Face Fix in SwarmUI: Sharpening Faces Without ADetailer

What is face fix, and why do you need it?

If you've used CivitAI's built-in generator, you'll have seen the Face Fix option -- a post-process step that detects faces in your image and runs a dedicated restoration pass on just those regions, sharpening features, correcting softness, and cleaning up compression artifacts.

SwarmUI doesn't have a one-click face fix button, but you can achieve the same result using a combination of seed locking and a face-trained upscale model in the Refiner. It takes two generation passes but gives you full control over how aggressively the face is refined.

---

How it works

The technique relies on a simple fact: the same seed always produces the same image. So the workflow is:

1. Generate your image and note the seed
2. Re-generate with the same seed to reproduce the identical output
3. On that second generation, enable the Refiner with a face-focused upscale model
4. The refiner sharpens and restores face detail on top of the already-correct composition

You're not guessing -- you know exactly what base image the refiner is working from.

---

Step 1: Note your seed

After a generation you're happy with, check the seed displayed in the interface. Either copy it manually or click the lock icon to pin it before your next generation. With the seed locked, every subsequent generation will reproduce the same image.

---

Step 2: Choose your model based on image style

The right model depends on the style of image you're working with. Using a photo-trained model on illustrated output (or vice versa) will give poor results -- mismatched texture, over-sharpened linework, or an uncanny push toward realism.

Photorealistic and semi-realistic images

Use 4xFaceUpDAT by Helaman, a 4x upscaler trained specifically on the FaceUp dataset of real photographic faces. It comes in four variants:

Model	Best for
4xFaceUpDAT	Good quality input -- the default, start here
4xFaceUpLDAT	Lower quality or noisy input, includes denoising
4xFaceUpSharpDAT	Good quality input, produces a sharper result
4xFaceUpSharpLDAT	Lower quality input with sharper output

Download: https://openmodeldb.info/models/4x-FaceUpDAT

Illustrated, anime, and AI-generated art styles

For Illustrious, Pony, SDXL illustrated output, or any non-photorealistic style, use 4x UltraSharpV2 by Kim2091 instead. It was trained on a private dataset covering realistic images, anime, cartoons, and illustrations, and handles all of them well. It also includes deblur and compression removal, making it a strong general-purpose choice for AI-generated output.

Download: https://openmodeldb.info/models/4x-UltraSharpV2

---

Place whichever file you download into:

SwarmUI/Models/upscale_models/

Then do a Server > Refresh Models so SwarmUI picks it up.

---

Step 3: Configure the Refiner

Open the Refine / Upscale panel in the Generate tab and set the following:

Setting	Value
Refiner Control Percentage	0.15 -- 0.25
Refiner Method	Post-Apply (Normal)
Refiner Upscale	1.0 (no size change) or 1.5 if you also want to upscale
Refiner Upscale Method	4xFaceUpDAT (realistic) or 4x-UltraSharpV2 (illustrated/anime)
Refiner Model	(Use Base)

The key difference from a general upscale pass is the Refiner Control Percentage -- keep it low. You want the model to sharpen and restore, not to redraw. At 0.15--0.25 the composition and likeness are preserved while facial detail is tightened up.

---

Step 4: Re-generate

With the seed locked and the Refiner enabled, hit Generate. The output should be compositionally identical to your original, with noticeably improved face clarity.

---

Important caveat: whole-image vs. face-only

CivitAI's face fix uses face detection to isolate and process just the face region. The SwarmUI technique described here applies the upscale model to the entire image, not just the face. This means:

• Non-face areas (clothing, background, hair) may also be subtly sharpened or textured
• Keeping control percentage low (0.2 or under) minimises unwanted changes outside the face
• For images with small faces at distance, the effect will be less pronounced than on close portraits

For most generated images this is a non-issue, but it's worth being aware of if you're working with complex multi-figure compositions.

---

Quick reference

Seed:                       [lock your original seed]
Refiner Control Percentage: 0.2
Refiner Method:             Post-Apply (Normal)
Refiner Upscale:            1.0 (or 1.5 to also upscale)
Refiner Upscale Method:     4xFaceUpDAT (realistic) / 4x-UltraSharpV2 (illustrated/anime)
Refiner Model:              (Use Base)