Vampelium Research
Vampelium Dev Research is the internal research division of Vampelium Dev - a small, private team building experimental AI systems focused on generative models, anime media, and scalable inference.
We work at the intersection of deep learning research and applied engineering. Our work is not published for prestige - it is documented because rigorous documentation makes better research.
What We're Building
VampAI
VampAI is our umbrella for all generative AI research and tooling under the Vampelium ecosystem. It spans text-to-image generation, model architecture experimentation, dataset pipeline engineering, and inference infrastructure.
Our research is not derivative. We do not fine-tune existing public models and call it research. We design architectures, build training pipelines from the ground up, and test hypotheses that the broader community has not prioritized - particularly around native style and quality control in diffusion models.
Active Research
VampDiffusion-V1
A from-scratch text-to-image diffusion model for anime and illustration generation, built on a Diffusion Transformer (DiT) backbone with a dual Mixture-of-Experts system for native style and quality routing.
- 3.0M filtered images from Danbooru2023 + Danbooru2024
- 1.3B total parameters, ~875M active per forward pass
- 8 style experts + 4 quality experts routed in a single forward pass
- No LoRA. No HiRes fix. No external upscalers.
Current status: Phase 1 pretraining - active. Step ~48,000 of 800,000.
Principles
From scratch when it matters. Fine-tuning existing models produces entangled results. When we want to understand whether an architectural hypothesis holds, we need a clean baseline - not one that inherits assumptions from a model we did not design.
Private by choice. We are not a lab competing for citations. We publish this documentation for internal rigor and for the record, not for external validation.
Constrained hardware, deliberate design. We train on 2× RTX 4090s. Every architectural decision is evaluated against a real VRAM budget. Constraints produce better engineering.
Team
Four researchers and engineers operating under the Vampelium Dev Research Division. Individual identities are not disclosed.
Documentation is updated as research progresses. Nothing here is final until marked as such.