The Viggle team has released the Viggle/Viggle-Image-Swap LoRA adapter on Hugging Face for the Qwen/Qwen-Image-2.1 base model: in six inference steps, it fully replaces the person in the source frame—face, hair, body, and clothing—with a character from a character sheet, preserving the pose, expression, framing, background, and lighting. According to the authors, on an NVIDIA B200 accelerator, one frame is generated in approximately 2.2 seconds; however, the Qwen Research license currently permits only research and evaluation, excluding commercial products.
What happened
The adapter is published in the open repository Viggle/Viggle-Image-Swap along with the model card, demo examples, and the inference/swap.py script. Technically, this is a rank 256 LoRA module with alpha 256 in bf16 precision, trained on top of Qwen/Qwen-Image-2.1. The base inference mode fits into six steps entirely without CFG; in turbo mode, the adapter performs two steps, and the remaining detail is added by six steps of Qwen-Image-2.1-Turbo. The output resolution is approximately 1 megapixel. The authors provide benchmarks on an NVIDIA B200 accelerator: with loaded models, it takes about 2.2 seconds per image and 46 GB of peak memory; in turbo mode, it takes 4.1 seconds and 55 GB. The position of the person in the frame can be specified using plain words in the prompt—left, right, third from the left—which allows replacing multiple people in a single pass.
Context
Single-pass character swapping is a response to the classic character consistency problem: previously, replacing a person involved a chain of detection, segmentation, inpainting, and manual scene assembly. Here, the same task is reduced to a single few-step operation on top of a ready-made image model: the adapter does not introduce a new architecture but combines fast CFG-free inference with whole-figure replacement. The character is described by a 1536×1024 character sheet, with a portrait on the left and three full-body views on the right; such a sheet is created once for each hero and reused. The pipeline is supplemented with tone correction based on the unchanged background to prevent the frame from darkening or brightening after replacement. In terms of scientific novelty, this is incremental engineering work without a public training methodology—the value lies in the packaging of the technique rather than in new technology.
Why this matters for the industry
For studios and product teams, the adapter is a ready-made building block: character generation, merchandise mockups, stickers, and stylization no longer require fine-tuning for each hero, and character swapping becomes a parameter of the pass rather than a separate project. The economics change more noticeably than the technique itself: expensive multi-stage character redrawing pipelines are compressed into a single cheap operation, and protected value shifts from algorithms to rights—base model licenses and character IP. The main limiting factor is legal: the base model Qwen-Image-2.1 is distributed under the Qwen Research license, so this stack can currently be used to build prototypes and internal demos, but commercial launch will require separate licensing.
Why this matters for users
Practical scenario for the reader: download the adapter and the inference/swap.py script from Hugging Face, create your own character sheet with a portrait and three views for your hero, and then replace yourself or any person from a source photo with that character, preserving the entire frame. The placement of heroes in the scene can be described in words, and multiple people can be replaced in a single pass. The stated speed requires a B200-level accelerator with sufficient video memory; on simpler cards, generation time and memory consumption will increase significantly. Video demos and examples in the model card show the expected level of results. For now, this is a tool for research and experiments—it is too early to build a commercial product on it.
What is still unknown / limitations
The evidence base relies on a single self-reported source—the model card: the training dataset, methodology, and quantitative metrics for pose, background, lighting preservation, and similarity to the character sheet are not provided, and there are no independent benchmarks. Speed and memory figures are stated only for the B200; performance on weaker hardware can only be assessed qualitatively. The authors list known failures: the new character sometimes remains in the original person's clothing, the pose may shift, and in turbo mode, the model may return accessories or the entire person from the source. High speed alone does not prove quality—the adapter can only be verified by running it independently.
Sources
Author
Look at AI, editorial team