A Character Turnaround-Sheet Workflow

A diffusion model trained on internet images has learned to generate plausible-looking characters, but it has no persistent internal representation of "this specific character." Every generation call is independent. If you prompt "cartoon robot, front view" and then "cartoon robot, side view," you will get two images that both look like cartoon robots but differ in antenna shape, color distribution, arm proportions, and dozens of other details. For game assets or animation this is unusable — animators need the side view to be geometrically consistent with the front view because they trace or rig directly from the sheet.

The solution is to pass the front-view design as an explicit reference image into every subsequent generation node. By conditioning each angle on the approved front view, the model generates side, back, and three-quarter angles that reflect the reference design rather than hallucinating a new one. This technique is sometimes called image-conditioned multi-view generation, and it is the architectural backbone of the turnaround-sheet workflow in Floniks.

The turnaround workflow follows a two-phase structure. Phase one is the front-view generation node, which runs first and produces the canonical design. Its output becomes the reference input for phase two. Phase two consists of four parallel generation nodes — one each for side (90°), three-quarter (45°), back (180°), and optionally a low-angle or action-pose view. Each phase-two node receives the front-view image as a reference via an image conditioning input and uses a prompt that specifies only the angle: "same character, exact side view, neutral T-pose, flat lighting, white background, no shadow."

In the Floniks editor, wire the front-view node's output port to the reference image input port on all four parallel nodes simultaneously using Shift-click batch wiring. Set a dependency edge from the front-view node to each parallel node so they do not begin execution until the front view is complete and available. This guarantees the reference image is present before any angle generation starts, which is a requirement — not an optional optimization.

Angle specification in prompts is more nuanced than simply appending "side view." Diffusion models often interpret "side view" loosely, producing a three-quarter angle rather than a true 90° orthographic projection. To get precise angles, use explicit degree notation and reinforce with compositional constraints: "exact 90-degree side profile view, character facing right, both arms visible, legs parallel, no foreshortening."

For the back view: "exact 180-degree rear view, character facing away from camera, full body visible, same color palette as reference, no new design elements." For the three-quarter: "45-degree front-right three-quarter view, character's right shoulder closer to camera, face partially visible." Each prompt should also contain the universal constraints: "white background, neutral flat lighting, no cast shadow, character centered in frame, same scale as reference image." Adding negative prompts such as "perspective distortion, different costume, different colors, different proportions" further anchors the model to the reference design.

Most image-conditioned generation models expose a reference strength or image influence parameter — sometimes called IP-Adapter weight, reference image strength, or conditioning scale. For a turnaround sheet, set this value high (0.80–0.95) to maximize fidelity to the front-view design. A value below 0.70 gives the model too much creative latitude and will result in color or proportion drift between views.

If the model you are using in Floniks exposes a style lock parameter separately from a structure lock parameter, set structure lock high and style lock to a moderate value (0.65–0.75). This keeps the character's silhouette and proportions faithful to the reference while allowing the model to naturally adapt the surface rendering to the new angle — which is necessary because lighting and material behavior legitimately change between front and back views. Test this calibration by running the workflow with a known reference character and spot-checking that the ear shape, shoulder width, and belt line are identical between the front and side views before committing the settings to the template.

After all four parallel view nodes complete, route their outputs to a Style Consistency Validator node. This node computes a perceptual similarity score between each view and the front-view reference. Configure a minimum threshold of 0.72 — outputs below this threshold are likely to have significant design drift. Route passing outputs to the final Output Collector node and failing outputs to a Retry Branch that modifies the reference strength upward by 0.05 and regenerates the failing view.

For the final output, arrange the four views in a 2×2 grid or a horizontal strip using a Composite Layout node configured with equal spacing and a thin guide-line overlay. This composite is the deliverable turnaround sheet, suitable for handing off to an animator or rigging artist. Save the full workflow — including the composite layout settings — as a named template so future character projects only require updating the front-view prompt and the reference image input.