Progress update,

Okay, I need to correct part of my last update. Saying a LoRA alone CANNOT fix this was too absolute. The way we train it to behave on Turbo matters too.

v2 is still usable if you already downloaded it, but honestly the gains were not really worth it, so I have moved on to a different approach.

What I am working on now is taking the existing v1 LoRA, which was trained on the non-distilled model, and distilling its behaviour for Turbo. Basically, I run the original LoRA on Turbo with more evaluations, then train a copy of the LoRA to reproduce that longer result in fewer steps. The Turbo model itself stays frozen; only the LoRA weights get updated.

And yeah, this is actually giving me results worth working on. The goal is to keep the concept together through the denoise instead of having it show up initially, then shift around and end up with those weird textures. This is the direction I am focusing on now after the progress I have been seeing.

I am also testing this with character LoRAs, since the method can work with their existing trained layers too.

Still a work in progress. I want to see how consistently this holds up across more prompts and seeds before sharing the next version, but this is looking much more promising than where I was with v2.

V1.0

I recommend loading this LoRA with this LoRA Loader I just created :
https://civitai.com/models/2844700/comfyui-sigmasync-lora

Buy me a coffee :)

# 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.