HyperDetailer with Metadata | Invoke | Page 1

worthy swan Mar 2, 2025, 2:42 AM

#

The workflow from hipsterusername with Linked Metadata Nodes.

#

Fixed Version with right Model in Metadata.

📎 Hi-Rez_Fix__Detailing_SDXL__Linked_Metadata.json

worthy swan Mar 2, 2025, 3:08 AM

#

Nodes needed: https://discord.com/channels/1020123559063990373/1133465385182699582

placid haven Mar 2, 2025, 2:28 PM

#

you can avoid using the xygrid nodes if you use the new denoise latents + metadata node that is now part of main in place of the last first two denoise nodes it will capture most of the metadata fields automatically. The only things it does not capture are the prompts so capture them in the same way as you do now and feed them into the final denoise node.

#

this also makes the metadata collection much simpler.

#

here is my attempt at reorganising the workflow with those changes but I've never used it before and it will need double-checking as I've never used this workflow before and I got a slightly screwy output but I might just be using it in an odd way.

📎 Hi-Rez_Fix__Detailing_SDXL__Linked_Metadataskunk.json

placid haven Mar 3, 2025, 3:52 PM

#

I was wondering why I was getting poor results in the upscale phase of the workflow so I added a tile controlnet to that part and it really improved things and now keeps consistency with the first image generation. I also played around with the linear UI part to tidy it up a bit. It could do with some text to explain things a bit. Any thoughts?

📎 Hi-Rez_Fix__Detailing_SDXL__Linked_Metadataskunk.json

#

It would be nice to add the ability to easily switch between random and fixed seeds. The only way I can think of doing this easily is with Stretches Integer Selector node. https://discord.com/channels/1020123559063990373/1278825261432967169

median belfry Apr 30, 2025, 12:05 PM

#

Hmmm how come generating with this workflow the same prompts and same image size immensely slower ?

#

am irunning out of vram

#

I made this frankensetin version which inputs the seed cfg prompts etc from image metadata

📎 Hi-Rez_Fix__Detailing_SDXL_face_darcow.json

#

I dont think the loras are working properly here tho

worthy swan Apr 30, 2025, 7:36 PM

#

median belfry I made this frankensetin version which inputs the seed cfg prompts etc from imag...

This is a working Version of your workflow. i removed most of the metadata stuff and this should work fine. (i did not test the lora because my test image has no lora)
it did some work in the image but maybe not the best example (because the test image was a good one)

📎 HyperDetailer_from_existing_image.json

median belfry Apr 30, 2025, 10:33 PM

#

worthy swan This is a working Version of your workflow. i removed most of the metadata stuff...

Oh wow thats a lot better and faster than my janky one

elfin viper May 5, 2025, 7:39 PM

#

Interestingly enough, some of the models work, some of them produce a black square around a face.

worthy swan May 5, 2025, 9:52 PM

#

maybe you need to change the Scheduler for some models

soft orchid May 15, 2025, 10:51 PM

#

Made some tweaks to the most recent version posted by Pfannkuchensack. This update fixes the positive and negative prompts being swapped going into the denoise node, which was producing some wonky output. I also added a fp32 toggle (having it off was causing black boxes for me), made it so you can change what part of the image is being detailed directly on the form, and pulled the original metadata through to the final output image so that it can be used as an input for further detailing.

📎 HyperDetailer_from_existing_image.json

#

Also pulled most of the denoise config values back out of the metadata so that you don't have to modify the workflow when you change image generation parameters.

hasty leaf May 20, 2025, 7:02 PM

#

Im trying to figure out this workflow but keep getting this error "NodeInputError: Node 364b0742-112d-4987-8dae-d4db486ab370 has invalid incoming input for cfg_scale"

#

Ok actually i found where to edit that, Still trying to see how to utilise this workflow properly so any tips are very welcome

soft orchid May 22, 2025, 2:00 AM

#

The main use of the workflow is to add extra detail to faces in images (but you can detail other things, too). Since stable diffusion essentially has its "attention" spread across the entire image, things with fine detail tend to get a bit . . . wonky. Distorted faces are especially common when it's a character at a distance.

This workflow uses a "Grounding DINO" node to find a specific element in an image (defaults to "Face", but you can pick other objects too). It then re-generates just that portion of the image, focusing only on that element (as if you had generated a super-zoomed-in version of it). And it doesn't regen it from scratch, but uses the initial image as a starting point and adds about 35% of the noise back in, regenerates it, and saves the new image to the gallery.

The end result is that faces (or other objects with fine detail) can have extra detail added back in sort of "automatically" without the user having to use inpainting. In fact, this is essentially "automatic inpainting" for objects that DINO can find in the image.

blissful marlin Sep 28, 2025, 7:56 AM

#

thanks a lot! included this workflow into my regular one! works like magic!

blissful marlin Oct 1, 2025, 1:37 PM

#

However! 😄 Now I found that on some models detailer creates slightly mismatched in color images.

I tried to play with a workflow, changing the mask blur, but nothing. Color correct node also bugs out on me, I don't know why. Thinking about applying curve fix node. Maybe someone have other ideas. Please share!