Free Online Face Swap Models: Which One Should You Use?

Start with Balanced. It is the default (hyperswap_256_b), it uses a 256 px face crop, and it is the variant to try first on an ordinary portrait. Move to Robust for angle or hard light, Accurate only when both faces are sharp, and Baseline when a clear front-facing photo only needs a first pass.

That is the whole decision. The rest of this guide gives you the rule for each symptom, plus the file limits that decide whether a swap runs at all.

Key takeaways

• Four models, one workflow. Baseline is inswapper_128 at 128 px. Robust, Balanced and Accurate are hyperswap_256_a, hyperswap_256_b and hyperswap_256_c at 256 px. Balanced is the default.
• Crop size is not output size. 128 px and 256 px describe the aligned square the model receives, not the resolution you download.
• Two sliders, one layer. Edge blur runs from 0 to 1 with a default of 0.4. Fusion area runs from 0 to 50 with a default of 50. Both change the generated face's alpha after generation.
• Free, no account, no daily quota. The limits are technical: 6 swaps per minute per session, 12 per minute per network, one face per image, and the file caps below.

Decision rule: keep Balanced until you can name the symptom. Then change one control at a time.

Which model should you start with?

Display name	Model id	Face crop	Start here when
Baseline	inswapper_128	128 px	The photo is clear and front-facing, and you want the smallest crop
Robust	hyperswap_256_a	256 px	The head is turned, the light is uneven, or glasses and hair cross the face
Balanced	hyperswap_256_b	256 px	Default. A normal portrait, or a first attempt with an unfamiliar pair
Accurate	hyperswap_256_c	256 px	Both faces are sharp and large, and you will inspect eyes, lips and skin texture

Start with Balanced because it is the default and because 256 px gives the model more structure than the Baseline crop. On a clean, front-facing portrait, the other variants mostly add variance rather than information.

Rule: if you cannot describe what is wrong, do not change the model. A different variant will not fix a blurred source or a face that the detector cannot find.

When should you switch models?

What you see	Try	Why
A profile, a turned head, or one side of the face in shadow	Robust	It is the variant positioned for wider angles and harder light
Glasses, hair or a hand crossing the face	Robust	The same angle and occlusion case
A large, sharp face where you will check eyelashes and lip edges	Accurate	It is the detail-oriented variant
A clear, passport-style photo and a first pass	Baseline	Its 128 px crop is the smallest input
The whole result looks soft, not just the boundary	Fix the source photo first	No model recovers detail the source does not contain

Rule: change the photo before you try a third model. Angle, blur, compression and heavy filters set the ceiling for every variant, and Robust cannot invent a face behind a hand.

What do Edge blur and Fusion area change?

Control	Range	Default	What it changes
Edge blur	0 to 1	0.4	Softens the boundary of the generated face
Fusion area	0 to 50, integer	50	How much of the generated face layer is kept before blending; the product converts it to inverse padding

Both controls are applied in the browser after the generated face arrives. Moving either one does not start another generation. Only alpha changes: position and color are not re-aligned, so a slider cannot fix a misplaced eye.

Rule: use Edge blur for a visible seam and Fusion area for how much of the face layer survives. If the feature is in the wrong place, fix the source crop instead.

What does the tool check before the model runs?

Check	Limit	Where it applies
Picker format	JPG, PNG or WebP	Browser
Picker size	20 MB	Browser
Browser downscale	4096 px on the long side and 16 MP	Browser
API upload	6 MiB, 16 MP, 8192 px per side, one frame	Server
Faces	One face per image	Browser detector
Pace	6 per minute per session, 12 per minute per network	Server

These limits are product specifications, not measured results. Detection and alignment run in the browser before anything is uploaded. The detector uses MediaPipe's FaceLandmarker, configured for one face, and downloads its model from Google's storage on first use.

The public photo tool sends an aligned face crop. The product also defines a Full image mode that uploads the complete photos. In both cases, a file that passes the picker can still be refused by the API.

Rule: resize for convenience before the picker, not to dodge the API caps. The browser already scales large photos down to 4096 px or 16 MP, and a converted file still has to fit the server limits.

Why does a swap fail?

Message or symptom	Cause	Fix
"No face detected"	No clear single face in one of the images	Use a sharper, front-facing source and remove heavy filters
"Choose a valid JPG, PNG or WebP image under 20 MB"	Wrong type, empty file, or over 20 MB	Convert the format or reduce the file size
A softer result than the original	The browser scaled the photo to 4096 px or 16 MP before cropping	Start with a sharper source and inspect at full size
The API refuses the image	Over 6 MiB, over 16 MP, over 8192 px per side, or more than one frame	Export a single-frame image inside those caps
A model name is not available	The request used an id outside the enabled list	Pick one of the four models above
A rate-limit refusal	More than 6 swaps a minute from one session, or 12 from one network	Wait for the next minute; there is no daily swap count

Rule: read the error before changing the model. File, face and pace errors do not improve with a different variant.

How do you run a clean first swap?

1. Open the online photo tool and add the photo you want to change.
2. Add one clear source face with open eyes and no heavy filter. The detector needs visible landmarks.
3. Run it once with Balanced and inspect the result at full size.
4. Change one thing: the model if the structure is wrong, Edge blur or Fusion area if only the boundary is visible.
5. Download the result. For video, use the Windows app, because the online tool accepts still images only.

Rule: one change per run. Otherwise you cannot tell which control helped.

What do readers ask most?

Which model should I try first?

Balanced. Hold both photos constant, run it once, and only then compare Robust for angle and light or Accurate for a large, sharp face.

Is Accurate always better than Robust?

No. Accurate is the detail-oriented variant, while Robust is positioned for angle and difficult light. On a profile shot Accurate has less to work with, and a soft source limits both.

Why does the result still look wrong after changing models?

Because the source photo sets the ceiling. A small, blurred, filtered or strongly angled source limits every variant, so fix the source before trying a third model.

Can I use the online face swap on a phone?

Yes. It runs in a modern mobile browser with no installation. Detection and blending happen in the browser, while generation runs on a server, so a larger screen helps you inspect the result but is not required.

Do I need an account?

No. Photo swaps are free, with no account, no sign-in and no daily swap count. What remains are the technical limits above: the file caps, one face per image and the per-minute pace.

Rule: if changing models did not help, change the source photo. That is the variable every model shares.

Bottom line

Use Balanced first and switch only for a named symptom: Robust for angle and light, Accurate for sharp detail, Baseline for a clear front-facing photo. The sliders fix the seam, not the geometry, and the file caps decide whether a swap runs at all. If you are choosing between the browser and a desktop app, read the AIFaceSwap.io alternative comparison.