What is AI Depth to Image?

Upload a reference image. The AI computes a depth map (foreground vs. background, what is closer vs. farther from the camera) and uses that 3D layout as a constraint while rendering a brand-new image in any style or content you describe.

When should I use depth vs. canny vs. pose?

Depth preserves SCENE GEOMETRY (the room, the camera angle, the relative distances). Canny preserves OUTLINE. Pose preserves BODY POSITION. Use depth when the spatial composition matters — e.g., restyling a product shot keeps the product in the same place even if you swap the background.

How is the depth map extracted?

MiDaS (Intel ISL) — a monocular depth estimator. It looks at a single 2D image and predicts per-pixel depth, no actual 3D camera or LiDAR needed. The depth map becomes the SDXL Depth ControlNet input.

What kind of reference images work best?

Anything with clear foreground/background separation. Indoor scenes, product shots, outdoor landscapes, room interiors all work. Flat compositions (e.g., a graphic poster) yield uninteresting depth maps; depth control is most useful when the source has 3D structure.

Yes. ~1,200 tokens per render (1,000 base + 20% ControlNet surcharge). Signed-in users get 5K free daily — about 4 depth-controlled renders per day at the free tier.

Can I use this to relight a scene?

Indirectly — the depth map keeps the geometry, the prompt controls the lighting ("at sunset", "under harsh studio lights", "by candlelight"). Output keeps the spatial layout, swaps the lighting cues. Great for "same room, different time of day" effects.

How does it differ from img2img?

img2img re-paints over the input directly — colors, edges, AND geometry. Depth-to-image throws away the colors and edges, keeps just the geometry. Same composition, totally fresh content. img2img tends to be subtle; depth lets you radically swap the subject.

Can I tune how strictly depth is followed?

Yes — ControlNet strength slider. Default 0.7. 1.0 = strict adherence (output looks like a depth-mapped re-render); 0.4 = loose (more creative variation). Tune per scene.

What output sizes are supported?

512×512 default. Up to 1024×1024 on standard SDXL aspect ratios. Larger sizes use more VRAM + tokens.

Is the depth map output viewable?

Today only the final rendered image is returned. The depth map is computed on the GPU + discarded. If you want to expose the raw depth output as a separate download we can wire that — file an idea via /contact/.

Are the uploaded references stored?

Processed immediately, depth-extracted, then discarded. Only your prompt + final render survive on /account/?tab=history. Never used for training. /privacy/ for the full policy.

Yes — POST multipart to /v1/image/generate/ with model=sdxl, prompt, control_image (file), control_type=depth. Bearer auth, 10K free tokens/month. Curl at /api/.

وێنە (هەژمارەی کۆنتڕۆڵ)

کێشانی هێڵ، یان هەر وێنەیەک دابگرە -- ئێمە کێشانی دەرئەنجامەکەی دەردەهێنین (دۆزینەوەیەکی جوانی لاوە) و بەکاری دەهێنین بۆ کۆنترۆڵکردنی وێنە دروستکراوەکە. زۆر باشە بۆ گۆڕینی کێشانی نەخشە بۆ هونەری تەواو لەکاتێکدا پێکهاتەی خۆت دەپارێزیت.

وێنەی کۆنتڕۆڵکردن (نقاش / شێوەی کێشان / وێنە)

ئێمە دەستنیشانکردنی لاشەی کاننی لەسەر ئەم کار دەکەین. نەخشە پاکەکان کۆنتڕۆڵی پاکتر دەدەن.

پرسیار

هێزی کۆنتڕۆڵکردن

وازهێنان 0.7 زۆرتر

ڕێژەی ڕوونما

~1,200 tokens (SDXL × 1.2 ControlNet)

وێنە (هەژمارەی کۆنتڕۆڵ)

ئەنجام

١٠,٠٠٠ تەن بەبێ پارە بەدەست بهێنە

چاوەڕێ بکە - ١٠K ی بێ نرخ بەدەست بهێنە!

زیاترت دەوێت؟