Kroma released: a Krea-2 fine-tune trained for 3 months locally on RTX Pro 6000 GPUs

KadahCoba · reddit · 2026-10-11

Developer LodestoneRock released Kroma, a fine-tune of the Krea-2 base model, now downloadable on Hugging Face.

Unlike the previous Chroma run, the Krea team provided the real, non-collapsed base model, enabling direct OPD (RL) on top and good out-of-the-box generation quality.
Training cost roughly 10,000 RTX Pro 6000 hours (3+ months on a 4-GPU local machine), including failed experiments and bugs; ComfyOrg donated 2 RTX Pro 6000 cards.
Next steps: focus on better datasets ("about 90% of quality comes from data"), then attempt video models like h3 — but fine-tuning video is astronomically expensive, so the project is fundraising for 4 more RTX Pro 6000 ($60k total).

More from Multimodal