Granola Bowl

The first attempt used the materials I had set up for the scene in Blender: speckled ceramic bowl, wood table with fine grain. Canny at defaults, thresholds 50 / 150. The result was poor — a half-empty bowl, a little fruit, and strange speckles reminiscent of poppy seeds. Almost nothing from the detailed prompt came through.

The error is visible in the Canny image: this isn't an edge map. It's a noise field. Two sources flood it. The entire background is covered in fine lines — the wood grain, which Canny reads as thousands of tiny edges. And the interior of the bowl is covered in white dots — the speckled glaze, also read as edges. What actually matters — the bowl rim and spoon silhouette — drowns in all of it.

That explains the weak result. At high strength, ControlNet tells the model: structure must go here, and here, and here, everywhere. The background is nailed down by the grain noise. And the bowl interior is defined by the speckles as a filled, textured surface — even though it's empty. The model simply can't fill the bowl freely. The geometry says: empty, dotted bowl in a noisy environment.

There are two ways to fight Canny noise. You can filter it downstream by raising the thresholds — low-contrast micro-edges drop out. That works, and it's the fallback when you can't change the scene.

The cleaner approach is to avoid the noise at the source. In this run the thresholds stayed at 50 / 150 — unchanged. Only the materials changed: the speckled glaze out, a smooth, matte, neutral bowl in. The fine wood grain replaced by a material with clear floorboard joints instead of dense surface texture.

The result, with identical Canny settings: same subject, same thresholds. Only the materials are different. The noise field has become an edge map: the bowl rim as a clean ellipse, the spoon silhouette, the rim of the honey dish. And the crucial point — the bowl interior is black. Not a single speckle left. That's exactly where the model can now place granola, yogurt, and fruit. Nothing forces an empty surface on it anymore.

This is the core lesson, and it's model-agnostic: Canny is calculated before any model renders. Keep the control surfaces smooth and neutral. Texture is what the model should generate — stochastically, from the prompt. It doesn't belong in the geometry channel.

A side effect makes this concrete. A completely empty surface gives Canny nothing, and the model loses its sense of perspective. Hence the floorboard joints: a few clear, long lines in the direction of the vanishing point define the table plane and the perspective anchor. They survive even high thresholds because they're high-contrast and long. The fine grain, on the other hand, can disappear. You control the geometry of the surface; you leave the texture as a material to the model.

The prompt is just as decisive for a food shot — and Flux requires something different here than SDXL.

SDXL works with comma-separated keywords. Flux doesn't. The T5XXL encoder responds poorly to keyword stacks and well to descriptive sentences. More on the separate content architecture of CLIPTextEncodeFlux in 2.4 — Prompt Architecture — what counts here is the result: a food subject doesn't become appetizing through tags, but through the description of sensory surface qualities. Moisture. Gloss. Freshness.

The prompt for the final result:

The first Flux run was too hard. ControlNet strength 0.7 with the InstantX Canny adapter pushes edges through so forcefully that the granola looked dry and crumbly — forced edge structure instead of soft material freedom. Flux holds the composition at significantly lower strength because the base model is geometrically more stable than SDXL. So: lower the strength, raise the guidance: