Flimmer – open source video LoRA trainer for WAN 2.1 and 2.2 (early release, building in the open)

Alvdansen Labs, led by developers including Timothy Bielec, has launched Flimmer, an open-source video LoRA (Low-Rank Adaptation) training toolkit in early release. Built specifically for WAN (Weights & Biases) 2.1 and 2.2 models for text-to-video (T2V) and image-to-video (I2V) generation, Flimmer offers a comprehensive pipeline that automates the entire workflow from raw footage to a trained checkpoint. This includes scene detection and splitting, frame rate normalization, captioning using Gemini and Replicate backends, CLIP-based clip triage, dataset validation, and VAE + T5 pre-encoding. The project is being developed transparently, with the team actively soliciting community feedback and bug reports.

The toolkit's major technical differentiator is its support for phased training, a multi-stage approach where each phase has distinct learning rates, epoch counts, and datasets, with checkpoints automatically carrying forward. This enables advanced curriculum training strategies. Most significantly, it provides the infrastructure to properly handle WAN 2.2's Mixture-of-Experts (MoE) architecture. Unlike current trainers that treat the model's two experts as a single entity, Flimmer's phased system allows for a unified base training phase followed by forked, specialized training phases for each expert with tuned hyperparameters. While the data preparation tools are standalone and output formats compatible with popular trainers like kohya and ai-toolkit, the MoE specialization feature remains experimental. The team plans to add support for the LTX model next and is open to community requests for other model integrations.