Workflow Philosophy

This workflow is designed as a reference implementation, not a one-click automation.

The objective is to expose every major stage of the image generation pipeline while keeping the workflow organized, modular, and easy to understand.

Reference Workflow Goals

Educational and easy to understand.
Modular and easy to customize.
Validated on low-VRAM hardware.
Suitable as a foundation for building your own production workflows.

Pipeline

Model Setup
      ↓
Character LoRAs
      ↓
Style LoRAs
      ↓
Prompt Guidance
      ↓
Sampling
      ↓
HiRes Fix
      ↓
Output

Validated Reference Configuration

Checkpoints Validated

Pony Realism v2.2
Pony Realism v2.3 Ultra

Sampler

DPM++ 3M SDE

Scheduler

Karras

Base Resolution

832 × 1216

HiRes Fix

Upscaler: ESRGAN 4x+ Anime6B
Scale: 2×
Refinement Steps: 20
Denoise Strength: 0.40

Clip Skip

2 (Recommended by the model author.)

VAE

Embedded VAE
External SDXL VAE

Both configurations have been validated with this workflow.

LoRA Architecture

Character Stack

The Character Stack is intended exclusively for identity-specific LoRAs.

During testing, I found that certain Illustrious Character LoRAs remain compatible with Pony Realism and can, in some cases, produce more realistic results.

Personal observation: This is based on my own testing and may vary depending on the checkpoint, prompt, and LoRA combination.

Example

<lora:Illustrious\Character\Suzu_Airantou_ILL:1.0>
<lora:Pony\Character\Suzu_-_Airantou:1.0>

Style Stack

The Style Stack is intended for LoRAs that influence rendering quality rather than character identity.

Example

<lora:Pony\Style\Expressive_H-000001:1.0>

Separating Character and Style LoRAs makes experimentation easier while preserving character identity across different visual styles.

Prompt Parser

By default, this workflow uses CLIP Text Encode++ with the A1111 parser enabled.

This provides prompt parsing behaviour similar to AUTOMATIC1111 / Forge, including support for prompt weighting and multi-conditioning.

If you prefer native ComfyUI prompting:

Open the CLIP Text Encode++ node.
Change the Parser dropdown from A1111 to ComfyUI.
Continue writing prompts using ComfyUI's native prompt syntax.

Note: Prompt behaviour may differ slightly between the A1111 and ComfyUI parsers, particularly when using prompt weighting or advanced prompt syntax.

GPU Validation

Validated on:

NVIDIA RTX 2080 (8 GB VRAM)
Low VRAM Compatible