Three-View Character Sheets: Stepwise Generation and Stitching Practice

In character concept design, costume variant validation, and modeling reference image production, three-view sheets are an important intermediate asset. A three-view sheet usually includes a front view, a side view, and a back view. It is used to express body proportions, garment structure, hair accessory placement, color relationships, and key identifying elements in a stable way. In practice, costume variants such as magical girl outfits, nun habits, and school uniforms are well suited to rapid comparison through three-view sheets. When using general text-to-image services to generate three-view sheets, a common approach is to generate multiple views at once on a single canvas. However, this approach often leads to cross-view inconsistency, distorted hand structure, discontinuous outline lines, local color patches or noise, and other issues.

Based on a general image generation workflow, this article introduces a more controllable approach: split the three-view sheet into three single-image generations to obtain the front view, side view, and back view separately, then complete stitching through unified resolution, baseline alignment, color correction, and local repainting. This approach trades an increased number of generation calls for greater correctability and more stable final quality. It is suitable for concept design workflows that require high character consistency.

1. Background: Why One-Shot Multi-View Generation Is Unstable

The core problem with one-shot three-view generation is compounded task complexity. The model must handle multiple views, multiple poses, costume details, and background constraints at the same time. The following issues are common:

1. Insufficient cross-view consistency: Hair accessories, belts, buttons, or patterns present in the front view may be missing, deformed, or shifted in the side and back views.
2. Anatomical distortion: Hands, fingers, joints, and garment folds are high-frequency error areas. When multiple views appear on the same canvas, these errors are more likely to be amplified.
3. Discontinuous lines and outlines: Outline thickness, closure, and smoothness may be inconsistent across views.
4. Local color patches and noise: Small areas on a large canvas are prone to abnormal color blocks, which affect subsequent stitching and use.
5. High correction cost: Once one view has a problem, the entire image often needs to be redrawn, making it difficult to repair only a local area.

Therefore, decomposing a complex task into multiple lower-complexity subtasks is a direct way to improve controllability. With per-view single-image generation, each view can be checked independently, retried independently, and locally repaired independently before entering the stitching stage.

2. Overall Workflow: Single-View Generation + Unified Specifications + Post-Stitching Processing

The recommended workflow is as follows:

Character specification -> Front view baseline generation -> Side view generation -> Back view generation
        -> Quality check -> Size and baseline alignment -> Color correction -> Local repainting
        -> Three-image stitching -> Final review and archiving

The front view serves as the baseline view. It should first establish the character’s height proportions, head size, shoulder width, waistline, main garment colors, and key identifying elements. The side view and back view should reference the same character specification in their prompts and should keep the same canvas ratio, background strategy, and lighting description as much as possible. The stitching stage should not simply place images side by side. It should first unify dimensions, foot baseline, head center, and background, and then perform necessary local repairs.

The following schematic pseudocode uses placeholder interface names only:

function generate_three_views(roleSpec):
    frontPrompt = buildPrompt(roleSpec, view="front")
    front = imageService.generate(frontPrompt, seed=roleSpec.seed, size=roleSpec.canvas)
    sidePrompt = buildPrompt(roleSpec, view="side", reference=front)
    side = imageService.generate(sidePrompt, seed=roleSpec.seed, size=roleSpec.canvas)
    backPrompt = buildPrompt(roleSpec, view="back", reference=front)
    back = imageService.generate(backPrompt, seed=roleSpec.seed, size=roleSpec.canvas)
    checked = qualityGate([front, side, back])
    aligned = alignBaselineAndScale(checked)
    corrected = colorCorrect(aligned)
    repaired = localRepaint(corrected, focus=["hands", "edges", "colorPatch"])
    return assembleHorizontally(repaired)

3. Prompt Design: Break the Character Specification into Reusable Fields

Per-view generation does not mean describing each view arbitrarily. Instead, it requires a unified character card. It is recommended to split the prompt into the following fields:

• Character identity: Age impression, body type, hairstyle, hair color, eye color, and signature accessories.
• Costume style: Magical girl outfit, nun habit, school uniform, etc., with clear main colors, secondary colors, materials, layers, and decoration placement.
• View description: Front view, side view, or back view, with clear orientation and focus of presentation.
• Pose constraints: Standing pose, arms naturally down or slightly spread, avoiding occlusion of key structures.
• Canvas and composition: Full-body framing, margin ratio, foot baseline, and head center position.
• Background and lighting: Solid-color background, even lighting, low shadows, reducing color differences after stitching.
• Negative constraints: Extra limbs, fused fingers, broken outlines, abnormal color patches, and complex background interference.

When generating different views, change only the “view” field and keep the other fields as consistent as possible. The side view should emphasize the side silhouette, garment thickness, and the side structure of hair accessories. The back view should emphasize back garment construction, the back of hair accessories, and color continuity. Field-based prompts reduce description drift across views.

4. Consistency Control: Use the Front View as the Anchor

Consistency control is key to stepwise generation. A “front-view anchor method” is recommended:

1. Establish the baseline first: The front view should pass quality checks first, confirming that proportions, costume structure, and hands have no obvious errors.
2. Reuse the character card: The side view and back view should reuse the same character card without adding new decorations on the fly.
3. Reference the baseline image: If the generation service supports reference images or image-conditioned input, the front view can be used as a reference to reduce color and structural drift.
4. Fix the random seed: If the service supports a fixed seed, keep the seed consistent during debugging to observe differences caused by view changes.
5. Build a checklist: Check item by item whether hair accessories, neckline, cuffs, belts, skirt hem, footwear, and back decorations are consistent across views.

Note that a fixed seed does not guarantee complete consistency. It is only an auxiliary means of reducing random variation. The truly stable approach is a combination of character cards, reference images, quality gates, and local repainting.

5. Stitching and Post-Processing: Align First, Then Blend

Stitching is not simply placing three images horizontally. The recommended processing order is as follows:

• Unify the canvas: Use the same resolution and ratio for all three images, with consistent character height proportion.
• Align baselines: Use the sole of the foot as the horizontal baseline and the head center as the vertical reference to avoid height mismatch.
• Crop margins: Keep consistent spacing between views and avoid areas that are too dense or too sparse.
• Color correction: Unify white balance, brightness, and saturation to reduce background color differences.
• Edge inspection: Check outline continuity and perform local repainting when necessary.
• Priority repair: Prioritize hands, cuffs, skirt hem edges, and color patch areas rather than redrawing the entire image.

Local repainting should be limited to specific regions and prompts so that repairing one area does not damage other parts that have already passed inspection. After repair, run the quality gate again to form a closed loop.

6. Quality Assessment and Cost Trade-Offs

Quality assessment should cover five dimensions: anatomical structure, costume details, cross-view consistency, outline smoothness, and color stability. Pass/fail criteria can be defined for each dimension, and failure samples should be retained for prompt iteration.

Generating three single images means roughly three times the generation calls of a one-shot approach, with a corresponding increase in cost. Whether this is worthwhile depends on the intended use. If the three-view sheet is only used for sketch reference, standards can be relaxed appropriately. If it is used for modeling reference or costume variant review, the correctability brought by stepwise generation is more valuable. From an engineering perspective, cost can be controlled through caching approved views, batching task queues, setting retry limits, and applying tiered quality strategies.

7. Engineering Implementation Recommendations

In team collaboration, the workflow can be split into general capabilities such as task queues, object storage, logging services, and monitoring and alerting. Source images, prompts, seeds, parameters, and review results should be archived together. Failure rates and average retry counts should be included in monitoring. Version records should be retained for each modification to support traceability. For costume variants, maintain separate configurations for character cards and costume cards. After the front view passes, batch-generate the side view and back view.

8. Conclusion

The approach of generating views separately and then stitching them is essentially about trading task decomposition for controllability. It cannot completely eliminate structural errors in image generation, but it can significantly reduce the complexity of each generation and make hand distortion, uneven outlines, and local color patches easier to detect and repair locally. Combined with field-based prompts, a front-view anchor, unified stitching specifications, and quality gates, it can form a reusable production workflow for virtual character three-view sheets. For scenarios that require stable output of character concept images, costume variant images, and modeling reference images, this is a general technical path worth practicing.