Blazing Fast Face Detailer Workflow for ComfyUI

First of all, this workflow seems very complicated, but it's actually very easy and you barely have to tweak any of the nodes!

Face Enhancer Ultra is a precision upscaler/enhancer workflow for ComfyUI that targets only the faces in an image. Instead of re-generating the entire picture, it detects faces automatically, creates accurate masks, crops them for high-resolution enhancement, and blends them seamlessly back into the original.

It’s ultra fast because it works only in the low sigma range, which is the final stage of the diffusion process where the image is already well-formed. By skipping the high sigmas, the model focuses entirely on adding fine detail rather than reshaping the image, allowing faces to be enhanced in just a few steps without altering their original structure.

To run Face Enhancer Ultra, you will need the latest version of ComfyUI, either installed locally or on a cloud platform, and a GPU with at least 8GB of VRAM. Lower VRAM cards can still run the workflow by reducing the crop resolution in the face size node and/or using a smaller model such as the GGUF version of Flux Krea Blaze.

Below is a table with download links for the workflow and all required models. Place each file in the correct ComfyUI folder before running the workflow.

This workflow uses several custom nodes, and these do not download automatically. The simplest way to install them is directly inside ComfyUI after loading the workflow.

First, open ComfyUI and load the workflow file. If you see any nodes outlined in red, it means those nodes are missing from your installation. You can install them through ComfyUI’s manager:

• Click the Manager button in the ComfyUI interface.

• Choose Install Missing Custom Nodes.

• A list of missing nodes will appear — click Install on each one and select the latest version.

• When all are installed, restart ComfyUI to register them.

After restarting, reload the page and the red outlines should be gone. You only need to install these nodes once.

💡Note: The bbox detector model and the SAM detector model will download automatically the first time you run the workflow.

In this stage, the workflow automatically detects the face, builds a mask, and crops the area for enhancement. BBOX Detector finds the face, SAMDetector converts the detection into a detailed segmentation mask, and the mask is refined before resizing for processing. While this runs smoothly out of the box, a few settings let you fine-tune results:

• BBOX Detector Threshold – Lower if a face isn’t detected; raise to reduce false positives.

• SAMDetector Dilation – Expands the mask to include more of the surrounding area, like hairlines or ears.

• Mask Blur – Softens mask edges for a smoother blend into the original image.

• Face Size – Default 1024 for maximum detail; lower values reduce VRAM use and speed up processing.

If the face is small or far from the camera, keep mask blur and dilation low to avoid over-blending. For close-up portraits where the face fills more of the frame, increasing these slightly will help create a smoother, more natural transition between enhanced and original areas.

Model Selection & Loading

Once the mask and crop are ready, the face is enhanced using the Flux Krea Blaze model. Use the FP8 version for maximum quality, or the GGUF version if you need lower VRAM usage. The key difference is that FP8 loads through the standard diffusion model loader, while GGUF requires the UNET loader. Make sure to also select the correct CLIP models and VAE.

💡Tip: The FP8 version offers slightly better detail but uses more VRAM. The GGUF version runs on lower VRAM cards with a minor trade-off in sharpness.

Prompting for Flux Krea Blaze

Flux Krea Blaze only accepts a positive prompt — there’s no negative prompt field. This means you should focus on clearly describing the look and details you want to add to the face. Keep prompts short and precise to avoid unwanted style changes. For example:

ultra realistic, sharp eyes, detailed skin texture, natural lighting

Avoid long descriptive chains or unrelated terms, as Flux Krea Blaze will interpret everything you give it.

Tip: Start with a simple 4–6 word prompt and adjust wording between runs to fine-tune the enhancement style.

Denoise & Latent Noise Injection

Denoise strength controls how much the face is changed, while latent noise adds subtle variation to avoid an overly smooth or artificial look. Since this workflow operates only in the low sigma range, even higher denoise values will keep the original face structure intact.

Optional – max_shift: In the ModelSamplingFlux node, the max_shift parameter controls how far the sampler can deviate from the base signal. Lower values give more consistent, predictable results, while higher values (up to 2) can slightly boost detail and variation. Use with caution — pushing it too far can cause artifacts. I recommend leaving this at 2.

Tip: For portraits that already look decent, use low denoise (0.1–0.3) with minimal latent noise (0.05–0.1) to sharpen details without introducing new artifacts.

Sampler, Scheduler & Steps

For Flux Krea Blaze, the following settings are recommended for fast, high-quality face enhancement in the low sigma range:

• Sampler: uni_pc

• Scheduler: sgm_uniform

• Steps: 4 (enough to refine detail without changing structure)

• Sigmas: Automatically calculated

These settings allow Flux Krea Blaze to quickly add fine detail without unnecessary computation. If you choose to use a different model, check its download page for the recommended sampler, scheduler, and step count, as these can vary significantly between models.

Once enhancement is complete, the workflow automatically pastes the improved face back onto the original image. This is done seamlessly, using the blurred mask edges for smooth blending. The final image is saved in your output folder with the filename prefix set in the SaveImage node, so no manual editing or compositing is needed.

You can also easily compare the before and after images in the Image Comparer node by rgthree.

Here are a few before-and-after comparisons showing the results of Face Enhancer Ultra. In each case, facial details such as skin texture, eyes, and lighting are noticeably improved, while hair, background, and clothing remain untouched. The enhancements are subtle enough to look natural, but strong enough to bring out definition that was missing in the original.

Example 1 - Full Body Portrait

Example 2 - Medium Shot

Example 3 - Medium Close-Up

If a face is not detected, try lowering the BBOX Detector threshold to make the detector more sensitive. Sometimes increasing the Dilation in the BBOX detector also helps.

When the mask cuts off too close to facial features, increasing the SAMDetector dilation value will capture more of the surrounding area, such as hairlines and ears.

Visible blend edges can usually be fixed by raising the mask blur slightly for smoother transitions.

For systems with low VRAM, reduce the face size resolution in the cropping stage to prevent OOM errors — this will slightly lower detail but keep the workflow running. You can also switch from the FP8 model to a GGUF version of Flux Krea Blaze, or use a compatible SDXL model, both of which use less memory while still delivering strong results.

This Face Enhancer Workflow delivers fast, targeted face restoration in ComfyUI by combining automated detection, precise masking, and low-sigma enhancement with the Flux Krea Blaze model. By refining only the facial area, it keeps the rest of the image untouched while adding sharpness, texture, and clarity where it matters most.

Whether you’re polishing AI-generated portraits or restoring real photographs, this workflow offers a reliable and efficient way to upgrade facial detail in seconds. With the right balance of denoise strength, latent noise, and mask settings, you can achieve results that look natural yet noticeably improved.