Character Creation Log #1 | Creating a Consistent Character | Midjourney/Nano Banana/Kling Elements/Higgsfield Soul ID
Tokii here.
I have been researching AI music and AI artists, but
Will AI artists be loved by humans?
・Analysis #1: Tata Taktumi
・Analysis #2: HAVEN.
・Analysis #3: Xania Monet
Rather than just doing research, I thought, "I just have to do it myself!" so I started creating a character.
This is a personal record, but I want to make it as meaningful as possible by sharing the points where I got stuck, steps I didn't need to take, efficient production steps, and actual prompts for creating a consistent character.
Step 1: Image Generation
photo realistic Japanese 20-year-old beautiful woman with long black ponytail wearing a pastel blue and pastel pink vanity jacket, hoop earings, pastel pink lipstick, tight blue denim jeans and stylish boot heels. Tokyo Tower visible in the distant background. --ar 3:4 --raw --profile 9hpl985 --v 7
Midjourney keeps making them look like different people🐱
The prompt for the first image was "photo realistic〜", and despite having photorealistic personalization ON, it turned out illustrative. Maybe the pastel-related words had an effect? or perhaps the AI perceived it that way because I imagined a character with a new-retro/city-pop feel? might be the reason.
Photorealistic materials were also produced with the same prompt, but I was drawn to this character, so I decided to use these images as a base for photorealism.
Step 2: Generating Variation Images
This will be the standard for later profile, angled, waist-up, and full-body images
First of all, it is efficient to create a bust-up (or close-up) image where the front face is clearly visible. I do not recommend starting from a full-body shot because the face will shift with every generation.
Also, it is super important to choose one where the gaze is firmly fixed on the camera!!
the model ended up looking to the left the whole time... (hours of loss⏰️)
So, for the reference image, let's choose a miraculous shot that is bust-up (or close-up), front-facing, and looking at the camera.
Once you have a bust-up image, use that image as a starting point to select the best shot of the profile.
Using these two images as references to create images from different angles, and then adding those to your reference images to expand to waist-up and full-body shots, makes it easier to maintain facial consistency.
Creating variations of your preferred expressions beforehand also makes it easier to maintain your preferred expressions during later image/video generation.
there is no need to combine them into a single sheet like this 📝
(though it is easy to understand when showing it to people).
Step 3: Elements (Character) Registration
Once you have images from various angles and framing, upload and register those images to Kling's "Elements" or Higgsfield's "Soul ID".
Kling Elements → Up to 4 images can be registered *Variation generation is also possible from a single image
Higgsfield Soul ID → 20 to 80 images can be registered
This time, I registered using only images (I need to create a voice for it too 🗣️).
While Kling can generate variations from a single image, the generation is a bit of a gamble, so it is overwhelmingly recommended to prepare your preferred profile and full-body images.
(Preparing the images is quite a task...)
(Generation platform is Flow/Higgsfield).
Create a photorealistic studio image of the same woman shown in the reference images.
Use image 1 as the main facial identity reference.
Use image 2 as the left-profile facial structure reference.
Use image 3, image 4, image 5, and image 6 as the full-body proportion, outfit, hairstyle, and silhouette consistency references.
Preserve the exact same character identity across every image: same face, same high ponytail, same large gold hoop earrings, same pink satin bomber jacket, same white crop top, same pink jogger pants, same white sneakers, same slim body proportions, same realistic skin texture, same clean white studio background, same soft diffused studio lighting.
Do not invent a new face. Reconstruct only the same woman from the reference images. Keep the anatomy realistic, the head size proportionally correct, the shoulders natural, the hands accurate, and the outfit unchanged.
Single subject only, no text, no collage, no props, no background elements, no stylization, no anime look, no cartoon rendering.
image 1 = Bust-up left
image 2 = Bust-up front
image 3 = Full-body right
image 4 = Full-body back
image 5 = Full-body left
image 6 = Full-body front
Elements (Character)bybindingit,consistency is maintained
Here is a video generated with Kling Video 3.0 using the registered Elements↓
By the way, without Elements, not only the man in sunglasses but also the woman's face will become distorted↓
Google tends to have a Veo accent and Veo face, and Sora makes even a woman without sunglasses look like a different person... (Sora 2 even has the actress laughing at the end, lol)
At the end of last year, I gave up on character consistency in live-action and started using anime characters as leads instead↓
Finally, we can now maintain character consistency in videos too.
Summary: Tips for creation, time-saving points & precautions
・Start character images from a bust-up (or close-up) front view
→It is difficult to maintain facial consistency if you start from a full-body shot
・Make sure the base image is looking directly at the camera
・Nano Banana Pro/2 is the best for generating variation images
→I think it's pretty much the only choice (though Seedream can be good at times)
・Increasing reference images when generating variations makes it easier to maintain consistency
・Remove items like sunglasses or headphones that aren't always worn
・You don't need to be too picky about fashion when creating the character
・A single character sheet is not needed for Elements registration
→Because you upload one character image per file
・Kling Elements allows registration of 4 images
→Kling automatic generation (AI Multi-Shot) from a single image is difficult to keep consistent
・Higgsfield Soul ID Character allows registration of 20 to 80 images
・You can also register a voice (Planning to try this next time)
It was a more voluminous task than I thought, but once registered, you can call it up easily, so for those who want to create consistent character videos, Kling's Elements feature and Higgsfield Soul ID are the best options for now.
The more things you can do, the more the root question of "what are you going to create?" is asked.
(If you find computer-related words in their names, leave a comment😽)
Next, I want to create a voice, and before that, I want to make music,
and I also need a story for the two of them...
I have no choice but to do it, right?
So, I'm going to use these two not just for experiments and verification, but to create music and stories.
The adventure continues...
Tokii
