TokenFlow

Converts source video frames into latent representations using Stable Diffusion's VAE encoder, then applies DDIM inversion to compute noise maps that can deterministically reconstruct original frames. This preprocessing stage extracts temporal sequences as latent codes and inverts them through the diffusion process, enabling frame-by-frame consistency tracking during editing. The inversion produces both latent tensors (for editing) and an inverted video reconstruction (for quality validation before proceeding to editing).

Propagates diffusion features across video frames by computing optical flow or patch-based correspondences between consecutive frames, then using these correspondences to enforce consistency in the diffusion feature space during editing. During the reverse diffusion process, features extracted from one frame are warped and injected into neighboring frames based on computed motion vectors, ensuring that semantic edits (e.g., 'change dog to cat') apply consistently across the temporal sequence without flickering or temporal artifacts.

Decodes edited latent tensors back to pixel-space video frames using the Stable Diffusion VAE decoder, converting 4-channel latent representations (8x downsampled) to 3-channel RGB video frames at the original resolution. The decoder is applied frame-by-frame to edited latents, producing the final edited video output. This stage is the inverse of the VAE encoding step in preprocessing, enabling the full latent-space editing pipeline to produce viewable video output.

Estimates optical flow between consecutive video frames to compute inter-frame correspondences, which are used to guide feature propagation during editing. The optical flow maps represent pixel-level motion vectors between frames, enabling the system to warp features from one frame to the next while respecting the underlying motion. This correspondence estimation is a prerequisite for the feature propagation mechanism, ensuring that edits follow the original video's motion dynamics.

Manages video frame sequences as batches during preprocessing and editing, enabling efficient processing of multiple frames in parallel on GPU. The system handles frame extraction, batching, and sequence management, allowing users to process videos of arbitrary length by chunking them into manageable batches. Batch processing reduces per-frame overhead and enables GPU parallelization, improving throughput compared to frame-by-frame processing.

Implements feature and attention injection at configurable diffusion timestep thresholds, allowing selective replacement of UNet features and cross-attention maps with values from the inverted source video. During the reverse diffusion process, features are injected at early timesteps (high noise) to preserve structure and at later timesteps (low noise) to allow text-guided semantic changes. This technique balances fidelity to the original video structure with adherence to the target text prompt through threshold-based switching.

Implements SDEdit-style editing by controlling the noise level (number of diffusion steps) applied to the source video before running the reverse diffusion process with a new text prompt. Lower noise levels preserve more of the original video structure; higher noise levels allow more dramatic semantic changes. The technique works by adding Gaussian noise to the inverted latents for a specified number of steps, then denoising with the target text prompt, effectively interpolating between structure preservation and text fidelity.

Integrates ControlNet guidance into the diffusion editing pipeline by extracting edge maps from the source video and using them as structural constraints during the reverse diffusion process. The edge detection (typically Canny or similar) creates a structural skeleton of the original video, which is fed to a ControlNet model alongside the text prompt. This ensures that edited frames maintain the same spatial structure and object boundaries as the original, even when applying dramatic semantic changes.

Generates an inverted video reconstruction during preprocessing to enable visual assessment of temporal consistency and reconstruction fidelity before proceeding to editing. The inverted video is created by decoding the DDIM-inverted latent tensors back to pixel space using the VAE decoder, producing a frame-by-frame comparison against the original. Users can inspect this reconstruction to identify temporal artifacts, flickering, or structural degradation that would propagate through downstream editing steps.

Provides YAML configuration files (e.g., config_pnp.yaml, config_sdedit.yaml) that specify all editing parameters including technique selection, hyperparameters (thresholds, noise levels, step counts), model paths, and I/O specifications. The configuration system decouples parameter tuning from code, enabling users to experiment with different editing strategies by modifying YAML files without touching Python code. Each editing technique has a dedicated config template with documented parameters and sensible defaults.

Orchestrates a three-stage pipeline (preprocessing → technique selection → editing) with separate entry points for each editing technique (run_tokenflow_pnp.py, run_tokenflow_sdedit.py, run_tokenflow_controlnet.py). The pipeline manages data flow between stages, handles intermediate file I/O (latent tensors, inverted videos, configuration files), and provides a unified command-line interface for executing end-to-end workflows. Users specify the editing technique via configuration or command-line arguments, and the pipeline automatically routes to the appropriate editing implementation.

Provides a command-line interface (CLI) for executing preprocessing and editing stages with arguments for specifying input/output paths, prompts, and technique-specific parameters. The CLI uses Python argparse to parse command-line arguments, with sensible defaults for common parameters and validation for required arguments. Users can invoke preprocessing (preprocess.py) and editing (run_tokenflow_*.py) scripts directly from the terminal, making TokenFlow accessible to non-Python developers and enabling integration with shell scripts or workflow automation tools.

Integrates Stable Diffusion 1.5 and 2.1 models as the core diffusion backbone, using the pre-trained UNet, VAE, and text encoder from these models without requiring fine-tuning or additional training. The integration abstracts model loading, device management (CPU/GPU), and inference through a unified interface, allowing users to specify which Stable Diffusion version to use via configuration. The system leverages the pre-trained text-to-image capabilities of these models for video editing without modifying model weights.