It's hard to describe the reasons for releasing it

I just want to make models for use in face restore theAfter testing theIn testing baIn testing the base While testing the base model, I appended the previously trained female facial expression lora to the official Refiner model using comfyui's ModelMergeAdd. It turned out that it produced some changes.

I recall that not long after sdxl was released, the refiner feature was abandoned by everyone in the face of sniping from the many excellent basemodels.

So I spent some time putting together the bare parts reg images I used in creating the ???? lora, trained a lora, and added it to the original refiner model.

Tests afterward showed a lack of positivity in doing this, but I wanted to put it here anyway, maybe it's just because I have so few training images, maybe there's a master out there who can make the refiner model work again?

Creating Methods:

1. Official basemodelV1.0 + sdxl-wrong-lora (STRENGTH: -0.5)

2. LyCORIS was trained using ???? images on the model from step 1 and merged to the checkpoints. *Includes full body, half body, close-up of both men and women

3. Merge the previously trained facial expression Lora into the checkpoints from step 2. *The dataset has only females and the age range of the characters is very narrow

4. original refiner model + (Checkpoints for step 3 - Official basemodelV1.0). *clip-G has 40% from Checkpoints for step 3

The example images are the tests I performed, the non-refiner, the original refiner, and the refiner after appending 2 Lora's. The images can be dragged into comfyUI to see the test workflow.

The refiner's ascore function is still valid in the test.

*Turn on return_with_leftover_noise in the 1st KSampler of comfyUI to verify the Refiner function. I'm not sure what the normal flow of Refiner in auto1111 looks like.

Again: personally I don't think it makes much sense, download with caution.