InfiniteYou

Generates photorealistic images from text prompts while preserving a person's identity from reference photos. Uses InfUFluxPipeline to orchestrate the FLUX Diffusion Transformer base model, injecting identity features extracted from reference images via InfuseNet's residual connections throughout the diffusion process. The pipeline coordinates face analysis, identity feature extraction, and controlled diffusion sampling to balance text-image alignment with identity similarity.

Provides two pre-trained model variants (aes_stage2 and sim_stage1) that represent different points on the identity-preservation vs. aesthetic-quality spectrum. The aes_stage2 variant applies supervised fine-tuning (SFT) to improve text-image alignment and visual aesthetics, while sim_stage1 prioritizes identity similarity. Users can select the variant at runtime based on their specific use case requirements.

Supports composition with OmniControl for multi-concept personalization, enabling simultaneous control over multiple identity-related or style-related concepts in a single generation. The pipeline can integrate OmniControl's multi-concept conditioning alongside InfuseNet's identity injection, allowing users to generate images that preserve identity while also incorporating other personalized concepts (e.g., specific clothing, accessories, or artistic styles).

Exposes diffusion sampling parameters (guidance scale, number of steps, sampler type) as user-configurable options within the InfUFluxPipeline. Users can adjust these parameters to control the balance between identity preservation, text-prompt adherence, and generation quality. Higher guidance scales strengthen text-prompt following; more steps improve quality but increase latency. The pipeline supports multiple sampler implementations (e.g., DDIM, Euler, DPM++).

Supports seed-based reproducibility for image generation, enabling users to generate identical images by specifying the same seed, reference image, prompt, and parameters. The pipeline manages random number generation across PyTorch, NumPy, and other libraries to ensure deterministic behavior. This is critical for debugging, evaluation, and creating consistent results across different runs.

Analyzes reference photos to detect faces and extract identity-relevant features that are injected into the diffusion process. The Face Analysis Module performs face detection (likely using MTCNN or similar), extracts facial embeddings or feature vectors, and passes these to InfuseNet for integration into the generation pipeline. This enables the system to understand and preserve the identity characteristics of the reference person.

InfuseNet injects identity features into the FLUX Diffusion Transformer via residual connections at multiple layers of the model, rather than concatenating embeddings or using cross-attention. During the diffusion process, identity feature vectors are transformed and added to the DiT's hidden states at strategic points, allowing identity information to flow through the generation without disrupting the model's ability to follow text prompts. This architectural pattern preserves identity semantically within the learned representation space.

Provides multiple memory optimization strategies to enable inference on GPUs with limited VRAM (16GB or less). Supports flash-attention for reduced memory footprint during attention computation, 8-bit quantization for model weights, gradient checkpointing, and selective layer freezing. Users can enable/disable optimizations via configuration parameters, trading off memory usage against inference speed and generation quality.

Supports optional composition with LoRA (Low-Rank Adaptation) modules for style transfer (e.g., Realism, Anti-blur LoRAs) and ControlNet for explicit pose, composition, or style guidance. These extensions are loaded and applied within the InfUFluxPipeline without modifying the core identity-preservation logic, allowing users to layer additional control signals on top of identity-preserved generation. The pipeline handles LoRA weight merging and ControlNet conditioning at the appropriate diffusion steps.

Provides a test.py CLI script that enables programmatic and batch image generation without GUI overhead. Users specify reference images, text prompts, model variants, and optimization settings via command-line arguments or configuration files. The CLI handles model loading, inference orchestration, and output saving, making it suitable for automated workflows, CI/CD pipelines, and server-side generation.

Provides an interactive web UI (app.py) built with Gradio that enables real-time image generation with live preview, parameter adjustment, and result gallery. Users upload reference images, enter text prompts, select model variants and optimization settings, and see generated results immediately. The interface handles model loading, inference, and result caching to provide responsive user experience.

Provides native ComfyUI nodes that integrate InfiniteYou into ComfyUI's node-based workflow system. Users can compose identity-preserved generation workflows visually by connecting nodes for image loading, identity extraction, prompt input, and generation. The integration handles model loading, parameter passing, and result routing within ComfyUI's execution graph.

Supports swapping the underlying FLUX base model with alternative variants (e.g., FLUX.1-schnell for faster inference) while maintaining identity-preservation capabilities. The InfUFluxPipeline is designed to be model-agnostic at the base level, allowing users to substitute different FLUX checkpoints without modifying the InfuseNet identity injection logic. This enables tradeoffs between inference speed, quality, and memory usage.