Illuminate AI — SDXL Modular Workflow

A modular SDXL ("Illustrious, Pony") workflow designed for high-quality image generation, concept exploration, detailed refinement, and flexible image-to-image editing.

The workflow is organized into clearly separated modules, allowing you to activate only the parts you actually need. This keeps the setup easier to understand while still offering advanced quality options.

Main Features

• SDXL text-to-image generation

• Fast Concept Batch with four variations

• Selectable concept image using Batch Index

• Two-stage HiRes refinement

• PAG applied only to the base generation pass

• Optional Face Detailer for improved facial quality

• Independent Image-to-Image pipeline

• Separate Img2Img checkpoint, CLIP Skip, LoRA loader, samplers, VAE path, and upscale model

• Two-pass Img2Img refinement

• Optional SeedVR2 ultra upscaling

• Built-in example prompts and workflow guidance

• Modular group controls for easier activation and bypassing

• Optimized for high-end GPUs such as the RTX 5090

Concept Batch

The optional Concept Batch generates four fast variations before running the full quality pipeline.

Use the SELECT CONCEPT IMAGE node to choose the preferred result:

• Batch Index 0 = Image 1

• Batch Index 1 = Image 2

• Batch Index 2 = Image 3

• Batch Index 3 = Image 4

Keep the length set to 1.

After generating the batch, change the sampler seed from randomize to fixed before selecting an image. Otherwise, a new batch will be generated during the next run.

Image-to-Image

The Img2Img section is fully independent from the main SDXL generation pipeline.

It includes:

• Separate checkpoint loader

• Separate CLIP Skip

• Separate Power LoRA Loader

• Separate upscale model

• Two Img2Img sampling passes

• Optional Face Detailer

• Optional SeedVR2 upscaling

The first sampling pass controls the strength of the main change. The second pass restores fine details and should use a lower denoise value.

Recommended first-pass denoise values:

• 0.25–0.35: minor color and detail changes

• 0.40–0.50: expression, hair, and lighting changes

• 0.50–0.65: clothing changes or stronger redesigns

• 0.70+: major reinterpretation

For best results, describe the complete desired final image instead of writing commands such as “remove jacket.”

Recommended Workflow

1. Generate four concept variations.

2. Select the preferred image using Batch Index.

3. Run the main SDXL generation.

4. Apply the HiRes refinement pass.

5. Enable Face Detailer if needed.

6. Use SeedVR2 only for the final ultra-resolution output.

7. Use the independent Img2Img pipeline when editing an existing image.

Requirements

This workflow may require the following custom node packages:

• rgthree-comfy

• ComfyUI Impact Pack

• ComfyUI Impact Subpack

• ComfyUI SeedVR2 Video Upscaler

• ComfyUI Easy Use

• TinyTerraNodes

Required models, checkpoints, LoRAs, detector models, and upscale models are not included.

Replace unavailable models with compatible SDXL alternatives installed in your own ComfyUI setup.

Notes

SeedVR2 Torch Compile is intentionally not required. On some Windows portable installations, Torch Compile can cause issues related to missing Microsoft C++ Build Tools or CUDA memory allocation.

Disabling compilation does not reduce image quality. It only removes the optional compilation speed optimization.

The workflow is intended as a flexible modular foundation and will continue to receive improvements and additional modules in future versions.