Krea2 TextFusion Refusal-Reduction [by Capitan01R]
No idea what happened with Capitan01R
I will happily take the model down if Capitan01R decides to come back. Until then yellow buzz generated with this gets tipped to him. If he deletes his account I'll throw the buzz to some random dude.
But this is a great LoRA and it should be available to the community.
Update: Capitan01R politely declined the yellow buzz, so it will be random users.
# 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
FAQ
Comments (31)
Thanks for the reupload, man. 👍
indeed. @CraigToomy we gotta never take things for granted and must all stay greedy and download and hoard everything. who knows if civit . red will remain intact long term.
Can someone say wich one is better? 3 diff versions to do the same?
No, it's technical, but the bottom line is the higher versions "go deeper" in the model substructure, so they provide a deeper coverage at the price of more skew to the gen output.
The thing is, with Krea2 that skew to the gen isn't necessarily a bad thing, 'cause the innate NSFW refusal is so strong it actually skews gen output by default, which is why base Kerea2 actually has worse prompt adherence than when u add a refusal reduction /any NSFW lora to it.
So... it's complicated.
And each gen is different, no way to give a hard rule which to use.
But still, If you have no clue, I'd advise use v1, or v2 is also good, but if you see your particular prompt/lora combo needs more NSFW unlock then you'll need to go with the deeper fix.
@ravemry9 deeper fix means the v2 full rank, right?
@ralphjx yeah, v2 full-rank is deepest, then goes v2, and v1 is lightest.
@ralphjx I've just heard from @oneluckyluke that v2 (non full-rank) has an artifact problem
Yep, In order of usefulness: V2 full rank, V1, then V2. For it's size, V1 gets me about 95% of all the unlocks I need so it's hard to justify using the full rank version given its size.
Also, you should know that the base krea 2 model only has so much knowledge of NSFW concepts, you can't unlock what the model doesn't know. Finding a good finetune (e.g Kroma ODP) that has a huge amount of NSFW knowledge baked in should be the place to start.
Lastly, this helps with more than just NSFW, it also helps with prompt adherence to facial expressions, actions and even art styles. For example, krea2 turbo struggles showing people sneezing or spitting or furious, this lora enables the model to do that. There is a wide range of things that the model interprets as "unsafe" for no good reason and it will refuse to portray it or simply dampen its effect.
@oneluckyluke thx.
BTW, you mention Krea2 turbo so I just wanna add that I've had the same prompt adherence struggles (including facial expressions) with full Krea2.
P.S. I love the Krealism and Gonzolamo fine-tunes, you have any idea if they have the added NSFW range on par with KromaODP?
@ravemry9 I haven't tried those fine tunes yet but I will check them out soon! Thanks for putting me onto them
@oneluckyluke So, is this LoRA—in its v1 or v2 full-rank version—better than the "bypasslora" that’s also available here on Civitai?
@ralphjx In my limited experimentation, I think so. I started out using the bypass lora and switched to this one because I was getting better results with it.
@oneluckyluke @ralphjx
same here. I did a bunch of tests with several NSFW-unlock loras before deciding this is the best. (but some of them weren't poorer quality, just heavier style bias or similar subjective preferences. I do think capitan01r's is the most comprehensive unlock with the least side effects; objectively).
@ravemry9 I tried version v1, and the difference in terms of adherence was night and day compared to BypassLora. I’m definitely sticking with this one. Thanks, guys, for the response.
V1, hands down. other 2 must be overtrained or something because for me they give boring results.
Does this help if im already using mysticxxx?
Yes, it helps, but it's usually not necessary, their use-cases overlap.
It's a matter of personal taste, but imo mystic has more of style and composition bias and this lora is more of a pure NSFW-unlock.
In my experiments, mystic works great for NSFW concepts that the base model doesn't have. This lora helps with prompt adherence for things that are damped by the base model's safety filters like facial expressions, open/closed mouths and other non-NSFW poses. Mystic tampers with art styles since it's mostly trained on realistic content.
Best combination I've found for NSFW is using kroma-ODP and this refusal lora. Kroma gives you more NSFW concepts than mysticxxx and incredible prompt adherence including styles.
@oneluckyluke This is very useful info, thx.
I hadn't realized the krea problem with facial expressions would be fixed by NSFW unlocks! stupid of me, should've realized. Thanks a lot. I've been fighting with krea for a long time to get grumpy cats, almost never works, do you know if animal expressions are in the same boat?
Also, I'm into bondage (light; just basic ropework) and base krea is pretty bad at it, do you know if mystic is better at it?
@ravemry9 Yes, animal expressions are definitely in the same boat, I tested this with krea2 turbo and refusal v1. Tried to generate angry 3d monkeys. Without the refusal Lora, best I could get is "mildly annoyed", the safety filters soften the expressions to make them non-threatening. Once I added the refusal lora, the monkeys are way more furious and menacing, bearing fangs and claws, angry skin folds on the face etc. The difference is night and day.
I have no idea if mystic is any good at bondage, I dropped it as soon as I tested Kroma-ODP as my base model for NSFW. It is trained on a much wider range of NSFW concepts. There's also "Sick ollie" and "Lustify" that might work for your use case.
@oneluckyluke got it, all the info is much appreciated XOXO
@oneluckyluke May I join your discussion? I have a question: I recently discovered the Crea2 Turbo model and this LoRA. I wanted to ask—if I use them to generate anime-style NSFW content, what might happen? I also noticed a mention of "kroma-ODP"—what is that? Thanks!
@Sputnik_Art I'm not sure what you mean by "what might happen", I'm assuming you are asking about the quality of the generated image? If so, it depends on a lot of things, Krea 2 turbo generally produces decent anime quality images with only prompt descriptions. The model contains a "safety lock" that will steer away from NSFW images. If you use this LoRA combined with Krea 2 turbo and a good anime prompt or anime LoRA, you will get an NSFW anime image. This LoRA will remove the "safety lock" so the Krea 2 Turbo model can generate NSFW.
"Kroma ODP" is a model based on Krea that has been trained on more NSFW and anime images so it "knows" about a lot of different NSFW and anime concepts. Krea 2 turbo may understand human anatomy well and some basic sexual concepts but Kroma understands more advanced and complex sexual concepts. However, because it was trained in such a way that it preserves a lot of the original Krea's capabilities, it sometimes suffers from the same "safety lock" issue so you may need to unlock it (not always though). Kroma also produces more variety for the same prompt and it understands prompts a lot better than Krea 2 turbo. I have noticed for example that it understands style prompts much better.
Hope this helps.
@oneluckyluke Understood, thank you for the response. I used a translator, as English is not my native language. Thank you.
i looked at @Capitan01R github because there was some other good stuff there too, and that's also totally wiped. all 14 repos gone. bummer.
Does anyone know what did happen to Capitan01R? was it a mod action, or he decided to close up shop himself?
I don't know, but announcement on his profile said a few days ago that his reddit account was banned or something like that.
@RisingV damn.
Hope the peanut gallery didn't drive him away, he made good work and seemed a good guy -- afaict this is the best krea2 unlock (for pure nsfw unlock, the others like mystic are great but have so much style/concept bias)
Thanks for keeping this critically important LoRA online!
bless whoever made this. this isnt just about nsfw it just straight up seems to unlock parts of the model ive never seen. they all seem to do slightly different things, im just going to use them all 😈
UPDATE: I am REALLY starting to prefer version 1 to all the others quite honestly. It's so freaking good. Its got the most character and feels the most 'natural' imo. Other 2 are boring in comparison, the full rank one is glitchy.
Is there any difference between this model and https://civitai.red/models/2993700/krea2-textfusion-refusal-reduction-loralowrank?modelVersionId=3394995? Or are they both just reuploads of Captain's.