AI Model Recommendations

This documentation is for update 2.0+, applied to 5.4+.

This guide provides a list of recommended AI models, VAEs, ControlNets, and LoRAs specifically for generating high-quality, realistic PBR textures with the AI Generation Job. The right combination of these assets can significantly improve the detail, realism, and consistency of your materials.

Note: Workflows and models are generally not cross-compatible between Stable Diffusion 1.5 (SD 1.5) and SDXL. Ensure the models you choose match the base model of your selected workflow.

1. Main Models (Checkpoints)

The Main Model is the foundation of your generation. For PBR textures, you want a model that excels at photorealism and fine detail.

Juggernaut XL

Model Name	Model Type	Why it's Recommended for PBR	Where to Find It
Realistic Vision v6.0	SD 1.5	Excellent photorealism and fine details.	Civitai
DreamShaper	SD 1.5	Versatile, great for complex prompts.	Civitai
Juggernaut XL	SDXL	Top-tier SDXL model for detail.	Civitai
RealVisXL	SDXL	Maximum realism for photorealistic textures.	Civitai

1.1 Notable Mentions for Stylized Textures

While the models above focus on realism, the following checkpoints are excellent for creating stylized, hand-painted, or non-photorealistic (NPR) textures suitable for games like Valorant, Fortnite, or Genshin Impact.

Stylized

Realistic

Model Name	Model Type	Why it's Recommended for Stylized Textures	Where to Find It
Rev Animated	SD 1.5	Clean, vibrant, illustrative style.	Civitai
ToonYou	SD 1.5 / SDXL	Cartoonish or Ghibli-esque style.	Civitai
Animagine XL	SDXL	High-quality, polished anime style.	Hugging Face

2. Custom VAEs

A VAE (Variational Autoencoder) translates the AI's internal "latent" image into the final pixel image. A good VAE prevents washed-out colors and enhances fine details.

VAE Name	Model Type	Why it's Recommended	Where to Find It
vae-ft-mse-840000-ema-pruned	SD 1.5	Industry standard; prevents washed-out colors.	Hugging Face
sdxl_vae.safetensors	SDXL	Official VAE for best SDXL results.	Hugging Face

3. Upscale Models

An upscaler increases the resolution of your generated image. These models are generally compatible with both SD 1.5 and SDXL outputs.

1024x1024

2048x2048

Upscaler Name	Model Type	Why it's Recommended	Where to Find It
4x-UltraSharp	Universal	Adds significant sharpness and detail.	Civitai
4x_NMKD-Siax_200k	Universal	Popular, high-quality, clean results.	OpenModelDB
SwinIR_4x	Universal	Natural detail restoration for organics.	OpenModelDB

4. ControlNet Models

ControlNet models are essential for guiding the generation process to create specific structures. You must use ControlNet models that match your base model version (SD 1.5 or SDXL).

SDXL Depth Control Net

ControlNet Models (for SD 1.5)

ControlNet Model	Use Case for PBR
control_v11p_sd15_normalbae	Essential for 3D shape/depth.
control_v11f1p_sd15_depth	Controls height differences (walls, tiles).
control_v11p_sd15_canny	Defines hard outlines (grout, planks).
control_v11p_sd15_softedge	Defines soft shapes (fabric, grain).
You can find the official SD 1.5 ControlNet 1.1 models on Hugging Face.

ControlNet Models (for SDXL)

ControlNet Model	Use Case for PBR
control-lora-depth-rank256	Controls height and structure.
control-lora-canny-rank256	Enforces strong structural lines.
control-lora-recolor-rank256	Guides color for patterned textures.
You can find these and other official SDXL ControlNet LoRAs on Hugging Face.

5. LoRAs for PBR Texturing

LoRAs (Low-Rank Adaptations) are small files that inject a specific style or concept. They are not cross-compatible; a LoRA made for SD 1.5 will not work with an SDXL model, and vice-versa.

LoRA Name	Model Type	Why it's Recommended	Where to Find It
Add More Details	SD 1.5	Enhances micro-details and sharpness.	Civitai
epiCRealism Helper	SD 1.5	Guides models towards photorealism.	Civitai
SDXL Detail Tweaker	SDXL	SDXL version of "Add More Details".	Civitai
Texture-Specific LoRAs	Both	Improves results for specific materials.

You can find these and many more LoRAs on websites like Civitai. Remember to check which base model (SD 1.5 or SDXL) they were trained on.