how vibe coding is a thing at all 🤷♂️
@green_anger Maybe to be a bit more specific: the lora you mentioned is just the same as the filter bypass lora. It just set a single weight to 0. So it is not really any original work. It is a bit like these lora merged checkpoints. People create a new merged checkpoint and merge-in other people's lora. This is legit, if you invest time in experimenting which loras to merge and how the merge improves over each individual one. The original work is then not the loras themselves (they were made by other people) but the experiments you did. The fedor filterbypass is having a wall of text but it didn't mention any experiments. The whole thing was done by Claude and it is very likely that it just hallucinated stuff.
I'm also sure that this lora here contains a lot of vibe coded stuff. The difference is that it a) works very well and b) makes sense. The whole idea of this lora here is to not try to make as few changes as possible (keep parameter count low) but to make many changes but keep these changes very small (high parameter count but low weights). The result is that the lora seem to have less impact on the final image beyond the uncensoring behaviour.
Thx!. Yes, it's true. This thing really works well and doesn't affect the style.
For some reason, using it with my character LoRA (which works perfectly on the Turbo model) results in worse character likeness than using the original model without the LoRA.
Seems to still be very limited beyond simple nudity.
I think that's about the limit of Krea 2's knowledge. This lora doesn't really turns KREA2 to nsfw model and it doesn't introduce any new concepts. If you want NSFW with KREA2 you still want NSFW lora.
This is not an nsfw lora. It removes some of the builtin bias in Krea against blood, violence, nudity and many other things. It let Krea follows your prompts more clearly and it also improved prompt following in cases that were totally sfw and safe and without violence for me.
Use it as a base for better prompt following. If you want nsfw, add an additional lora for it.
PS: in my opinion it is totally fine if a base model cannot do porn and is not trained on porn. Totally understandable. What is annoying, is that the censoring methods often harm other capabilities of the models to an extend, that they cannot do anatomy correct anymore or ignore the prompt. Its good to have loras like this that circumvent that.
This is great. It's a small size (trivially so); & it does what it says on the tin. Namely, I think people are going in expecting full blown porn....even when the creator said that's not what it's for; the pertinent part "lower refusals without changing style, etc".
Here's a real world example; I had a prompt with "realistic vampire fangs"...... it's tame as hell (to the point you wouldn't even think that'd get censored), but yet.....that was censored! lmao. I switch this LoRA on, & now I'm getting the vampire fangs (& some other super tame as hell concepts as well that I wasn't before). For people that prioritise/want these things; & without other details getting obliterated or overridden in the process, then this is awesome!
I for one extremely appreciate what this does (I'm sure plenty of others do, too); so just ignore the derpy "Why this noes do p0rns" comments, because it'd be sad if you thought there wasn't a considerable amount of people that were interested in exactly what you did with this fantastic LoRA.
Again, much appreciated addition to the AI/Krea library. Cheers ^^
PS - to clarify why you want this; sometimes I'd need to use large NSFW LoRAs to generate non-sexual content (light horror, sci-fi wars, etc). These would often changes faces, etc, massively, as well as eat up VRAM. This doesn't! So it's somewhat of a no brainer in most circumstances to just use it, regardless. There's another LoRA claiming to do the same thing (also very small size)..... it doesn't; not in my experience. It would change the image with the same seed I was using often dramatically!
I also appreciate this is a superior technical feat; & whilst not as overt in its output as other clumsy LoRAs that do explicit content, it was likely much harder to do without changing other things inadvertently in the process.
Before/After LoRA Comparison
https://civitai.red/posts/29793114
Image 1 https://civitai.red/images/136770776
Image 2 https://civitai.red/images/136770787
Added images 3 + 4 (To showcase it can do bare breasts & pubic area). Some may want this (& only this) for tasteful, artful nudes; without changing anything else.
https://civitai.red/articles/32627/krea-2-nsfw-guide KREA 2 NSFW Guide (I ended up turning this rather long comment into a full guide, if anyone is interested). It covers a lot more (& I'll be updating it a lot, too).
Understand these examples are to demonstrate what this LoRA will produce again, that was censored before! It's NOT to show that it doesn't change other details! Because even without the LoRA, if you add things to a prompt/change it (like "city background", instead of a "Jungle background"), then it will most likely somewhat change the face, pose, etc, too. EVEN in the base model just BY ITSELF!
Again, this is to showcase things that would consistently not be included in the base Krea 2 checkpoint, that now are. & I did it in a way to demo a massive amount of changes all at once (that will naturally change the image composition noticeably.....that is a given to some degree).
Apologies for the CAPS.... but some people are just too stupid to read properly, & I'm not wasting my time with the inevitable "But hurr, durr, this changed the image too much as well" comments because I didn't spell it out with crayons for them. Under real world scenarios, if I just wanted "vampire fangs", then I would get just that, & without it changing the face likeness beyond the obvious thing I just altered!
Image 1 (Base Krea 2/No LoRA)
"realistic vampire fangs" = Censored ❌
"sticking tongue out" = Censored (Often considered sexual in nature; even though it's clearly not in this context) ❌
"vampire like fingers" = Censored ❌
"Her hand is covered in dark glistening blood" = Censored ❌
"extremely large breasts" = Extremely small, more like. Totally censored. ❌
"predatory look" = Censored ❌ (blank expression)
Image 2 (With the LoRA)
"realistic vampire fangs" = ✅
"sticking tongue out" = ✅ (Considered sexual)
"vampire like fingers" = ✅
"Her hand is covered in dark glistening blood" = ✅
"extremely large breasts" = ✅ (I'm not into this sort of fake looking thing, but it was just to show a very obvious comparison lol)
"predatory look" = ✅ (Probably not extremely obvious for most people to tell, but pretty much). This is frequently considered 'sexual' in nature; if you're wondering, & will entirely block an image generation on a lot of cloud models just by itself (on the same seed). If it goes through, then it's supposed to look like a narrowing of the eyes/scrunching, sensual gaze.
Conclusion
The LoRA added everything in the prompt that I put there! I tried the censored prompts over multiple images in base KREA 2; it didn't produce them a single time!
Again, these are tame comparisons; but if I put something like "throat cut & bleeding" (which I never showed, because I thought it'd be too extreme), then it would add her neck bleeding consistently, etc (with the LoRA). It would add other stuff like that (that again, I won't demo or show).
List of things that are included, that weren't before (timid things you wouldn't even think would get censored).
Image 1 (Without the LoRA)
"realistic vampire fangs" = Censored ❌
"sticking tongue out" = Censored (Considered sexual) ❌
"vampire like fingers" = Censored ❌
"Her hand is covered in dark glistening blood" = Censored ❌
"extremely large breasts" = Extremely small, more like. Totally censored. ❌
"predatory look" = Censored ❌ (blank expression)
Image 2 (with the LoRA)
All the above that was censored, is now uncensored. ✅
I'm also using this (it does little by itself btw) Qwen3-VL 4B Abliterated (ComfyUI / Krea 2 Text Encoder) BF16 / FP8
Sometimes an uncensored text encoder makes a massive difference (not in Krea 2), but it's worth mentioning if people have experience with using them elsewhere.
Quick question, does this Lora alters or helps also those finetunes that are being focused on ns+fw or directly apply ns+f+w concepts? Or it isn't necessary for those?
You will benefit from this, regardless. It's a small download, anyway.
Most NSFW LoRAs use a brute force strategy/image reinforcement to enable concepts; but they need to be trained on those concepts explicitly (excuse the pun). If they weren't, then you will confuse it outside of its training data (which may not include what you just asked for).
Example (this is just for instance) "Very Large Breasts" = Censored. It's highly likely this would be in the training data for most NSFW LoRAs, but if it wasn't, then it might not include it (or to a larger degree). If you ask for it in base Krea 2, then you'll get small breasts regardless. This would bypass it.
In real scenarios; You see people complain about bodily fluids in NSFW LoRAs....again, if this wasn't in the training, then it wouldn't be included; I've asked for timid things, & it refused to produce them (instead overriding them with the NSFW LoRA concepts in question instead). So yeh, this can help add things, without creating a spaghetti mess/tug of war scenario where multiple LoRAs try to pull multiple different image concepts in different opposing directions, & 'fight' over what it's trying to visually produce. This will uncensor things, without breaking other things, basically. That is my testing, anyway.
PS - A better example; a lot of facial expressions are also censored btw (in a sexual/violence related context). This will help you out there (& with some other similar things you may not have thought of that are also surprisingly censored).
@ShadowCell oh okay thank you, sadly I'm getting for the first time 3 arms with the Lora on at 1:00 as the creator recommend,idk why
@Neon_signs Might be from what I just said. If you ask for two opposing things; it might try to give you both, even thought it doesn't make any sense/is contradictory. With the way LoRAs work, you might be asking for them without even knowing it.
For instance "Showing ass" = LoRA thinking 'Ok, they want doggy'.....even though that's not what you wanted in the given context, & it wasn't an overt request; it will still aggressively try to force it along those lines on the limited data it was trained on.
The base anatomy in KREA 2 is really good, so it's highly likely it's just trying to fight over two different opposing concepts.
Upload the image with what you actually wanted/intended, & I can probably tell you what confused it, or it went wrong.
Just another comment of appreciation of this marvelous tool!
❤️
Thank you for creating it! Thank you for sharing it!
I don't think I can glaze this LORA enough. Thanks for it. It works incredibly well.
Works pretty good ! if you want more than nudity just fuse it with other loras around.
this is the best filter bypass!
This lora seems excellent and i have feeling it improves prompt adherence greatly.
Has anyone found a way to "remove" the censorship when using it?
I tried using different loras with different weights (snofs, mystic and realism engine) and it doesn't seem to work well with them (genitals remain censored), unless i increase their weight and and then it hurts the composition of the image.
Maybe someone have success and mind sharing?
I'm in the process of actually creating a guide to do precisely that..... but it's not ready yet... ¯\(ツ)/¯ lol
But I have before & after images with prompts for turning base Krea 2 images into NSFW, & keeping it otherwise the same. So the face, pose, style, etc, just stay the exact same (but with genitalia added afterwards). It's in conjunction with a position/pose prompt cheat list with the prompts necessary so it just consistently does them...
I hate to do the 'Tune in next week to find out what happens....' cliffhanger routine here.... but it's not a quick process to just describe, & it's silly when I've already started on the exact same thing in the guide.
Though I'll be honest; right now the images are in a realistic digital painting style; it's a subtle realistic 'noir' style, but I'll have to see if they also just convert easily to full realism without issues, & finish off the repertoire of positions.
To do what I'm describing, the prompts actually need to be completely, clinically air tight; & totally unambiguous structurally (which is why that's part of the guide, so people will have those prompt templates); if they're not, then any scaffolding holding the composition together, etc, will come undone the moment you add other LoRAs that try to tug image concepts/vague wording in different directions. I also still need to test the 'Pick n Mix' copy & paste format for it.....which is time consuming for a number of reasons (something that works perfect in one aspect ratio, can completely fall apart in another), so I kind of need to balance that for lower end users, too.
If that's a bit unclear, then part of the guide is actually explaining the proper prompt template structure; which is critical. You can't just bang random details on the end of the prompt, or have them randomly all over the place.
I will try to dig out some demonstrations of it.
@ShadowCell Thank you for the response, I'm looking forward to the guide then! Of course i also read earlier your comment and comparisons with/without lora.
So you put it needs to be a tight prompt engineering? well...i'll wait for you to properly finish the guide with some examples.
Amazing! This LORA is the only one that does not effect image quality with NSFW and other blocked contents being permitted. Almost all other methods that permit NSFW effect image quality, except this LORA. Great work.
this is the best by far
Is there anything to be done about Ideogram's censorship?
You need to use the json prompt format
This lora is like magic, genuinely. Comparing images with and without it, the prompt adherence is genuinely night and day.
I've been using it with an NSFW lora, and it has totally changed Krea for me.
Work great! Thank you for sharing!
生成男性器官时会有马赛克，有什么方法可以避免吗
Love your lora and it has become one of the few that I always leave on, but I have a noob question: from what I understood the lora targets just the "refusal layer" of Krea2 and in theory the same seed generation results should stay unchanged with the lora on and off IF the very simple prompt had nothing that Krea2 would refuse, But the results are different, composition, details, etc. I don't mean worse in any way, just different. So the lora does affect a bunch of stuff?