Krea-2 Depth LoRA Released: Two-Stage Training for High-Quality Image Depth Maps
JasonNickSoul · reddit · 2026-07-29
A developer has released a new Depth LoRA designed for the Krea-2 model, capable of converting any input image into a high-quality depth map.
The LoRA was trained using a two-stage, progressive-resolution strategy:
- Stage 1 (512 Resolution Pretraining): Trained on a 2,000-pair public dataset with VLM-regenerated captions to enhance text-image alignment, establishing a solid depth-structure prior.
- Stage 2 (1536 High-Resolution Refinement): Focused on detail reconstruction, optimizing edges and fine geometry using a cosine learning rate schedule for stable convergence.
Important: This LoRA requires the ComfyUI-EditUtils plugin to function correctly. For optimal results, input images should have a longest edge between 1024 and 2048 pixels.
More from Multimodal
- Microsoft removes Mage Flow and points to a more efficient Qwen-VL-class encoder — Dante_77A · 2026-07-29
- Opus 5 users are building full games, FPS prototypes, and 3D scenes in one shot — eyishazyer · 2026-07-29
- Seedance 2.0 prompt turns text-to-video into a chaotic bodycam street scene — techhalla · 2026-07-29
- A burger-cooking AI demo splits viewers between detail accuracy and ad-like polish — victor_explore · 2026-07-29
- SCAIL 2 survives character swaps, object permanence, and physics tests better than expected — blackmixture · 2026-07-29
- a16z’s Justine Moore says AI micro-dramas could widen who gets to create — a16z Podcast · 2026-07-29