The upscaler changed your subject's face
A generative upscale redraws identity. Keep face denoise at or below 0.35, mask the face out of a hotter pass, and check a 100 percent crop before you commit.
The face that comes back from an upscaler is not a sharper version of the face you sent. It is a new face, sampled from whatever the model thinks a high-resolution version of those pixels should look like.
That is why a hero still can look sharper at a glance and wrong at a client review. Pores appear, eye spacing shifts, the mouth is a different mouth. You did not ask for a new person. The pass was still a generation, so identity got re-decided with the fabric weave.
If the job is a landscape or a packshot without a label, this is usually acceptable. If the job is a specific person, it is the failure. Keep a denoise ceiling on the face, mask identity-bearing pixels out of a hot pass, and check a 100 percent crop before you treat the file as a master.
A generative upscale is a generation
Traditional resizing spreads the pixels you already have. A generative upscaler predicts the missing high-frequency detail and draws it. The result is genuinely sharper and genuinely invented.
Denoising strength is how much of the original the model is allowed to throw away. Low values leave the picture visible, so the model has little freedom. High values bury it, and you get a new picture that only shares a silhouette. Identity breaks first. Faces stop being the same person well before layout stops being the same layout.
That is the opposite of inpainting a broken hand. There you want a redraw: a 20 to 30 pixel mask margin and denoise in the 0.5 to 0.7 band, because the original fingers are not the thing you are keeping. A face you have already approved is geometry you must keep. Raising denoise because the upscale still looks soft does to the face what that hand pass is supposed to do to the hand.
Dedicated upscalers sit on the same trade. Topaz Upscale Image is billed as detail-faithful enlargement. Clarity Crystal Upscaler regenerates fine texture rather than stretching pixels. Faithful is the default when a specific person has to survive. Versely's upscale_image path takes a model and an image URL. It does not expose a denoise slider. The ceiling below is for any pass that does (img2img, tiled diffusion, an inpaint used as an upscale). When the slider is missing, the equivalent controls are model choice, a face mask, and compositing the original face back.
On video the same invention happens per frame, which reads as crawl once the clip is playing. Over-sharpening makes it worse: in Topaz, Sharpness above 50 amplifies the flicker the pass was meant to hide. Check upscaled footage by playing it, never by scrubbing.
The denoise ceiling for a face
Treat 0.35 as the working ceiling when a specific face has to survive a generative upscale. Below that, the original geometry still dominates and you are mostly adding texture. At 0.4 and above you are asking the model to redraw the person. The exact break point moves a little by model and by how many pixels the face occupies, so bracket rather than trusting one number: run 0.2, 0.3 and 0.35 on a crop of the face and pick the highest value that still looks like the same human.
That ceiling is an operator rule, not a published measurement. It follows from how denoise is defined (low keeps the original visible, high keeps only the shape) and from identity failing before layout. Do not borrow the 0.5 to 0.7 hand-inpaint band. Those values exist to discard the original anatomy.
| Pass | Denoise band | What you are asking for |
|---|---|---|
| Identity-preserving upscale | 0.15 to 0.35 | Texture and edge, same person |
| Expression or grade tweak | 0.3 to 0.45 | Small change, still the same face |
| Restyle that keeps composition | 0.45 to 0.6 | Same silhouette, new surface |
| Hand or defect inpaint | 0.5 to 0.7 | Replace the original pixels |
If the upscale is still soft at 0.35, you have a source problem. The face never contained those frequencies, and more denoise will invent them. Generate the portrait tighter, or crop the face, generate it at portrait scale, and composite it back. Whole-frame upscaling of a distant face is how you get a reconstructed identity that survives a thumbnail and fails a client.
Two other settings interact with the ceiling. Tile size: tiled diffusion denoises each tile independently, so a face split across a boundary can disagree with itself (one eye sharper, a cheek seam). Overlap tiles and feather the mask, or keep the whole face in one tile. Step count: very few steps at even moderate strength leaves grit, not preserved identity. If you lower denoise, do not also starve the steps.
When to mask the face out of the pass
Mask the face when any of these is true:
- The upscaler does not expose denoise, and the model you want is a texture-regenerating one.
- You need more than 0.35 of denoise on the rest of the frame (a soft background, a product box, a set that has to sharpen) and the face is already correct.
- The face sits on a tile boundary you cannot move.
- A previous unmasked pass already shifted identity, and you are trying to recover rather than re-roll the whole still.
The move is inpainting used in reverse. Instead of masking the defect, you mask the thing that must not change. Invert the face mask so the upscaler is allowed to work on hair, cloth, the room, everything except the identity-bearing pixels. Feather 15 to 25 pixels around the hairline, jaw and ears so you do not get a cut-out. Then, if the original face looks soft against the new texture, run a second pass on the face only, denoise at or below 0.3, with a prompt that names the person as they already are: age band, expression, gaze, light direction. You are asking for texture, not a new reading of the face.
Same woman, mid-30s, three-quarter view, window
light from camera left, closed mouth, sharp eyes,
natural skin texture, no change to identity,
no change to expression
That prompt restates the current state. It does not ask for "more beautiful" or "more detailed face," which the model will treat as a different person.
If you would rather not draw a mask, element re-render versus masked inpainting is the other fork: naming "the face" lets the model decide which pixels count, including shadows. For identity lock, the hard boundary is safer. A named re-render is right when you want the change to propagate.
The last option is mechanical: composite the original face back. Upscale the frame, keep the source face (or a low-denoise face-only pass), and blend. Grain and colour have to match or the composite reads as a sticker.
Check identity before you commit the render
Do not judge the upscale at the size it will be posted. Judge it at 100 percent on the surfaces identity actually lives on.
- Crop the eyes. Spacing, limbal ring, the catchlight. If the catchlight moved, the head turned.
- Crop the mouth. Tooth edge, lip thickness, the philtrum. This is where "same person, different day" usually is a different person.
- Crop the hairline and ears. Generative upscalers regularise. A widow's peak becomes a smooth line. An ear loses a crease.
- Flip between source and upscale. A two-second A/B on the same crop is more honest than staring at the new file. If you have to talk yourself into "it's close enough," it is not the same face.
- On video, play at 1x. A paused frame can look locked while the face crawls. Upscale the still first on the hero frame, then decide whether a whole-clip upscale is even the right job. Generating at full resolution versus upscaling the winner is the spend question. This check is the identity question, and it comes first.
Save the source. An upscale you cannot compare is an upscale you cannot undo. Compare on a crop before you run the same settings over a whole set. The photo editor is the right surface for the masked pass: protect the face, or instruct the edit to leave it untouched. Do not prompt "make it sharper" and hope the model knows which pixels are off limits.
FAQ
Why did the upscaler change the face if I didn't ask it to?
Because a generative upscale is not a resize. It predicts high-frequency detail that was never in the source. Faces are a high-scrutiny, low-pixel-share region, so the prediction is a new reading of the person rather than a recovered one. You did not have to ask. Invention is the method.
What denoise value should I use on a face?
Stay at or below 0.35 when the person has to remain the same person. Bracket 0.2, 0.3 and 0.35 on a face crop and pick the highest value that still matches the source at 100 percent. If the tool has no denoise control, pick a reconstruction upscaler, mask the face out, or composite the original face back. Do not use the 0.5 to 0.7 band you would use to replace a bad hand.
Should I mask the face on every upscale?
No. Mask when the rest of the frame needs a hotter pass than the face can survive, when tiles split the face, or when the upscaler has no strength control and tends to regenerate texture. A low-denoise whole-frame pass on a close-up is often enough. A wide shot where the face is small should not be upscaled for identity at all.
Can I fix a face the upscaler already changed?
Sometimes. Mask the new face and run a low-denoise pass toward the original, or composite the source face back if you still have it. If the source face was twenty pixels wide, there is nothing to restore. Generate that region at portrait scale and composite. Do not raise denoise again and hope the model wanders back.