ComfyUI Exteriors: Depth, Fill or Img2img?

Which ComfyUI workflow fits architectural exteriors?

Choose by the source image and the edit it needs. Use depth guidance when the input defines the intended massing, a masked edit when a region needs replacing, or an image adapter when the task is a reference variation. BFL documents separate structural models, a Fill model for image regions and a Redux adapter for image variations. A longer node graph is not a selection criterion.

Work through geometry, materials, site context, then lighting. Treat these as review stages: keep an accepted output before starting the next operation. This order is an editorial recommendation based on the documented operations. Keep the model, encoders, VAE and conditioning components compatible with the chosen example; the FLUX and SDXL model-choice guide explains the separate routes.

1. Set the structural input before surfaces

FLUX.1 Depth [dev] generates images guided by a depth map. Start with the intended camera and massing. Where the source model can export depth, use that map as the proposed structural input and inspect it before generation. This recommendation follows the model's conditioning mechanism; it is not a claim that every opening will survive unchanged.

If the input is only a rendered image, MiDaS estimates depth from that image. An estimate and a model export are different sources of information. The auxiliary preprocessor pack creates hint images, including depth and edges; it does not replace the weights that use those hints.

Import the corresponding ComfyUI FLUX example, load its required files and supply the source input. Use the example's compatible components before adding custom nodes. BFL's full Depth model and its Depth LoRA are distinct options; the LoRA route requires FLUX.1 [dev]. Neither is an SDXL ControlNet.

Save a candidate and compare its silhouette, parapets and window openings with the source. Make that comparison a decision point before changing surfaces. The stage-checkpoint guide expands this review method.

2. Study facade materials from an accepted image

ComfyUI's img2img documentation describes loading an image, encoding it with a VAE and sampling that latent with reduced denoise. Lower denoise adds less noise and generally changes less of the input. This is a general img2img example, not a FLUX material-study graph. It explains the operation without establishing one correct value for concrete, glazing or brick.

Use the accepted structural image as the starting point for a material study. “Board-formed concrete, pale grey surface” is a prompt example, not a tested recipe. For a change confined to a spandrel panel or podium cladding, choose a compatible masked-edit workflow. A full-image instruction does not establish that adjacent surfaces will stay unchanged. Review the boundary as well as the material. A convincing texture can still belong to the wrong wall.

Sampler values belong to the selected model example. The FLUX fp8 checkpoint example, for instance, specifies Load Checkpoint with CFG 1.0. Do not copy an SDXL sampler recipe into it. ComfyUI also warns that fp8 can reduce quality; it recommends the full 16-bit version when resources permit. This is not a hardware capacity guarantee.

3. Choose the operation for site context

For a local edit, FLUX.1 Fill [dev] accepts an existing image and a mask to fill a region according to a text description. The ComfyUI Fill example includes access to MaskEditor from Load Image. Mark the region intended to change, run the documented example and compare the edited edge with the source. The Fill card also warns of colour shifts outside the filled area; inspect those surroundings too. A mask is an instruction to the workflow, not evidence that the resulting context is geographically correct.

For an image-reference variation, Redux is a FLUX.1 image adapter. Treat it as a different operation from filling a selected region. The reference image supplies visual information; it does not verify site conditions. Use a reference you have permission to use and keep its origin with the output.

The IP-Adapter model card places ip-adapter-plus_sdxl_vit-h in its SDXL family. That adapter is not the FLUX Redux component. A matching-sounding filename cannot make those families interchangeable.

Which public files does each route load?

The locations below follow the ComfyUI FLUX examples. Load the supporting text encoders and VAE named by the chosen example too. These are separate model routes, not a single plug-and-play chain.

The dev model card, Depth/Canny table, Fill card and Redux card identify the dev non-commercial licence. Generated-output permissions are a separate part of the terms. Check the licence for the intended use and BFL's commercial licensing options; local weights are not a blanket commercial-use permission.

4. Review lighting as another image change

For a lighting variation, return to the accepted context image and choose a compatible img2img workflow for that model. The general example linked below explains the operation; it does not supply a FLUX lighting graph. “Overcast daylight” and “warm evening light” are illustrative prompts. The documented denoise control governs how much noise enters the starting image; it does not promise a lighting-only edit.

Compare the candidate with the same baseline before accepting another variant. Check whether the material request still reads correctly and whether openings changed. Keep the original control inputs with the workflow; the second-pass guide discusses their use in later passes.

Keep the workflow reproducible

ComfyUI-Manager can save installation snapshots, but its documentation says snapshot support is incomplete for custom nodes not managed by Git. Record versions and required files alongside the workflow; do not treat a saved graph as a complete environment backup.

Custom packs are optional until a specific graph needs them. Advanced-ControlNet exposes control strength and sampling-time scheduling. Those controls are not a mandate to stop conditioning at one particular percentage, and the pack is not required simply because this guide has “advanced” in its URL.

For enlargement, Ultimate SD Upscale documents tiled img2img inside ComfyUI. Inspect it as another generative edit. There is no basis here for a claim that an external upscaler is required for architectural work.

Save the source image, model names and each accepted stage. The next pass should have a reason to exist.