Highland Cottage

The Blender render shows what this workflow asks of Blender — and what it doesn't. Camera position, focal length, and proportions are precisely defined and reproducible. The house stands where it should. The hill, the hedge on the left, the chimney — everything is in place.

Materials, textures, and lighting are deliberately absent. The walls are grey surfaces, the two spheres by the door are placeholders only. This isn't an unfinished render — it's the division of roles: Blender delivers the geometry, not the image. The AI may interpret — but not recompose.

A Canny edge map is calculated from the Blender render first. It's the contract between geometry and model:

A clean case: clear contours for the house, roof, windows, the hilline in the background. Only bottom left, at the hedge, does it get a little noisy — organic edges Canny can't resolve cleanly. Uncritical for this subject, because the architecture carries the image. (How badly texture-heavy surfaces can flood a Canny is shown in the food case study — here the geometry is the subject, and that's exactly ControlNet Canny's home territory.)

The first photographic pass, SDXL with ControlNet Canny. The result is sparse and honest: a grey stone house in the Highlands, slate roof, white door, moorland behind it and a loch on the left edge. No garden, no ivy, nothing embellished. The composition is exactly that of Step 1 — ControlNet held the geometry. The model added material, light, and atmosphere.

Two observations show where the boundary between deterministic and stochastic runs — and they show it from two sides.

First, the door: two round wooden discs stand there, reminiscent of whisky barrels. In Blender they were two spheres, boxwood placeholders. The model found the two round shapes in the Canny and gave them content. Where Canny specifies a form, it enforces the form — what becomes of it is the model's guess.

Second, the foreground: a dry-stone wall runs across the entire image. It didn't exist in the Blender render and wasn't in the prompt. SDXL placed it in the empty space on its own initiative. That's the other side of the same coin: where Canny specifies nothing, the model fills the space at its own discretion. Form is enforced by Canny — the empty remainder is negotiated by the model with itself.

For this image, the invented wall is a lucky find; it fits. But it's the one place in the image where the composition is not geometrically secured — its position depends on the model's chance, not on Blender. Which is exactly why it's deliberately written into the prompt in Step 3: what began as an accidental gift becomes a fixed specification.

What SDXL delivers as a base, Flux takes to a finished photograph. This is where everything the sparse version lacked comes in: ivy and Virginia creeper in deep autumn reds across the facade, a cottage garden of lavender, grasses, and roses, the two boxwood spheres flanking the white door. The atmosphere deepens — heavy cloud light, muted earth tones.

A change of mechanism worth naming: this step didn't run through ControlNet but as img2img in a separate ComfyUI workflow — no Blender, no Canny, just the SDXL image from Step 2 as reference. What holds the composition here is no longer the geometry edge but the low denoise value: img2img stays close to the source as long as you don't let it noise too heavily. The phrase "Preserve exact composition" in the prompt supports what the moderate denoise already does.

House, wall, hills, the loch on the left — everything stays in place. What changed is only the visual execution. This is the point where the workflow shows what it can do: photographic materiality, without losing the image architecture.

The final step is a shift in perspective — not the camera's, but the season's. The same house, the same composition, but winter: hoarfrost on wall and garden, the autumn reds faded to brown stems, a cold silver light across the whole scene.

This one variant wasn't generated locally but with a high-end cloud model (Nano Banana Pro) — a single render. This is a deliberate exception and I want to name it openly: the workflow otherwise lives on local inference and small models. For a controlled season change with composition preserved, a large cloud model delivers a result here that doesn't justify the effort of a local rebuild. One render, one exception, made transparent. More on the thinking behind that on the position page.