ANVIL is being reforged.
ANVIL 1 is retired as of the latest site relaunch. ANVIL 2 — built on stable diffusion model training instead of the legacy GAN pipeline — is in active development alongside the next wave of Tensorpunk products. Newer releases are taking priority on the public schedule, but ANVIL 2 has not been shelved.
If you grabbed ANVIL 1 and you need help with it: I will still answer. Email [email protected] and I will personally get back to you.
ANVIL 2 is a stable-diffusion-based audio model trainer. Where ANVIL 1 wrapped GAN training, ANVIL 2 fine-tunes diffusion models — the same family of architectures powering MACE — so the things you train can actually be used inside the rest of the Tensorpunk pipeline.
The software itself is largely built. What is left is the slow, honest part: training my own models with it, tweaking the pipeline based on what those runs reveal, and feeling out how diffusion fine-tuning actually behaves under different audio domains. That work is intentionally not being rushed.
If you have hands-on experience training or fine-tuning stable diffusion models — especially on audio — and you want a say in how ANVIL 2 evolves, reach out. I’m interested in collaborators who have actually done the work and want to push on architecture choices, training strategies, dataset curation, or evaluation. Email [email protected] with what you’ve trained and what you’d want to see in a tool built around it.
No date. ANVIL 2 ships when the training side proves itself, not before. MACE remains the focused, supported, production product in the meantime.
If you want to be notified when ANVIL 2 enters early access, the dev log is the canonical signal: /dev-log/.