A Face-Swap Experiment: Replacing the Real Person with Hina

A Test of Face Swapping Starting Point: DE JuN〈Just how outrageous is today's AI face-swapping technology?〉—just one full-face photo can completely transform a person. We tested three different methods using our character, Hina. October 15, 2026 That's not a face swap—it's a complete redraw of the entire person. Face-swap frame onlyDeep-Live-Cam·FaceFusionIn real time; does not work on anime-style faces Full-Body + Light and ShadowWan Animate·KlingOffline: A few seconds of footage plays for several minutes The Entire SceneDecart Lucy·StreamDiffusionReal-Time Streaming Redraw Pink = areas that were replaced. In the video, the wrinkles in the clothes and the lighting all match up, so these fall into the latter two categories. Measuring Three Paths with the Same Ruler Driver ChipLTX-generated mannequin, 5 seconds (no real human face) Reference ImageHina's design sketch, cropped to the waist Criteria (Sort First)Action > Like Hina > FPS > Flaws MachineRTX 3090 24GB / Cloud Route Real-time A complete transformation The Price Results C Decart Lucy 2.5 (Cloud) Yes Yes $0.02 per second, screen goes dark Works B Qwen+Wan Animate 2 (Local) No Yes 5-Second Clip, 6 Minutes Best picture quality A StreamDiffusion (Local) It's ~10 fps No 512×512, just a filter Deleted BEdit the first frame first, then make it move First Frame of the Original Film →Qwen Edited Image 51 seconds Replace the person with Hina; the lighting in the room remains unchanged. Wan Animate 2: Motion capture based on the original footage (314 seconds). Original footage on the left, animated version on the right.Wan Animate 2: Motion capture based on the original footage (314 seconds). Top: Original footage; Bottom: Animated version Realistic version. Flaws: The hair color has been bleached to a golden shade; the hair clip disappears when she turns her head. The model does not have a "Keep Background, Replace Person" mode → First, hide the background in the reference image (by editing the first frame), then pass it to the animation model. CA reference image; switch players instantly The image on the left is from the camera; the one on the right is a real-time output from Lucy 2.5 (generation begins about 3 seconds after connection). The screen for free accounts displays a moving “AI Generated” watermark.The top image is from the camera; the bottom shows real-time output from Lucy 2.5 (generation begins about 3 seconds after connecting). The screen for free accounts features a moving “AI Generated” watermark. Real-time, full-body replacement—including lighting and shadows—is currently only possible in the cloud. The cost: $0.02 per second for 720p video, which must be uploaded to the cloud. Document verification is used solely to remove watermarks; the API terms permit Decart to use the input and output data to train models. AOn-device processing is currently limited to filters only. Weak noise reduction (9.8 fps): It keeps up with the gestures, but the hoodie and the room are still visible. Strong Noise Reduction + Pose ControlNet (5 fps): Like Hina, the motion is lost The img2img pipeline is like a seesaw: you can only choose one side—either the action or the character. It performs a style conversion, not a character swap. Delete it after 6 rounds of adjustments. Take your pick Want to Select Instantly, the whole person is replaced C Cloud Lucy (Video Streaming, Billed by the Second) Best picture quality B Local WAN (non-real-time) The Real Deal, Zero Cost 3D/2D Models + Camera Motion Capture (Hina and Hanano already have these) Real-time + Open Source + Local None at this time Still to do: Run B using my own video, create reference images for "Hua Ye," and test Lucy using live-action footage. Wan-Animate Paper Decart Price List Decart API Terms daydreamlive/StreamDiffusion

Route	Real-time	A complete transformation	The Price	Results
C Decart Lucy 2.5 (Cloud)	Yes	Yes	$0.02 per second, screen goes dark	Works
B Qwen+Wan Animate 2 (Local)	No	Yes	5-Second Clip, 6 Minutes	Best picture quality
A StreamDiffusion (Local)	It's ~10 fps	No	512×512, just a filter	Deleted

Want to	Select
Instantly, the whole person is replaced	C Cloud Lucy (Video Streaming, Billed by the Second)
Best picture quality	B Local WAN (non-real-time)
The Real Deal, Zero Cost	3D/2D Models + Camera Motion Capture (Hina and Hanano already have these)
Real-time + Open Source + Local	None at this time