Repair a small face during generation Face fixing with GFPGAN

1. Prompt

   1 · Keep the face small in the frame

   The face fixer only helps when the face came out badly, and faces come out worst when they are small in the picture. This example asked for a full-body shot so there would be something to repair: a wide timber skyharbor walkway in bright morning daylight, a huge zeppelin airship envelope in the sky beyond it, a woman in a patched teal canvas coat standing at the railing far down the walkway, facing the camera, her face lit by the sun, (small distant figure:1.45), (wide establishing shot:1.3), the walkway stretching away, whole body visible from her boots to her head, wide angle 24mm photograph, deep focus, natural light, highly detailed, with close-up, portrait, headshot, bust, upper body, cropped, face filling the frame, large face, person near the camera, silhouette, backlit, dark, night, dusk, sunset, facing away, back turned, blazer, suit, indoors, forest, trees, cartoon, illustration, painting, lowres, blurry, deformed hands, extra fingers, watermark, text in the negative prompt.At that distance the model has very few pixels to put two eyes, a nose and a mouth into, and a small face is the first thing a picture loses.If the picture you want fixed is already made, stop here and repair it where it stands. This pass is asked for before a picture exists, so reaching for it now would mean making a different picture instead of mending the one you have.

2. Model & style

   2 · Pick a photoreal model

   This example used ICBINP - I Can't Believe It's Not Photography.GFPGAN was trained on photographs of faces, so it has the most to give a model that draws photographic ones, and less to give a drawn face. Whichever model you pick, it works on the face it finds.

3. Size

   3 · Set the size, and keep the same seed

   Size 512×768, a portrait shape at the model's native scale. Seed matters more here than in most examples, because this technique is two runs compared against each other, and everything but the face fixer has to stay identical between them.Sampling was k_euler_a, 30 steps, guidance 7.5, seed 1234567890. Type the seed in by hand before you go on, and keep it for both runs.

4. 4 · Generate once, untouched

   Press Generate now, with Post-processing left alone.What comes back is the first half of the comparison. Look hard at the face before going on, and at how small it is on screen: the next step is only meaningful measured against it.How much there is to see later depends on how badly this face came out. A model that drew it cleanly leaves the pass with little to do.

5. Finishing

   5 · Turn on the face fixer

   Now open Post-processing and set Face fix to GFPGAN, with Face fix strength at 1.A face fixer is a post-processor: it runs over the finished image, after the picture has been made, so it cannot move anything or change the composition, only redraw the faces it finds.The strength blends the restored face back over the original, so lower values keep more of the model's own. The two are set together, because the strength slider does not exist until a fixer is chosen.

6. 6 · Generate again, seed and all

   Press Generate a second time with everything else exactly as it was, the seed included.The second request is here to be compared and for no other reason. Two fields separate it from the first, so any difference between the two images comes from the repair pass.This is only for the comparison. In normal use you switch the face fixer on before the first request and send it once.Expect anything from a few seconds to a few minutes.

Everything here happens before the picture exists. If the one you want mended is already made, sending the request again would only hand you a different picture, so repair the face in the picture you have: Fixing a face in a finished picture