Minimax-h3_Singularity-x-10Eros-Max_PRISM: Minimax H3 Merge Model Combined Using the PRISM Algorithm
OliviaRossi/Minimax-h3_Singularity-x-10Eros-Max_PRISM
About the project
A merge model combining two Ref2VA fine-tuning models based on Minimax H3. It uses WarmBloodAban's Singularity and TenStrip's 10Eros-Max as the original sources and applies a self-developed PRISM algorithm instead of existing merge tools. It provides weight files in fp16 and int8HQ versions.
It undergoes a 4-stage mathematical process, including spectral denoising, polar geodesic interpolation, sign conflict mediation, and magnitude re-anchoring, rather than simple weight averaging. The core involves separating structured and residual parts using randomized SVD and applying adaptive gamma values to reduce noise. This allows for a more precise fusion of the strengths of both models.
Various test update files, such as an 8-step turbo mode, a hybrid beta version, and a PIKA intensity adjustment version, are also released. The license follows the original models, so the conditions of each source model must be checked before use. The download count for the last month was recorded at 99,353.
It is suitable for developers seeking better video generation quality in the Minimax H3 ecosystem. It is useful when you want to combine the strengths of different models while maintaining the characteristics of specific fine-tuning models. Additional materials can be referenced via the CivitAI mirror link.
OliviaRossi/Minimax-h3_Singularity-x-10Eros-Max_PRISM
The original page has no description.
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.