InfiniteYou
https://github.com/bytedance/infiniteyou- Category
- AI Tools
- Rank
- No. 1036Tools index
Previous survey · No. 1041 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- bytedance
- GitHub
- 2.7k stars
- Date
About
ByteDance's identity-preserving photo recrafting — change pose, scene, or style while keeping the same face identity intact.
What it does
InfiniteYou builds portraits from a face image, a text prompt, and an optional pose reference. It extracts facial identity features, projects them into the FLUX generation process, and applies them through residual conditioning. Optional facial keypoints steer composition without supplying the full reference image.
Why it's ranked here
The design exposes useful choices instead of pretending identity fidelity and prompt adherence are identical goals. One model favors facial similarity, while the default favors prompt alignment and aesthetics. Adjustable conditioning, optional pose guidance, quantization, and CPU offloading make the research implementation unusually practical, though still hardware intensive.
What's good
The two model variants make a real quality tradeoff explicit. Users can tune when identity conditioning begins and how strongly it applies. The tool accepts deterministic seeds, optional five-point facial control, realism and anti-blur adapters, multiple output sizes, and several memory-saving configurations. Both command-line and visual interfaces are provided.
Tradeoffs
Full inference needs about 43GB of GPU memory. Quantization plus CPU offloading reduces that to about 16GB, but this remains a substantial requirement. The implementation requires CUDA, supports only one InfiniteYou-FLUX version, and depends on access to FLUX.1-dev. When several faces appear, it uses only the largest detected face.
How to use it well
Use it for local portrait experiments or node-based image workflows where identity consistency matters across varied prompts. Choose the default model for stronger prompt alignment and aesthetics, then switch variants when facial similarity matters more. Start with default conditioning before tuning. Do not choose it for CPU-only generation, modest GPUs, or preserving several identities in one input.
Technical notes+
test.py provides the argparse inference surface and writes generated PNG files. app.py wraps InfUFluxPipeline in Gradio, downloads model snapshots, switches between aes_stage2 and sim_stage1, and manages optional LoRAs. pipelines/pipeline_infu_flux.py detects the largest face with InsightFace, creates an ArcFace embedding, projects it through Resampler, and loads FLUX plus InfuseNet conditioning. pipelines/pipeline_flux_infusenet.py extends Diffusers' FluxControlNetPipeline. pipelines/resampler.py implements Perceiver-style attention for identity projection. requirements.txt pins the Python dependency set, while LICENSE contains Apache License 2.0 terms.
Observed
- License
- Apache License 2.0
- Primary language
- Python
- Installation surface
- Dependency installation through pip using a pinned requirements file
- Interfaces
- Command-line inference script and local Gradio application
- Workflow integration
- The documentation links an official ComfyUI node implementation
- Platform support
- Inference implementation requires CUDA-capable GPU execution
- Base model dependency
- Uses FLUX.1-dev and requires users to obtain access to it
Read from README.md, requirements.txt, app.py, test.py, pipelines/resampler.py, pipelines/pipeline_infu_flux.py, pipelines/pipeline_flux_infusenet.py, LICENSE.
What it can do
Change pose in photo while preserving identity
Photo of a person → Photo with same face but different pose
Change scene background while keeping face intact
Photo of a person → Photo with same face in different background/scene
Change photo style while maintaining facial identity
Photo of a person → Photo with same face in different artistic or visual style
Recraft photo composition with identity preservation
Original photo and desired modifications → Modified photo maintaining original face identity
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.