<h1 id="depth-lora">Depth LoRA</h1><p>Convert depth maps into high-quality images</p><h2 id="introduction">Introduction</h2><p>This is a depth LoRA that converts depth maps into high-quality images. It was trained in multiple stages on a variety of datasets to balance overall structural accuracy with fine-grained detail reconstruction.</p><p></p><p>HF download:</p><p><a target="_blank" rel="ugc" href="https://huggingface.co/lrzjason/Krea2-Lora-Collection/tree/main/loras">https://huggingface.co/lrzjason/Krea2-Lora-Collection/tree/main/loras</a></p><p></p><p>EditUtils</p><p><a target="_blank" rel="ugc" href="https://github.com/lrzjason/ComfyUI-EditUtils/tree/main">https://github.com/lrzjason/ComfyUI-EditUtils/tree/main</a></p><h2 id="training-details">Training Details</h2><p>The training followed a two-stage, progressive-resolution strategy:</p><h3 id="stage-1-512-resolution-pretraining">Stage 1 — 512 Resolution Pretraining</h3><p>Trained on a 2,000-pair subset sampled from a public depth dataset, with all captions regenerated by a VLM (VLM re-caption) to improve text-image alignment. This stage used a constant learning rate of <code>1e-4</code> to establish a solid depth-structure prior.</p><h3 id="stage-2-1536-high-resolution-refinement">Stage 2 — 1536 High-Resolution Refinement</h3><p>Continued training on a 1536-resolution depth dataset, focusing on detail reconstruction (edges, thin structures, and fine geometry). This stage switched to a cosine learning rate schedule for more stable convergence.</p><p>Stage 1 512 res, 2,000 pairs (public depth dataset + VLM re-caption), constant LR <code>1e-4</code>Stage 2 1536 res depth dataset, cosine LR scheduleTotal Epochs50+</p><h2 id="usage">Usage</h2><p>⚠️ <strong>This LoRA must be used together with the EditUtils plugin:</strong><br />👉 <a target="_blank" rel="ugc" href="https://github.com/lrzjason/ComfyUI-EditUtils">https://github.com/lrzjason/ComfyUI-EditUtils</a></p><ol><li><p>Install the ComfyUI-EditUtils plugin.</p></li><li><p>Load this LoRA in your workflow.</p></li><li><p>Feed in the source image to obtain the corresponding depth map.</p></li></ol><h2 id="notes">Notes</h2><ul><li><p>Works better on 10240- 2048 with long captions; with higher-resolution inputs yielding richer depth details.</p></li><li><p>Will not work as intended without the EditUtils plugin.</p></li></ul><h2 id="rh-workflow">RH Workflow</h2><p>Experience link: <a target="_blank" rel="ugc" href="https://www.runninghub.ai/post/2082077636234313729/?inviteCode=rh-v1279">https://www.runninghub.ai/post/2082077636234313729/?inviteCode=rh-v1279</a></p><h2 id="contact">📬 Contact</h2><ul><li><p>𝕏 <strong>Twitter</strong>: <a target="_blank" rel="ugc" href="https://twitter.com/Lrzjason">@Lrzjason</a></p></li><li><p>📧 <strong>Email</strong>: <a target="_blank" rel="ugc" href="mailto:lrzjason@gmail.com">lrzjason@gmail.com</a></p></li><li><p>💬 <strong>QQ Group</strong>: 866612947</p></li><li><p>💬 <strong>WeChat ID</strong>: fkdeai</p></li><li><p>🎨 <strong>CivitAI</strong>: <a target="_blank" rel="ugc" href="https://civarchive.com/user/xiaozhijason">xiaozhijason</a></p></li></ul><p></p>

Files with SHA256: 94b0a04db0a4504f9fffc39d5191f3bdeb18f97e554ff975086979694c704671