Face swap? Cute. Viggle wants to replace the ENTIRE person.
Meet Viggle Image Swap, a new LoRA for Qwen-Image-2.1 that takes a character sheet and replaces someone in an existing photo.
Not just the face. Hair, body, and clothing too, while trying to preserve the original pose, expression, camera angle, lighting, and background.
And the interesting part is how much control you get.
Here's how it works:
Character sheets: Start with a 1536 × 1024 reference image. A portrait goes on the left, with front, side, and back views on the right.
Full character replacement: Give it a source photo and identify the person you want to replace.
Group photos: Use --left to target specific people and --kept to leave others untouched.
Local inference: Run the included inference/swap.py script using Hugging Face Diffusers.
And yes, you can use a real face photograph in the portrait section of the character sheet, provided you have the necessary permission.
Now for the technical stuff.
Under the hood, Viggle Image Swap uses a rank-256, bf16 LoRA running on QwenImage21Pipeline.
Standard inference takes just SIX denoising steps, with CFG set to 1.0.
There's also a --turbo mode.
It runs two steps with the swap model, then switches to Qwen-Image-2.1-Turbo for another six steps to produce sharper details.
And the developers made a rather clever optimization.
They replaced a heavy Conv3d patch-embedding operation in the vision encoder with equivalent matrix multiplication.
The result?
Approximately 30 seconds saved per input image, according to Viggle.
No changes to the output. Just less unnecessary computation.
But here's a small reality check.
Viggle reports generation times of approximately 2.2 seconds per image on an NVIDIA B200.
That's impressive.
It also requires around 46 GB of peak GPU memory in the benchmarked configuration.
So don't expect exactly the same performance from your gaming laptop.
There are also some limitations.
The original person's clothing can occasionally survive the swap. Poses may drift, and Turbo sometimes brings back unwanted details from the source image.
Still, I really like where this is going.
Instead of generating an entirely new scene and hoping the composition stays consistent, you can preserve an existing image and replace individual characters.
For visual storytelling, character concept art, and compositing experiments, that's a much more controlled workflow.
And with support for multiple people, it becomes even more interesting.
Imagine replacing the entire cast of a photograph while keeping the original composition.
Hollywood, but with considerably fewer casting meetings.
One important detail: the model weights use the Qwen Research License, which limits use to research and evaluation. Commercial applications require separate permission. The inference code itself is Apache 2.0.
Viggle Image Swap, model weights, examples, and inference code:
https://huggingface.co/Viggle/Viggle-Image-Swap
#Viggle #ImageSwap #QwenImage #LoRA #AIImageEditing #CharacterDesign #GenerativeAI
YouTube / ABVX
