Furniture AR Pipeline Quest: True-to-Scale AI Image Compositing Still Unsolved
New-Piglet-5207 · reddit · 2026-10-09
A Reddit developer building a product to preview online-store furniture in users' own rooms asks for help with an image pipeline.
Current stack: parse product page for dimensions → GPT-5.4 vision plans the scene (target box + edit prompt) → MoGe-2 depth estimates pixel width → crop edited with gpt-image-2.5 on Azure, pasted back with tone correction.
Pitfalls tested:
- Whole-room edits (ChatGPT-style) redraw 7–30% of the room and often get scale wrong
- Crop-and-paste keeps the rest intact but leaves visible seams
- Two-step erase-then-add is cleaner but still distorts
- gpt-image with mask: 15–25% of pixels outside the mask still change
- Pasting real product cutouts: seamless but wrong perspective (store photos shot from above)
- Gemini 3 Pro Image good for masks/erasing, not drawing; Bedrock Stability keeps out-of-mask pixels but poor quality, no reference image; FLUX Kontext takes no mask and shifts framing
The question: does a pipeline exist that gets real dimensions, correct scale, and zero distortion — and what does each approach force you to give up?
More from Multimodal
- MiniMax-H3 hailed as favorite open release of the year, runs fully local — AdinaYakup · 2026-10-23
- Michal Malewicz experiments with depth maps in his creative workflow — michalmalewicz · 2026-10-23
- One-Shot Tunable Style Transfer on Flux3 Via Reference Latent + Node — Winter_unmuted · 2026-10-23
- PhysVidBench accepted at EMNLP 2026: do video gen models grasp physical commonsense? — aykuterdemml · 2026-10-23
- EMNLP 2026 paper teaches multimodal LLMs to get cartoons via incongruity-resolution supervision — aykuterdemml · 2026-10-23
- A gorgeous Midjourney style: sref code 820221253 for dreamy lavender watercolor looks — umesh_ai · 2026-10-23