Everly Heights Character Maker 1.5

FULL DOCUMENTATION HERE

Create consistent characters in Stable Diffusion with the Everly Heights Character Maker 1.5, a custom version of the standard Everly Heights Character Maker. I've had several requests for this version, trained against SD 1.5 from people, so I finally went back and made one. This essential asset for animators and storytellers allows you to produce comprehensive views of your characters, facilitating animation or the development of a dataset to train a LoRA model for your original character.

The Everly Heights Character Maker Pony XL lets you generate a full character sheet with the three most common poses for animation (t-pose, back side, Side profile) in one image, or lock your seed and prompt each view for your character individually.

Find out more about the origins of this tool by watching our video AI Character Design for Animation.

Everly Heights Character Maker 1.5 – Every angle tells a story.

Tutorial

EVERLY HEIGHTS CHARACTER MAKER 1.5 FEATURES

• Multiple Angles Made Simple: Capture your characters in multiple stances with unparalleled accuracy. Whether you're designing for animation or static art, our specialized keywords like "t-pose," "Side profile," and "back side" allow for the creation of crystal-clear, perfectly positioned character images when you lock the seed. This ensures consistent quality and alignment across all your creative projects, making the Everly Heights Character Maker 1.5 an indispensable part of your artistic arsenal.

• Dynamic Character Sheets: Use the "4view" keyword to create detailed character sheets, ideal for ensuring consistency across animations.

• Style Adaptation with LoRA: The Character Maker LoRA works at a low strength, so you can dial back the Everly Heights aesthetic to fit your unique style while keeping the functionality.

• Consistent Character Engine: Character Maker is consistent. Lock in your seed and effortlessly switch between views without compromising your character's identity. You can even change the character's emotion by subtly modifying your prompt. This breakthrough ensures your characters remain true to your vision from every angle and emotion.

• Dataset Generator: Create a rich, diverse dataset for training a LoRA model using your Character Maker outputs to create 20-50 pictures of your character by mixing zoom levels, crops, and backgrounds. Mr. Matheson XL and Infinite Tina XL were both trained using this method.

• ControlNet-Driven Pose Consistency: Use our ControlNet OpenPose templates (below) to get the same poses for every character you generate, which is crucial when animating multi-character scenes. Every character will have the same base set of images you can build on with inpainting or reprompting with minor edits (i.e. 'a mad clown' instead of 'a sad clown').

Trigger Words

t-pose - Facing forward, arms out.
Side profile - Character stands 1/4'ed out/at 90 degrees.
back side - Shot from behind
4view - All three angles, plus a bonus Side profile.
Character sheet - Optional, uses Stable Diffusion's preexisting knowledge of character sheets
in front of a green screen - Ensures you get a bright solid green background every time. Otherwise, the background color will be a random color that contrasts the colors of the character. This phrase can also make the background cleaner.

INSTRUCTIONS

Everly Heights Character Maker 1.5 features three poses and a character sheet generator built-in.

Individual Poses

To prompt a particular pose, simply use one of the keywords t-pose, Side profile, or back side. I recommend using a resolution of 768x768. Other aspect ratios work, but tend to crop some part of a character.

Create your base prompt and use it with various seeds until you find one you like. Lock in the seed, then change out the keyword to get the other views of your character. Here's an example of a teen witch, all rendered with the same seed:

[gallery type="square" size="medium" link="none" ids="1651,1649,1648"]

Character Sheets

To generate a character sheet with four poses, just prompt using the keyword 4view. Sometimes you'll get more than four views if you use the keyword, and sometimes views will repeat. You can leave the keyword out and render at 1344x768 and you'll usually get at least 3 usable views. That being said, multiple versions of the same view can be useful, as you can cobble multiple versions together in post.

I recommend using a resolution of 768x512 as that's what I trained the model on, but other resolutions also give good results. You can also include the term "character sheet" in your prompt to leverage what SD 1.5 already knows about character sheets and make the results even better.

[gallery type="thumbnails" link="none" columns="4" size="medium" ids="1652,1653,1654,1655"]

Character Sheets with OpenPose ControlNet

Prompting alone will give you usable outputs, but if your animated project includes several characters you'll want to give yourself a consistent workflow for character rigging and animation. Use the ControlNet OpenPose templates below to lock in the character's poses. These were used to create the training images, so the model works very consistently with these. Some seeds/image resolutions will give a similar output with the 4view keyword, but it's a bit hit-or-miss.

[gallery type="thumbnails" size="medium" columns="2" ids="1656,1657"]

[gallery type="thumbnails" size="medium" columns="2" link="none" ids="1660,1662"]

You can still prompt using the 4view keyword (i.e. 4view, an evil clown) if you use the ControlNet to increase the consistancy of the characters. Sometimes dialing up the ControlNet weight to 1.05-1.2 will improve results with a complex prompt. I recommend stopping the ControlNet model before the generation is done to give the model time to tidy up. Here are the settings I use:

[caption id="attachment_1658" align="aligncenter" width="508"] ControlNet Settings for Everly Heights Character Maker[/caption]

ADetailer/Inpainting/Hires fix

The individual 768x768 poses come out with solid details (face, hands, etc.). You can improve them even more by rendering your image with Hires fix, which runs your image through an img2img upscale process before saving it out. Hires fix will greatly improve the quality of your character images. Here are the settings I use:

[caption id="attachment_1659" align="aligncenter" width="1024"] Hi-Res fix settings[/caption]

Hires fix goes a long way to cleaning up your character without more effort on your part, but inpainting with the same seed as your initial generation using the various pose keywords at a moderate denoising strength (.4-.7) improves the quality of your final artwork, especially when you're using Control Net Open Pose for the initial generation. PRO TIP: Running Hires fix before inpainting gives you a higher quality base to work with so you can spend less time make manual tweaks.

To inpaint manually, send your generated character sheet (including prompt, negative prompt, and seed) to img2img, then use the brush to select each pose individually. Change the 4view keyword in your original prompt to the pose you're inpainting (t-pose, side profile, back side). Play with the denoising to improve results, or switch the seed. Switching the seed can reduce the consistency of your character at higher denoising levels, but it can also give you better results.

[caption id="attachment_1665" align="aligncenter" width="544"] Inpaint poses to improve details[/caption]

You can also use the ADetailer plugin to automatically inpaint your character for you. Use your preferred hand, person, and face Adetailer models. Start at a low denoising strength (.3-.4), then dial it up until you find a good balance between quality and consistency.

The LoRA version of the Character Maker works well at a very low weight. Dial the LoRA weight down to .3-.5 to keep all the functionality of Character Maker without overwhelming your characters with the Everly Heights art style.

[caption id="attachment_1666" align="aligncenter" width="1024"] Reduce the weight of the LoRA to dial back the Everly Heights style. (Enlarge)[/caption]

This model has been a work-in-progress since January 2023, and I'm constantly looking for weaknesses to improve in the next iteration. If you run into any problems, or have any feature requests, shoot me an e-mail at [email protected].

THE STORY STUDIO TOOLKIT

About EverlyHeights.tv

Everly Heights is the brainchild of Bill Meeks. His goal is to tell stories in the small town of Everly Heights, based on the small town in West Virginia where he grew up. Along the way, he'll teach you the latest AI tech and techniques he's using to bring his world to life in the YouTube series Building Dreams, as well as keep you informed in his semi-monthly news show Stable DiNEWSion.

If you'd like to support the project and development of future models like this one, please consider donating on Patreon.

If you don't want to donate, that's cool! Come join our Discord and share what you're working on.

Everly Heights Character Maker LoRA Family