Our goal is to turn a prompt into a human character in LifeShare’s soft-realism / Folk style: rounded proportions, readable features and a warm, crafted feel.
Kindred is being built around reviewed Meshy Folk reference models. The intended workflow is prompt, generate, inspect and refine, then prepare the character for animation. A public generator, downloads and a release date have not been announced.
TRELLIS text XL · Sparse-structure flow
This stage learns the broad 3D structure. We completed 100 training updates using four reviewed Folk reference surfaces, targeting proportions and overall form.
This is an early fit to a small dataset. It does not yet establish reliable results for new prompts.
TRELLIS text XL · Structured-latent flow
This stage handles finer geometry. The fine-tuning pipeline is prepared, but no new detail training updates have run: the full-body training targets must preserve faces, fingers and clothing first.
We are testing that quality gate before teaching the model from those targets.
The CLIP text encoder, DINOv2 image features, structured-latent encoder and mesh decoder remain pretrained and frozen. They interpret prompts or convert between images, compact 3D representations and meshes. We are adapting Microsoft’s TRELLIS stages, not training Meshy’s proprietary model. This pipeline does not require an OpenAI API key.
Full-body reconstructions softened facial features and merged fingers. Enlarged head-and-neck and hand-and-wrist tests recovered five separate fingers, more facial structure and the upper garment neckline. One brow and eyelid still lost definition; the neck remained coherent.
The proposed service is a website connected to GPU inference on Runpod. A saved trained model would serve many prompts; training would happen separately when a reviewed dataset is ready.
This hosted service is planned, not deployed. Capacity and generation cost still need measurement.
Blender would handle editing and rigging, with Unreal used to validate animation and game integration. Those delivery steps still need end-to-end validation.
Local generation is also a possible option, but we have not validated the finished pipeline on a 16 GB GPU.
Kindred stays separate from Silhouette, our Blender character-authoring tool. They can complement each other without becoming one product.