Sparsity Anima

Sparsity Anima is an experimental diffusion model based on the Cosmos Predict 2 2B architecture and further fine tuned from the original Anima base model. The model was trained entirely on a single consumer GPU (RTX 5070 Ti) using a heavily customized full fine tuning (FFT) pipeline focused on maximizing efficiency and reducing VRAM usage. Multiple custom optimization techniques were used to make large-scale backbone training feasible on consumer hardware while preserving the flexibility and expressive power of full fine tuning. The fine tuning process was performed using a curated dataset of approximately 11,000 images sourced from Danbooru.

After the FFT process was completed, a custom post-training method called Sparse Parameter Merge (Parameter-Level) was applied to the model. Instead of performing conventional layer-wide merges, this technique alternates parameter inheritance directly at the parameter level between the original Anima base model and the fully fine tuned checkpoint using deterministic structured patterns such as A → B → A → B inside the tensors themselves. The objective is to restore portions of the original pretrained redundancy and generalization while preserving the learned adaptations from the fine tuning process, creating a structurally hybrid diffusion backbone.

A third auxiliary model (A → B → C → A → B → C) was also integrated using the same Sparse Parameter Merge technique with the objective of introducing additional concepts and newer character knowledge into the final backbone while preserving the stability of the original model structure.