How to Make AI Images Look More Realistic in ComfyUI

AI-generated images often have a telltale glossy or overly smooth finish that immediately gives them away. If you've struggled with getting truly realistic results from your outputs, this two-step workflow can help transform those images into something far more convincing.

The Problem with Single-Pass Generation

Most AI images fail at realism because of subtle issues—artificial sheen on skin, oversaturated colors, or textures that look too perfect. Running a single upscale pass or using one model alone typically isn't enough to fix these problems.

The solution is combining two different approaches: one for heavy lifting on realism, and another for fine-tuning textures.

Step 1: Initial Realism Pass with Qwen Image Edit

The first step uses Qwen Image Edit (2509) with specialized LoRAs to transform your image's overall appearance. This is where the major work happens.

What you'll need:

• Qwen Image Edit (2509) model
• Qwen Image VAE and Text Encoder
• Qwen Image Edit 8-Step Lightning LoRA
• A realism LoRA: either Anything2Real Alpha or Anime2Realism (2509)

Which LoRA you choose matters. The Anything2Real Alpha LoRA works well for images that already lean toward photorealism but need refinement.

For anime or heavily stylized illustrations, the Anime2Realism LoRA produces significantly better results.

The workflow itself is straightforward. Load your source image, use a simple positive prompt like "change the image to a realistic photo," and let the LoRA handle most of the transformation.

The key settings: 12 steps in the KSampler with the Euler sampler and Simple scheduler.

You can optionally add ControlNet nodes if you need more guidance over the output, but they're not required.

Important note: Use LoRAs specifically trained for Qwen Image Edit work better than those trained on Qwen Image. So if you are experimenting with LoRAs, be sure to note this subtle difference.

Step 2: Texture Refinement with SDXL

After the initial pass, your image will look significantly more realistic, but the textures might still feel slightly off. This is where Stable Diffusion XL comes in for fine-tuning.

This step uses a minimal workflow: just the Juggernaut XL checkpoint with a Touch of Realism LoRA (or similar realism LoRA of your choice), paired with the Ultimate SD Upscale custom node.

The critical setting here is the denoise value—keep it very low, between 0.05 and 0.15. This preserves character likeness and general compositional elements.

You're not trying to remake the image, just enhance surface details. Use a straightforward prompt that describes what's in your image, like "photograph of a man smiling outside."

The results are subtle but impactful. Fabric textures become more convincing, skin pores and wrinkles look natural, and that artificial glossiness disappears.

Details like denim weaves, leather textures, and facial features get the refinement they need.

Here's a comparison of Qwen Image Edit vs. SDXL for texture differences:

Why Both Steps Matter

You might wonder: why not just run the original image through SDXL? The answer is that SDXL alone isn't capable enough for the initial transformation. Too much gets lost in translation when you try to take a raw AI image and push it through SDXL directly.

But when you combine these two models—Qwen Image Edit for the heavy lifting and SDXL for texture polish—the results are significantly better than either model alone.

Getting Started

To run this workflow yourself, you'll need:

• ComfyUI installed and running
• Qwen Image Edit model files (2509) and VAE/Text Encoder
• Qwen Image Edit 8-Step Lightning LoRA
• Both realism LoRAs: Anything2Real Alpha (good for general use) and Anime2Realism (good for converting illustrations to photos)
• The Juggernaut XL checkpoint with a realism LoRA
• Ultimate SD Upscale custom node (essential for step 2)
• Optional: ControlNet custom node for more control of the pose and output