Kitsmith

Style memory

Single-shot generators give you a different look every time. Kitsmith keeps a per-project style memory so the fiftieth asset still matches the first. It has three parts: pinned references, nearest-neighbour references, and a style LoRA trained on your approved work.

Pinned references

Pinned references are the images that define the kit. They come from approved style sheet concepts, images you upload, or any approved asset you pin. Up to 4 pinned references are sent with every job at full weight (1.0). If more than four are pinned, the most recently approved four are used.

Pin sparingly. Four images that agree with each other work better than four that pull in different directions.

Nearest-neighbour references

Every finished asset is rendered from four fixed angles under the same studio lighting, and those renders are embedded with DINOv2 (a 768-dimensional image embedding). The average embedding of all approved assets is the project’s style centroid.

For each new job, Kitsmith adds the 3 approved assets closest to the centroid, preferring assets of the same kind: a new prop looks at your most representative props first. These references get a lower weight between 0.35 and 0.75, scaled by how similar they are to the centroid. Pinned assets are not counted twice.

Result: the references for a job are always your pinned set plus the three assets that best represent the kit as it stands.

Style score

Each generated asset gets a style score: the cosine similarity between its embedding and the project centroid. Higher means closer to the established look. The project page plots scores over time as a drift chart. A slow downward trend means the kit is wandering; a single low point is usually one odd asset.

The score is a guide, not a verdict. A tree will always score a little differently from a chair. Compare scores within a kind, and trust your eyes.

What approve and reject do

  • Approve: the asset joins the centroid, becomes a candidate nearest-neighbour reference, and counts toward the next LoRA training run.
  • Reject: the asset is excluded from the centroid, from references and from training. If it was pinned, it is unpinned.
  • Assets waiting in review and failed jobs are never used as references.

Style LoRA

After every 15 newly approved assets, Kitsmith automatically queues a style training job (40 credits) that fine-tunes an SDXL LoRA on the renders of all approved assets. Once trained, the LoRA is applied to concept generation for every later job at a scale of 0.8, on top of the image references.

  • Each training run creates a new LoRA version. Jobs record which version they used.
  • Only one training run per project can be queued or running at a time.
  • If your balance is below 40 credits, the automatic run is skipped and you can start it manually later.
  • Training needs at least a couple of approved assets with renders.

Keeping a kit coherent

  1. Lock the look first. Do not generate a large batch until the style sheet and your pins feel right.
  2. Approve only what you would ship. Every approval shifts the centroid.
  3. Generate in waves: a few assets of each kind, review, then the next wave. Early approvals guide later jobs.
  4. Keep the art style text stable. Put subject detail in asset prompts, not in the spec.
  5. Watch the drift chart. If scores slide, re-pin your strongest assets and reject the outliers.
  6. Use negatives for recurring problems. If every prop grows glowing gems, add glowing gems to the kit spec negatives.