Main prompt tags: coltbronco, centaur, solo, police uniform, necktie, belt, police hat, police badge, necktie, name tag, walkie-talkie <------ should always be included
Style control (IMPORTANT PLEASE READ):
for 2D style you MUST put 3d prompt into negatives
for 3D style use 3d in positives
Note: This is a 3d -> 2d LoRA. The goal was to decouple the character from the 3d render style. While you can generate images in the style of the movie its not recommended. You will probably need a good Illustrious checkpoint that can support a 3d pixar style. WAI isn't the best at 3D.
Recommended samplers:
- Euler Ancestral Scheduler: Normal CFG: 6.0 Steps: 20 Lora strength: 1.0
- DPM++ 2M SDE Scheduler: Karras CFG: 6.0 Steps: 20 Lora strength: 1.0 <--- better for 3D
Prompt tips (green - positive, red - negative):
- bipedal - discovered this new negative prompt which helps guide models to draw him correctly with 4 legs
- police uniform, necktie, belt, police hat, police badge, necktie, name tag, walkie-talkie, aviator sunglasses or wearing aviator sunglasses - full clothing
- (walk-talkie:0.8) - lower the weight if you get walkie-talkies on both sides of his shoulders
- (belt:0.8) - lower the weight if you get too much belt elements (e.g. around his centaur torso area)
- be mindful of using horse related tags - it can cause him to morph into a literal horse
The LoRa should support all kinds of Illustrious checkpoints. Tested working well with the following:
WAI-NSFW-illustrious-SDXL v12.0
The LoRa can be used with other Illustrious Centaur Loras to help stabilize the concept of centaurs further. I testing the following (at 0.75 weight):
To improve details:
- generate using higher base resolution (e.g. 1216x1216) or use latent upscaling / highres fix - especially when trying to do full body poses
- for fixing wonky face/eyes I use FaceDetailer node (comfyui) + Adetailer Face finder model
Last image in the gallery can be imported into ComfyUI with a basic workflow:
No custom nodes required anymore! Workflow includes a HiresFix pass with x1.3 latent upscaling by default.
For portraits/upper body shots I recommend lowering the upscaling ratio ('Upscale Latent By' node) - otherwise you might get distorted details.