AIAny
AI Model2026
Icon for item

Krea-2 Depth ControlNet-LoRA

Depth-conditioned LoRA for Krea‑2 that extracts a depth map from any input image and generates new images preserving the original 3D structure and composition while changing content and style. Single 862MB LoRA, works with Krea‑2‑Raw and Krea‑2‑Turbo.

Introduction

Why this matters Depth is one of the most reliable cues for preserving scene geometry across image-to-image edits, yet many adapters either ignore it or treat it as a loose hint. This LoRA turns a precise inverse-depth encoding into a hard control signal for Krea‑2, so you can feed any photo or render and get outputs that keep the same 3D layout while changing objects, materials, or style.

What Sets It Apart
  • Depth-first control: extracts an inverse depth map with Depth-Anything-V2 and encodes it into the model's latent space so depth information is injected at every denoising step, producing strong structure adherence rather than occasional alignment.
  • Lightweight adapter workflow: distributed as a single 862MB LoRA (rank-64 + expanded input projection) that leaves the base Krea‑2 weights frozen; works with both the distilled Turbo workflow (8 steps, no CFG) and the Raw base (longer sampling, CFG).
  • Measurable consistency: reports Pearson depth correlation ~0.98 with an empty prompt and ~0.99 when prompted, indicating the control reliably preserves relative depth ordering across outputs.
  • Practical controls: a --lora-scale style knob lets you trade strict structure adherence for creative freedom; lower scales loosen depth enforcement, higher scales tighten it at some image-quality cost.
Who it's for — and tradeoffs

Great fit if you: need edits that must keep the original perspective and spatial relationships (e.g., architectural modifications, object replacement inside a photographed scene, or stylistic re-rendering while preserving camera pose). Also useful for rapid prototyping with Krea‑2 Turbo when you want near-interactive turnaround. Look elsewhere if you: primarily work with flat 2D illustrations or scenes without meaningful perspective (the depth estimator yields nearly uniform maps there), need very high-resolution (>~1MP bucket) final images, or require a fully integrated ControlNet implementation rather than a LoRA surgery approach. The adapter is best applied to photos/renders with clear depth cues; 2D art will produce weak control.

Where it sits in a pipeline

Use this as a mid-step for structure-preserving image-to-image workflows: extract depth from a source, run the LoRA-enabled Krea‑2 pipeline to generate a geometry-consistent output, then optionally apply downstream upscaling or postprocessing. It complements per-pixel masks and prompt-based edits by anchoring results to the scene's 3D layout rather than only to text or 2D constraints.

Information

Categories

More Items

Hugging Face
AI Model2026

Runs text generation, summarization, and question answering directly on Windows devices as a 2.3B-parameter instruction-tuned model. It supports eight languages and a 32K context window, trading broad world knowledge and deep reasoning for local execution and privacy.

Hugging Face
AI Model2026

Handles text, image, and video understanding with sparse attention designed for contexts up to 1 million tokens. It has 770B total parameters with 49B activated per token, targeting long-document analysis, multimodal reasoning, and extended workflows.

Hugging Face
AI Model2026

Estimates metric depth for every pixel in a photo, video, or live feed, with sharp object boundaries and stable results across frames. Multiple model sizes are available; weights are restricted to non-commercial research and academic development.