Create Cohesive Images With 2 Subjects using 2 Different Reference Images
Vložit
- čas přidán 10. 07. 2024
- Welcome back, brilliant people! In today's video, we take our InstantID series to the next level by combining two reference images to create a single image with two subjects. Whether you're turning photos into characters like princesses, doctors, or comic book heroes, this tutorial will show you how to add a second subject into the mix!
Video Chapters:
0:00 - Introduction to InstantID with Two Subjects
1:15 - Setting Up Compute Grid
2:30 - Installing InstantID in ComfyUI
4:00 - Workflow 1: Basic Setup
7:15 - Workflow 2: Adding a Second Subject
10:00 - Workflow 3: Adjusting Head Positions and Masks
13:45 - Workflow 4: Enhancing with Custom Checkpoint Models
16:30 - Final Results and Outro
Key Points:
Compute Grid Setup: Learn how to start and configure your GPU instance for InstantID.
ComfyUI Installation: Step-by-step guide to installing and setting up InstantID with ComfyUI.
Two Subjects Workflow: Detailed process of adding a second subject, including image loading, keypoint positioning, and mask creation.
Advanced Tips: Fine-tuning with custom checkpoint models and adjusting conditioning parameters for optimal results.
Helpful Links
Comic Model I used in video: civitai.com/models/260599/pix...
Video Link on how to install Custom Models: • How To Upload & Run Cu...
Compute Grid: computegrid.ai
If you found this tutorial helpful, please consider subscribing to the channel and giving this video a thumbs up. Your support helps us continue to create valuable content for you. Let us know how it worked for you and what other tutorials you'd like to see in the future.
Thanks for watching! Keep exploring, keep experimenting, and we’ll see you in the next video!
Great tutorial, really love your work.
None of this makes since to me, but all I know is that you have a great ASMR voice.
These instantID tutorials are exactly what I needed, big thanks
I appreciate this explanation, I have been trying to do this with various different setups (FaceID, InstantID, Reactor) but don't get the results I'm looking for. Or blow up my notebook when running faceid.. Could be because I'm trying to rig Pony to behave with InstantID for consistent characters.
One thing I would like to recommend, post processing of the audio. Your audio has a lot of wet sounds, as if your mouth was either try or trying to keep gum in your mouth. Not sure how I can try to point it out in the kindest way I can, but its one of those sounds that sticks out to me and made it hard to stick with the video.
Afaik up to this date control net does not work/work well with pony.
Please share workflow download link
Very clear tutorial.
Maybe some chekcpoints models stylize so strongly that they wash out the features of the referenced faces... At least that's my brief experience with InstantID.
I always preferred face Id over instant Id, for example because of the key points needed.though they help to be able have two characters in one image.
Great idea for a tutorial but I couldn't get it to work quite right and think I noticed some issues in your workflow once I started previewing the face keypoints and masks. I noticed the masks were just giving me a full black and white mask instead of half and half. I think that is because you had multiply and it should be add for the MaskComposite node. Also I noticed when I previewed the invert image that's used for the face keypoints it was just inverting the colors on the image and not actually flipping it horizontal. I found the Image Flip node and that did the trick. But thanks for the tutorial either way, found some new tricks and learned some stuff! I might also play around with separate positive prompts but seems to work ok combined like this now that i figured out the masks and keypoints properly.
I apologize it didn't work out as well for you; but you are correct, the MaskComposite node should've been 'add'. Thank you for bringing that to our attention, so we can make a correction.
We are glad regardless though that you were still able to get good information from our video. 🙂
Also, when you play with the workflow you can always modify it so that not only does each subject get it's own prompt, but also its own pose.