Why concept images are intent, not geometry

AI image generators are astonishing at making pictures that look like products. But a picture is a flat arrangement of colored pixels. It has no millimetres, no wall thickness, no mating parts, and no guarantee that the thing in it could physically exist.

That is why Solid Vibes treats concept images as intent, not geometry: a guide to style and layout, never a blueprint.

What a concept image is good at

• Silhouette and proportion. Tall and narrow, or low and wide. Rounded or crisp.
• Style. Industrial, organic, playful, minimal.
• Layout. Where the buttons go, which side the handle is on, how many drawers.
• Material and finish cues. Walnut, matte white, brushed metal, translucent.
• Mood. Whether it looks like a toy or a tool.

What it cannot tell you

• Real sizes. The image does not know whether the lamp is 30 cm or 3 m tall.
• Thicknesses. The wall on a beautiful vase may be 0.2 mm in the picture. Nobody can print that.
• Internals. What is inside, where the screws go, how it opens.
• Mechanisms. A hinge drawn without a pin, a lid without a lip. Images often hide the engineering.
• Perspective-free truth. A wide-angle render exaggerates the foreground, so true proportions drift.

The workflow that respects this

1. Explore on the Concept Board. Generate four images, then refine with chips such as "Rounder" or "More industrial", mix two concepts, circle an area and describe the change, and pin photos, textures and colors to a moodboard. Every round is saved.
2. Lock the winner. Solid Vibes creates a clean turnaround sheet: front, side, top and three-quarter view on white. The front, side and top views are drawn straight on, like a technical drawing, in the same pose and at the same scale (only the three-quarter view is angled). This is the most useful artifact, because several views let the agent infer proportions far better than one angle. Then it reads the sheet's features and asks for a few real numbers. That usually takes under a minute. If the first read is slow, it takes a second, quicker look on its own, and the card tells you so; if that fails too, Try again is right there and your sheet is kept.
3. Hand off to Build. The turnaround and moodboard become visual intent for the agent.
4. Anchor to real dimensions. This is your job, and it takes one sentence.

Build the locked concept. Overall height is 320 mm. The base is a 120 mm circle. The bulb is a standard E27 socket. It will be 3D printed in two pieces.

With a few anchors, the agent can derive the rest sensibly: arm lengths, thicknesses, where the socket sits, how the pieces join.

Where real dimensions come from

The agent anchors to dimensions from three places, in order of trustworthiness:

1. Bought pieces. An E27 socket, an ESP32 board or an M3 screw has exact dimensions in the library.
2. Numbers you give. Overall size, key clearances, mounting hole spacing.
3. Defaults it chooses. Reasonable proportions drawn from the image. These show up with a "concept" badge on the Parameters, so you know they are assumptions.

Prompts that help

Ask the agent to be explicit about its assumptions:

Build this from the concept. Before you do, list the dimensions you are assuming from the image, so I can correct them.

Ask it to prioritize function over fidelity when they conflict:

Keep the overall look of the concept, but make the walls at least 2 mm and make sure it can be printed without supports. Tell me where you had to depart from the picture.

Organic concepts: the outline carries over

When the concept's form is the point (a smooth mouse, a vase, a figurine), the turnaround sheet does more than set the mood. The agent reads the subject's outline in the front, side and top views, builds a starting shape that matches those outlines at your real dimensions, and sculpts the model from it. Precise features (ports, screw posts, a socket, a flat base) go on top.

While it builds, the agent looks at its own work: after each build it puts the model next to the sheet, view by view, with the differences coloured in, and fixes what's off ("the tail is too short in the side view") for up to three more passes, keeping the version that matched best. After the build, the visual check compares the outlines once more and shows the match, for example "the outline matches the concept 88 %". Below 82 % in any view the agent makes one more pass to fix the biggest difference. The outlines are compared by shape, not size, so your real dimensions still win over the picture's proportions.

Your numbers are read for what they are: the object's own size ("overall length 180 mm") sets the scale, while a limit ("printer build height 220 mm", "fits on a shelf 120 mm tall") only caps it.

For characters, figurines, creatures and ornaments, the agent can go further when your workspace has the body generator: the sheet's three-quarter, front and side views go to an image-to-3D model, and the detailed body that comes back is turned to face the way the sheet shows and scaled to your real size. The generated shape follows the pictures closely (it is the one case where the image shapes the geometry directly), and the outline check reports how well it matches the sheet. See Turning a concept into a detailed 3D body.

Concepts that read as functional parts (a bracket, a box, a mount) are built the classic way, from sketches and extrusions, exactly as before. See Organic shapes with the agent.

Costs and limits

Each concept image costs credits (by default 5 in Fast and 8 in Smart), so a board of four costs four times that. It is a cheap way to find a direction, and it is cheaper than iterating on 3D geometry you are not sure about. Concept prompts are moderated, and the images you create count toward your storage.

When to skip Concept

If you already have exact dimensions, a sketch, a photo, or a drawing, skip to Build. The agent can start from those directly. See Prompts with images.

Try it

Run a short experiment:

Four concept images of a compact desk lamp with an adjustable arm, in matte white with a walnut base.

Lock one, then ask in Build:

Build this lamp. Give me a parameter table of the dimensions you assumed from the image and mark any you are unsure about.

Compare the table with the picture. This is the gap that your real-world numbers need to fill.