Adds details to the image based on Qwen 2.1
This workflow is not for generating images, but for enhancing images that have already been generated.
I found Workflow online as free.
I found the VAE on Hugging Face.
VAE
madebyollin/texture-fix-vae-for-qwen-image-2.1 · Hugging Face
It improves texture handling.
When working with anime and illustrations, it is recommended to move the second-to-last line—"real photo"—to the negative prompt. I didn't put it there myself because it sometimes yields interesting results without altering the character. HOWEVER, if the character changes drastically and you aren't happy with the result, then move that line to the negative prompt.
I do not recommend using the Qwen 2.1 model at a precision lower than INT8, nor should you use any "TurboLoRAs" (this is necessary to preserve the scene and ensure improvement rather than degradation). All examples were generated using the INT8 model.
Consequently, the minimum system requirement is 12 GB of VRAM; cards with 16 GB or more are recommended.
System RAM should be 32 GB or higher.
Description
Comments (2)
Hi, is this for Qwen 2.1? It's listed as Krea 2?
The Qwen 2.1 model with an improved VAE is used as the detailer. You can enhance any image, but you need to download the Qwen 2.1 text encoder and VAE from the provided link.
I’m updating the description to Qwen now.
In short, within this workflow, Qwen acts as an upscaler, detailer, and refiner.