Do the thinking before the 4K pass.

A current community question asks how to take an architectural render through FLUX and ComfyUI for “final polish” at 4K. The desired changes sound modest: better light, texture, reflection and planting. The requested pixel count is not modest at all.

That distinction matters. Enhancement asks the model to interpret the image. Upscaling asks it to enlarge one. A graph that pretends those are the same operation is a very efficient drawing-error machine.

Why should enhancement and upscaling be separate?

Diffusion has room to invent whenever it denoises. ComfyUI's official Stable Image Ultra image-to-image guide describes its image_denoise control plainly: values near zero retain more of the input, while values near one move toward text-to-image generation. The exact useful value depends on the model and input. The principle does not.

At working resolution, invention can be reviewed quickly. A shifted glazing bay or rubbery handrail is visible beside the source, and another queue run does not monopolize the machine. At 4K, the same mistake becomes slower, more expensive and harder to spot because the surface detail looks persuasive.

Upscale only after the architectural decisions survive. This is the same logic behind our render-enhancement change contract: approve what may change before choosing the node that changes it.

The first pass earns permission to add detail

Start from the cleanest conventional render or viewport export available. Keep the final camera and aspect ratio. Supply depth, line or edge control when the chosen model supports it, particularly where a facade grid, roof line or stair profile carries the design.

Then state a narrow job. “Improve late-afternoon light and material response while retaining all geometry” gives the graph a boundary. A prompt that re-describes the whole building invites a fresh building. FLUX, SDXL and partner API nodes differ in conditioning and settings, so there is no honest universal sampler recipe.

Use the smallest denoise strength that produces the required improvement. Test one representative frame before a set. If low denoise cannot make the requested change, decide whether the brief needs a targeted mask or whether the source render should be fixed upstream. Turning the global control upward is quick. So is losing the window schedule.

Save the accepted working-resolution output beside its source, workflow JSON, model name, checkpoint or API version, seed where the node honors one, and key settings. ComfyUI is valuable because the graph can expose a process. A screenshot of spaghetti is not that process.

What belongs in the second pass?

The upscale branch should have a dull brief. Reach the required dimensions. Retain edges. Add enough texture to avoid smeared surfaces. “Make it more photorealistic” belongs back in the first pass because it gives the upscaler creative authority over the whole frame.

A latent upscale with another diffusion stage may add attractive detail, but it is still generative. A conventional or model-based pixel upscaler usually has less license to revise the scene. Choose according to the defect you actually have. If the accepted image needs better paving joints, a targeted detail pass can be defensible. If it merely needs to fit a 3840-pixel-wide board, restraint wins.

Tile-based processing can make a large image fit available VRAM, yet every tile creates a local context. Long facade lines, continuous shadows and repeated bays can expose seams or inconsistent detail. Use overlap where the node provides it, then inspect across tile boundaries rather than admiring one brick at 200 percent.

ControlNet protects structure, not truth

Depth, edge and line controls are useful restraints. They are not certificates of dimensional fidelity. A depth map can preserve the broad recess while the model changes a frame count, material joint or balustrade spacing inside it.

That is why our guide to denoise and ControlNet treats controls as layers of resistance. Source pixels, low denoise, a structural condition and a mask each reduce the model's room. None relieves the author from comparison.

Control inputs also fail in recognizable ways. Fine line maps can overstate material joints. Depth estimation can confuse glass and reflection. Soft planting can become a hard silhouette. Preview the control image itself. If it misreads the architecture, faithfully following it is still failure.

How do you inspect the final 4K file?

Compare source, accepted enhancement and final upscale at matching zoom. Blink comparison is excellent for movement. Difference blending can expose broad changes, though legitimate lighting work will also register. The point is to direct attention, not produce an automatic pass stamp.

Inspect crop by crop: silhouette and roof edges; every opening; repeated facade modules; stairs and rails; signage; furniture legs; people near geometry; planting against walls; reflection lines; and the ground contact of every object. These are places where a beautiful texture can conceal a small architectural lie.

Check the delivery file too. Confirm dimensions, color profile, compression and absence of tile seams. A 4K label says nothing about useful detail. Four thousand pixels of invented mortar are still invented mortar.

For client work, record the accepted source and final hash or file identifier. Our client assumption log gives visual changes somewhere to live instead of leaving them buried in a graph.

ArchiGen's take: 4K is a delivery specification

The community desire for a one-graph finish is understandable. ComfyUI makes graphs look like machines, and machines imply that every output at the far right has passed through an orderly factory.

The useful graph has a stop in the middle. Someone approves the image before more pixels give it more confidence.

Make the design decision small. Make the file large.

Evidence note: the community prompt was identified in the 27 October 2026 sweep. ComfyUI's official Stable Image Ultra guide was checked the same day for its documented image_denoise behavior. Workflow advice is an editorial control method, not a report of ArchiGen hands-on testing or a universal setting claim.