Architecture and interior concept visuals in ComfyUI: control the layout, then label what is illustrative
Concept visuals from sketches, plans or 3D screenshots can use ComfyUI control images and masked edits, but the results are illustrative and not dimensionally accurate. This guide covers choosing control images, checking outputs against the source drawing, labelling images for clients, and recording licence and provenance.
Purpose and scope
For architects and interior designers making concept visuals from sketches, plans or 3D screenshots: what to control, record and not claim.
Source-supported context
Its authors describe ControlNet as adding spatial conditioning to pretrained diffusion models, demonstrated with edges, depth, segmentation and pose Source 5. Its repository lists Canny, M-LSD straight-line, depth and normal-map conditions among others, under Apache-2.0 Source 6. ComfyUI's tutorial applies control images with strength and start/end percent Source 1; its depth page lists architectural scenes (spatial structure and perspective of buildings) among use cases Source 2.
The SDXL base card says the model lacks perfect photorealism, cannot render legible text, struggles with compositional tasks (its example is a red cube on top of a blue sphere), and that using it for factual depictions of people or events is out of scope; its licence is CreativeML Open RAIL++-M Source 7.
On rights, the US Copyright Office's January 2025 report concludes that, with current generally available technology, prompts alone do not give enough control for authorship, while perceptible human-authored material and creative selection, coordination, arrangement or modification can be protected, case by case Source 8. US only; may change; not legal advice.
Before you start
1. Confirm you may use the source views and any third-party images.
2. Identify checkpoint, ControlNet model and preprocessor nodes, with licences.
3. Prepare storage for sources, outputs and records.
Steps and checks
The sequence below is an editorial procedure based on the cited context, not a diagnosis already confirmed for your environment.
Step 1. (Editorial advice, not from a source.) Decide what the source geometry must fix (massing, layout, camera) and what the model may invent (materials, light).
Step 2. Prepare and inspect a control image (line art, clay render, depth map). The tutorial says current Comfy Core nodes do not include all types of preprocessors and names custom node packs for the purpose Source 1; record any you use.
Step 3. Pick a control type that suits the source view. In a depth map, brighter areas are closer and darker areas farther Source 2. The apply node takes a strength (higher means more influence) and start/end percent (when guidance begins and stops during diffusion) Source 1. Changing one setting at a time is editorial advice.
Step 4. (Editorial advice.) Overlay each candidate on the source view; reject mismatched openings, walls or ceilings.
Step 5. For material or furniture swaps, mask the region; the inpainting page covers an image alpha channel or the mask editor as mask sources, the VAE Encode (for Inpainting) node and its grow_mask_by input, and shows an inpainting-specific checkpoint giving better results than a standard one in its example Source 3. Save each swap separately.
Step 6. ComfyUI can embed workflow and prompt in saved files, but re-encoding may strip them, they can be disabled, and they are not a digital signature and do not prove who created or modified a file Source 4. Keep an external model record too.
Step 7. (Editorial advice.) Label client-facing images, e.g. "AI-assisted concept visual, not to scale, not for construction", and send the source drawing.
Step 8. (Editorial advice.) Check model and ControlNet licences and contracts; for authorship, read the Copyright Office report Source 8.
Completion check
Each delivered image has saved source, control image, workflow, model record and label.
Limitations and cautions
A generated image is not shown here to be dimensionally accurate or buildable, and this site did not test how closely any model follows a plan. The metadata page calls embedded metadata optional and untrusted input and says copying a file copies its embedded workflow Source 4. Steps 1, 4, 7 and 8 are this guide's editorial suggestions, not source-backed requirements; the label wording is an example only. ControlNet checkpoint licences and depth preprocessing were not verified.
Original sources
- ComfyUI ControlNet tutorial — checked 2026-10-06.
- ComfyUI depth ControlNet tutorial — checked 2026-10-06.
- ComfyUI inpainting tutorial — checked 2026-10-06.
- ComfyUI workflow metadata — checked 2026-10-06.
- ControlNet paper (arXiv 2302.05543) — checked 2026-10-06.
- ControlNet repository (lllyasviel) — checked 2026-10-06.
- Stable Diffusion XL base 1.0 model card — checked 2026-10-06.
- US Copyright Office, AI Part 2: Copyrightability (Jan 2025) — checked 2026-10-06.
Prepared 2026-10-06 from the sources above. No GPU run, quality measurement or runtime guarantee is supplied.
Related troubleshooting and guides
Was this page helpful?
Anonymous. Only a yes/no counter is stored; no account, IP address or device details.
Sources & references
Editorial research from the cited public sources, dated 2026-10-06. This site did not run or test the workflow described.
01ComfyUI ControlNet tutorialSource checked: 2026-10-0602ComfyUI depth ControlNet tutorialSource checked: 2026-10-0603ComfyUI inpainting tutorialSource checked: 2026-10-0604ComfyUI workflow metadataSource checked: 2026-10-0605ControlNet paper (arXiv 2302.05543)Source checked: 2026-10-0606ControlNet repository (lllyasviel)Source checked: 2026-10-0607Stable Diffusion XL base 1.0 model cardSource checked: 2026-10-0608US Copyright Office, Copyright and AI, Part 2: CopyrightabilitySource checked: 2026-10-06Report an issue · dc8e08c4-2d1f-539a-a4b5-39899bd79d85
Related reading
Pages linked to this one by the editors.
Identify the expected edge, pose or depth representation and trace how its conditioning reaches the sampler. Different control models are not one universal node.
A painted mask is only part of an inpainting workflow. Check the image, mask interpretation and supported model/encoding path together.
Use the official basic image-to-image path to understand how input pixels influence a result. Denoise and seed serve different purposes; there is no single best value for every image.
A model name can cover bundled checkpoints, split weights, quantized variants, text encoders and VAEs. Preserve enough identity information to know which artifact a workflow actually used.