<h1 id="workflow-philosophy">Workflow Philosophy</h1><p>This workflow is designed as a <strong>reference implementation</strong>, not a one-click automation.</p><p>The objective is to expose every major stage of the image generation pipeline while keeping the workflow organized, modular, and easy to understand.</p><p></p><h1 id="reference-workflow-goals">Reference Workflow Goals</h1><ul><li><p>Educational and easy to understand.</p></li><li><p>Modular and easy to customize.</p></li><li><p>Validated on low-VRAM hardware.</p></li><li><p>Suitable as a foundation for building your own production workflows.</p></li></ul><hr /><h1 id="pipeline">Pipeline</h1><pre><code></code></pre><pre><code>Model Setup
      ↓
Character LoRAs
      ↓
Style LoRAs
      ↓
Prompt Guidance
      ↓
Sampling
      ↓
HiRes Fix
      ↓
Output</code></pre><hr /><h1 id="validated-reference-configuration">Validated Reference Configuration</h1><h3 id="checkpoints-validated">Checkpoints Validated</h3><ul><li><p> Pony Realism v2.2 </p></li><li><p> Pony Realism v2.3 Ultra </p></li></ul><h3 id="sampler">Sampler</h3><p>DPM++ 3M SDE</p><h3 id="scheduler">Scheduler</h3><p>Karras</p><h3 id="base-resolution">Base Resolution</h3><p>832 × 1216</p><h3 id="hires-fix">HiRes Fix</h3><ul><li><p> Upscaler: ESRGAN 4x+ Anime6B </p></li><li><p> Scale: 2× </p></li><li><p> Refinement Steps: 20 </p></li><li><p> Denoise Strength: 0.40 </p></li></ul><h3 id="clip-skip">Clip Skip</h3><p>2 <em>(Recommended by the model author.)</em></p><h3 id="vae">VAE</h3><ul><li><p> Embedded VAE </p></li><li><p> External SDXL VAE </p></li></ul><p>Both configurations have been validated with this workflow.</p><hr /><h1 id="lora-architecture">LoRA Architecture</h1><h2 id="character-stack">Character Stack</h2><p>The Character Stack is intended exclusively for <strong>identity-specific LoRAs</strong>.</p><p>During testing, I found that certain <strong>Illustrious Character LoRAs</strong> remain compatible with Pony Realism and can, in some cases, produce more realistic results.</p><p><strong>Personal observation:</strong> This is based on my own testing and may vary depending on the checkpoint, prompt, and LoRA combination.</p><p><strong>Example</strong></p><pre><code>&lt;lora:Illustrious\Character\Suzu_Airantou_ILL:1.0&gt;
&lt;lora:Pony\Character\Suzu_-_Airantou:1.0&gt;</code></pre><hr /><h2 id="style-stack">Style Stack</h2><p>The Style Stack is intended for LoRAs that influence rendering quality rather than character identity.</p><p><strong>Example</strong></p><pre><code>&lt;lora:Pony\Style\Expressive_H-000001:1.0&gt;</code></pre><p>Separating Character and Style LoRAs makes experimentation easier while preserving character identity across different visual styles.</p><hr /><h1 id="prompt-parser">Prompt Parser</h1><p>By default, this workflow uses <strong>CLIP Text Encode++</strong> with the <strong>A1111</strong> parser enabled.</p><p>This provides prompt parsing behaviour similar to <strong>AUTOMATIC1111 / Forge</strong>, including support for prompt weighting and multi-conditioning.</p><p>If you prefer native ComfyUI prompting:</p><ol><li><p> Open the <strong>CLIP Text Encode++</strong> node. </p></li><li><p> Change the <strong>Parser</strong> dropdown from <strong>A1111</strong> to <strong>ComfyUI</strong>. </p></li><li><p> Continue writing prompts using ComfyUI's native prompt syntax. </p></li></ol><p><strong>Note:</strong> Prompt behaviour may differ slightly between the A1111 and ComfyUI parsers, particularly when using prompt weighting or advanced prompt syntax.</p><hr /><h1 id="gpu-validation">GPU Validation</h1><p>Validated on:</p><ul><li><p> NVIDIA RTX 2080 (8 GB VRAM) </p></li><li><p> Low VRAM Compatible</p></li></ul><p></p>

Files with SHA256: d13ae821853c4dc0e729f16906ff5d88ebd4fd14b0ce14656e041e6282ae7f03