What happened
Viggle Image Swap is out, and it does one job very well: it replaces the person in a photo — face, hair, build, clothes, the whole package — with a character you hand it, while leaving the pose, facial expression, camera angle, lighting and background exactly where they were. Viggle built its reputation swapping people in video. This release drags that trick into the world of still images and ships it as a LoRA for Qwen-Image-2.1.
A quick translation, because LoRA sounds scarier than it is. A LoRA is a small add-on file that teaches a big image generator one new move. Picture a clip-on lens for a camera you already own: the camera stays the same, the shot doesn't.
Character sheet in, swapped photo out
The input is what Viggle calls a character sheet: a 1536x1024 canvas split with intent. The left third is a large close-up portrait. The right two-thirds hold a full-body, three-view lineup — front, side and back. That's the reference the model studies before it touches your picture.
Good news for anyone who isn't an illustrator: if your character is a real person, you don't have to build the whole sheet. Drop a headshot into the portrait slot and you're done. The rest — hair, build, outfit — gets pulled from the target photo's own context.
What comes back isn't a face swap. It's a whole-person swap.
Six steps, no CFG, and a turbo button
The numbers here are lean in a way that matters. The adapter was trained in bfloat16 (a compact number format that saves memory) at LoRA rank 256 with alpha 256 — in plain terms, a wide, high-capacity adapter rather than a thin one. It runs on the QwenImage21Pipeline inside the Diffusers library.
Then there's the speed. Generation takes 6 sampling steps with no classifier-free guidance — CFG scale 1.0 — on a custom sigma schedule. CFG is the dial that normally forces a model to obey your prompt, and it roughly doubles the work at every step, so switching it off is a big slice of the savings. Add --turbo and the model runs 2 opening steps on the base network, then passes the latent (its compressed internal sketch) to Qwen-Image-2.1-Turbo for 6 more. That second pass is where the micro-contrast and the razor-sharp fabric and skin texture come from.
Group photos: swap one person, keep the rest
Group shots work too. Two flags — --left and --kept — handle spatial positioning, so you can replace one or several people in a scene and leave everyone else in the frame untouched. One person becomes someone else; the rest of the wedding party, the team photo, the band, stays put. That single feature is the difference between a novelty and something you'd actually use on a client's picture.
A 30-second shortcut that ships with it
The open repository includes an optimized inference/swap.py script. The authors replaced the heavy Conv3d patch-embedding step in the visual encoder with an equivalent matrix multiplication (matmul), which trims roughly 30 seconds of pure prep time from every input image. It's not glamorous, but if you're working through a whole folder of photos, those saved seconds pile up fast. The weights are already live as viggle-image-swap-v0.1-lora.safetensors, and the release has been picked up by outside coverage (AlphaSignal) as another sign that image generation is getting dramatically cheaper to run.
What it means for you
Strip the jargon away and this is a tool that answers one question: what would this photo look like with someone else standing in it — same light, same pose, same place? Here's where that actually comes in handy.
At home: rescue the photo nobody loves
Someone blinked. Someone scowled. Someone's ex is lurking at the edge of the frame. Instead of reshooting or giving up, you swap the problem person for a character sheet and keep the shot everyone else looks great in. The lighting stays real because it is real — the model works around what's already in the picture.
At work: the stand-in becomes the real presenter
Plenty of teams shoot marketing photos with whoever happens to be in the office that day, then face a scheduling nightmare getting the actual spokesperson back in front of a camera. Now the stand-in holds the pose, and the real face gets dropped in afterward. You'll want written permission from the person whose likeness you're inserting, obviously.
For business: one model, an entire catalog
E-commerce folks pay a fortune to have the same model photographed in outfit after outfit, shoot after shoot. With a character sheet built once, the outfit changes and the person doesn't — and the pose, framing and studio lighting from the original product shot survive untouched.
For students: turn a name into a face
History and literature revision gets stickier when there's a visual to hang it on. A character sheet plus a period-appropriate photo gives you a consistent face for a figure you're studying, and the pose and expression of the original image keep it grounded instead of looking pasted together.
For creators: your character, every scene
Comic artists and story writers spend hours redrawing the same character from new angles. If you've got a ref sheet, this does the heavy lifting on background shots, crowd scenes and cover mockups — without your lead character changing jawline between panels.
For income: consistency at volume
Freelancers selling custom portraits, avatar packs or branded content live or die on consistency. One character sheet, many source photos, and the person stays recognizably the same across every delivery. And if you want to poke at AI image tools before committing to a local setup, there's a free collection at mykreatool worth a browse.
How to try it right now
The free route first: the model weights themselves cost nothing.
1. Download the weights. Head to Viggle Image Swap on Hugging Face and grab viggle-image-swap-v0.1-lora.safetensors.
2. Load the pipeline. Use QwenImage21Pipeline from the Diffusers library, in bfloat16.
3. Build your character sheet. A 1536x1024 canvas: left third, one large portrait; right two-thirds, full body from the front, the side and the back. Real person? Just paste a headshot into the portrait slot.
4. Run the script. inference/swap.py from the repo does the swap. It already includes the matmul fix, so you're not waiting around an extra 30 seconds per image.
5. Flip on --turbo. Two base steps plus six turbo steps for the sharper texture pass. Skip it if you want the plain 6-step result.
6. Working on a group shot? Use --left and --kept to say which people get swapped and which ones stay exactly as they are.
No GPU and no appetite for a terminal? Watch the API side — gptproto has been covering hosted access to Viggle tooling, which is the usual path for people who'd rather not install anything.
Upsides and what changes
The big win is that this moves character consistency out of the video-only bucket. Swap-the-presenter used to mean re-running footage; now a single still can be redone in seconds. Six steps with CFG off means you're not paying for a render farm either — this is the kind of thing that runs on modest hardware and quickly becomes a background step in a bigger workflow.
Preserving pose, expression, angle, lighting and background is the part people underestimate. Most swapping tools give you a floating head with mismatched shadows. This one keeps the scene intact and changes the person inside it, which is what makes the output usable rather than funny.
The open release matters too. Weights on Hugging Face plus an optimization script mean agencies and solo creators are working from the same starting line, and the prep-time cut per image is the sort of detail that only shows up when someone actually uses the thing.
Limitations
Be clear-eyed about a few things. This is a full-person swap, not a face swap — clothing, hair and anatomy all come from the character sheet, so you can't keep the original outfit unless your sheet matches it. That 1536x1024 sheet layout is specific, and stylized characters still need a proper three-view reference; a single selfie only works for real humans. Six steps without CFG buys you speed at the cost of prompt obedience, so elaborate instructions won't land the way they would on a long, heavily guided run. You need Diffusers, a working Python setup and a decent GPU, which is a real barrier for non-technical users. Group scenes with heavy overlap or occlusion will still need fiddling with --left and --kept. And the whole category carries a consent problem: swapping a real person's likeness into a photo without permission is a bad idea legally and ethically, no matter how clean the render looks.
Conclusion
Viggle's move from video to stills is a small technical step and a big practical one. Full-person replacement that respects the original pose, light and background, at 6 sampling steps, in an open LoRA you can download today — that's a tool worth having in the drawer.
One action for today: open the Viggle Image Swap model page, grab the LoRA file, and build a single 1536x1024 character sheet — portrait on the left, front-side-back on the right. Once that sheet exists, every photo you own becomes a candidate for a swap.



Comments 0