Krea2 Refusal-Reduction — v2 and v2 Full Rank

Okay, I am adding two versions this time: v2 and v2 Full Rank.

The goal is still the same: reduce refusal behavior while preserving as much of the model's existing visual knowledge and prompt adherence as possible. Getting a concept to appear initially is only part of it. I also want it to hold together through the denoise without drifting into something else or ending up with those weird textures.

There is also an important change to the setup if you use character LoRAs. Please read that part before loading everything together, because conflicts between them have been a recurring problem.

The Two Versions

v2

File: refusal_reduction_v2.safetensors

This is the smaller rank-64 version, approximately 30 MB. It includes updates to the four TextFusion blocks, the tap projector, and the two external text MLP layers that prepare the conditioning for the image transformer.

v2 Full Rank

File: refusal_reduction_v2_full_rank.safetensors

This is the larger full-rank export, approximately 2.6 GB. It covers the same main conditioning components and also includes selected normalization adjustments.

The matrix updates are stored at their full available rank in FP32, which accounts for the much larger file. “Full Rank” describes how those updates are represented; the changes are still confined to the text-conditioning components covered by this release.

Load one version only. Start at strength 1.00.

What These Versions Change

Krea2 takes several representations of your prompt from its text encoder and processes them before they reach the image transformer. These releases modify that processing, with the aim of making the model less likely to suppress the requested concept.

Both versions include:

• The two TextFusion layerwise blocks.

• The two TextFusion refiner blocks.

• The projector that combines the twelve text-encoder representations.

• The two external text MLP layers that feed the resulting conditioning into the image transformer.

The earlier description saying the projector and external text MLPs were excluded does not apply to these two files.

Neither file contains updates to the image-transformer blocks. They change the conditioning those blocks receive, which can still change the resulting image substantially.

The purpose remains refusal reduction. I am not aiming to impose a character, style, or preferred aesthetic with these releases.

Using a Character LoRA? This Loader Is Required

If you combine either version with a character LoRA, use Krea2T Character LoRA — Image Only to load the character LoRA.

One thing I have repeatedly run into is a character LoRA pulling an otherwise working setup away from the requested concept. You get the likeness, but other parts of the prompt start behaving differently or become less reliable.

Simply excluding weights from the dedicated text-processing layers does not fully separate their influence. Krea2 also has shared layers that process both the prompt and the image. A character LoRA applied normally to those layers affects both.

That is why I added this loader. It filters the supported character weights and applies their direct updates only to image tokens. The refusal-reduction LoRA stays active through its own loader.

Text and image still interact through the model's normal attention, so some interaction remains. The loader targets that direct overlap rather than treating every conflict as a reason to keep increasing LoRA strength.

Connect It Like This

Krea2 model
↓
Regular LoRA loader — refusal_reduction_v2 OR refusal_reduction_v2_full_rank
↓
Krea2T Character LoRA — Image Only — your character LoRA
↓
Sampler

• Load the refusal-reduction LoRA through your regular model LoRA loader.

• Load the character LoRA through the Image Only loader. Replace its previous loader; do not load the same character twice.

• Do not put the refusal-reduction LoRA into the Image Only loader. Its filtering excludes the conditioning weights these releases use.

• If you are not using a character LoRA, connect the refusal-reduction loader directly to the rest of your workflow.

Still a Work in Progress

I am continuing to check how consistently this holds up across prompts, seeds, and different LoRA combinations. The objective is useful refusal reduction with the concept intact through generation. I am not calling every changed image a success.

Buy me a coffee :)

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.