5-in-1 Multi-Model Realism Workflow

AETHERFORGE — MULTI-MODEL IMAGE GENERATION & DETAIL PIPELINE

Author: cyberonee

Contact: For corrections, improvements, bug reports, or suggestions regarding this workflow, please send a message to:

Email: [email protected]

A modular ComfyUI workflow designed for high-quality image generation, realism enhancement, upscaling, and targeted detail refinement across multiple model families.

SUPPORTED MODEL PATHS

• FLUX

• SDXL

• Pony 7

• Additional compatible model branches

CORE PIPELINE

1. Select the desired model branch.

2. Apply the model-specific LoRA stack.

3. Generate the base image.

4. Optionally apply DyPE for composition and prompt/detail preservation.

5. Upscale the image using the selected upscaler.

6. Apply targeted Eye and Hand Detailers when needed.

7. Apply additional enhancement/post-processing only when it improves the image.

IMPORTANT

• Different model families require different CLIP, LoRA, sampler, and prompt settings.

• Do not assume that a LoRA or setting optimized for FLUX will work equally well with SDXL or Pony.

• Keep Detailer denoise conservative to avoid changing the original anatomy or identity.

• Eye and Hand Detailers should only be enabled when the corresponding detector reliably identifies the target region.

• Upscaling and detail enhancement are optional stages; compare the result against the original before keeping them enabled.

• Avoid stacking multiple realism/detail LoRAs unless each one provides a clearly visible improvement.

• The goal is controlled enhancement, not aggressive regeneration.

MODEL BUS / LATENT BUS

Model and latent routing is organized through dedicated buses and reroute paths to keep the workflow modular and easy to expand.

QUALITY RULE

Always compare:

BASE → UPSCALED → DETAILED → FINAL

Keep an enhancement only when it produces a visible improvement without introducing:

• unwanted texture

• artificial sharpness

• altered facial features

• incorrect hands

• distorted eyes

• loss of identity

• excessive noise

• over-processed skin

This workflow is designed as a modular experimentation platform.

Individual branches can be enabled, disabled, replaced, or extended without rebuilding the entire pipeline.

DISCLAIMER

This workflow is continuously evolving. Model behavior, LoRAs, custom nodes, and recommended settings may change over time.

If you improve, modify, or extend this workflow, please consider sharing your improvements with the author so future versions can benefit from them.

© cyberonee