PRISM merge algorithm combines two Minimax-H3 finetunes, weights open-sourced
theOliviaRossi · reddit · 2026-10-25
Developer theOliviaRossi merged two Minimax-H3 Ref2VA finetunes (Singularity V2 and the 8-step Turbo 10Eros-Max) using a custom spectral/geometric algorithm called PRISM.
Why PRISM: merging turbo models into base models often introduces weight artifacts and signal degradation. PRISM addresses this by:
- filtering diffuse noise deltas with randomized SVD (rank ≤64)
- combining directional SLERP with log-space magnitude interpolation (polar geodesic)
- resolving sign conflicts and re-anchoring layer norms
The turbo block distribution is weighted toward time/head blocks (ttime 0.75, thead 0.70) with an average turbo share of 0.55.
Weights are published on Hugging Face in fp16, int8, and w6a8 formats; the author is soliciting test generations and feedback. The attached image shows a sample generation.
More from Multimodal
- ElevenCreative offers $100k for the catchiest AI ad, sharing the full maker workflow — AIwithGhotai · 2026-10-25
- GPT-6 Astra generates a full real-estate video tour from a single Zillow listing — venturetwins · 2026-10-25
- Kling 4 nails one-shot real estate videos with full reference-image consistency — venturetwins · 2026-10-25
- Flux 2 Klein 9B with LoRA freezes at 70% sampling on a 3090 Ti, user seeks troubleshooting help — FearlessLeader17 · 2026-10-25
- Higgsfield lands NVIDIA and TCV in Series B extension, crosses $1B annualized revenue — aakashgupta · 2026-10-25
- enGEN demo shows AI video with oil-painting cinematic look — jordanbayne · 2026-10-25