#Improve a Batch of SDXL images automaticaly with Flux

1 messages · Page 1 of 1 (latest)

latent fiber
#

Hi, i try to get a bit deeper into my very flat nodes knowledge.

So i try to automate a workflow, where i want to improve SDXL images with Flux, using varrying denoising strength. i import a batch and the metadata nodes extract the Prompt from the imported images. Furthermore, i want to be able to use varying sizes of imported images.. but here i struggle.. only my 1024 to 1024 images at the moment. the other sizes are ignored.

I dont understand where the Denose node of flux gets the size from, and how to tell it which size to use. So maybe someone can point me to whats wrong.

By the way sorry for my english.. german is easier... "i think"

Flux_img_to_img_1.png Flux_img_to_img_2.png 📎 gmp_-_img_to_img_-SDXL_batch_-_to_FLUX_-_auto_prompt___size.json
rough storm
latent fiber
# rough storm You need to resize your images to fit Flux requirements. You can use a Float to ...

Thank you ... i do not yet understand realy what you mean.. so better i put a screenshot..
if i got an image input of lets say 1200 width and 800 hight.. from SDXL.. so what are the Flux requirements you talked about. can it only create sizes at the multiple of 8 pixel ore so? just gessing.. so i would have to set the neares multiple to 8?

and where would i talke the size from? from size of the image batch node, the metadata node extracting the size, or from the imge to latend node.. ???? 🙂

Well blody beginner as you see..

image.png
rough storm
#

If you’re using flux, you’ll need a multiple of 16

#

You will need to use a resize node to transform your image to the correct size, and then use the resized image for everything downstream of that

latent fiber
# rough storm If you’re using flux, you’ll need a multiple of 16

Thank you so much... what a fun when finaly it works.. now ill integrate it in the bigger workflow..

so the way was :

  1. extract metadatas from image batch
  2. Extract metadata width and highth
  3. feed this to the Float to integer with : Multiple of 16 (to fit to flux) and Method " Nearest"
  4. feed the output to "Resize Image"
  5. feed the output from Resize image (Width and Hight) to the Denoise node of Flux..

at last this gave me now img to img

agile wasp
#

in this instance, you don't need the metadata from image and metadata to integer nodes you can take the width and height directly from the Image Batch node. Unless your metadata stores some width and height that is not the actual image size.

trail hawk
#

Hey guys. I'm trying to do this but with incorporating the object detection. I want to improve the faces and hands of a batch of sdxl images with flux basically. you know if this has been done before?

latent fiber
latent fiber
agile wasp
latent fiber
agile wasp
#

I vaguely remember there was a workflow around that did something like that but no upscaling

trail hawk
#

thanks for sharing this. this is a good starting point

#

in terms of improvement. i wanted to use flux inpainting rather than a detailer

trail hawk
#

so the top half inpaints the face and the bottom half makes a bounding box around the inapinted face?

image.png
#

where is the detailer?

#

@agile wasp @latent fiber @rough storm

latent fiber
agile wasp
# trail hawk where is the detailer?

its the whole lower part of the workflow from the Grounding Dino node onward. It isolates the bounding box returned by the Dino node scales it up to an ideal size, does an image-to-image at the higher resolution, scales it back down and finally merges it back into the original generation.