custom VAE for Krea2

[lillblues]

It seems that the Qwen VAE used fro Krea2 is not ideal. Many users have reported that replacing it with Wan2.1 VAE produces much better results. I tried to define a new VAE for Krea2 in Finetune but Wan2GP ignored it. Is there other way to use the Wan2.1 VAE for Krea2?

[diegomariop]

Did you create a finetune file?

[lillblues]

Did you create a finetune file?

Yes, the finetune file was created like this:

---

{
"model": {
"name": "Krea 2 Turbo with Wan2.1 VAE",
"architecture": "krea2_turbo",
"description": "Krea 2 is the most aesthetic open-source image model available according to its authors. Krea 2 Turbo is the distilled few-step Krea 2 checkpoint for fast text-to-image generation.",
"URLs": [
"Krea2Turbo_quanto_bf16_int8.safetensors"
],
"VAE_URLs": [
"Wan2.1_VAE_upscale2x_imageonly_real_v1.safetensors"
]
},
"prompt": "a fox walking in fresh snow at golden hour, natural light, detailed fur",
"negative_prompt": "",
"batch_size": 1,
"denoising_strength": 1.0,
"masking_strength": 1.0,
"num_inference_steps": 8,
"guidance_scale": 0
}

the ckpt and VAE were already downloaded in ckpt folder.