ARCHITECTURE

GENERATION

"Everything flows, nothing abides.”

Alci is the first digital art collection on Bittensor. Pieces are generated on mint - your piece is completely unique and never-before-seen upon creation, with the underlying automated process being heavily intricate. The intended piece output represents a 3D asset that was completely stripped of its mesh and its original point cloud, leaving a 3D network graph structure.

We seeded a collection of 256 unique assets and wrote 4 complex point manipulation algorithms. Upon mint, our smart contract executes a driver function that randomly selects an object from the 256 and randomly selects an algorithm to downsample the object and generate your piece. Upon execution, the object-algorithm pair is exhausted and can never be generated again.

SEEDS

"There exist only atoms and the void; everything else is opinion."

Structurally, we came up with 256 objects, where each object will go through each algorithm, giving us a total of 1024 pieces. Each object has two upper-level classifiers - the object’s category and the theme. So an example of this would be:

Theme: Animals & Plants
Category: Felidae
Object: Panther
Piece: Panther2 (algorithm 2 output)

The rarity breakdown of these objects is as follows:

We took our 256 objects and created prompts from them by attaching style modifiers and object-specific features. Each style modifier had a random weight between 1% to 10%, and was included in the prompt with 90% probability. The features were AI generated given the context of the object, and had an 80% chance of being added to the prompt.

Once we had our 256 prompts, we used SN-17 404.xyz/app, a text-to-3D generator originally intended for gaming assets, to programmatically generate our 3D objects. SN-17 allows you to use text input to guide 3D Gaussian Splatting and generate a PLY.

Miners of the subnet are rewarded for improving upon open-source SOTA 3D generation models, allowing the quality of 3D assets created with 404-GEN to improve as the industry improves. As a result, ALCI also serves as a snapshot of how good 404 generation was at the moment of this collection's launch.

While it's typically difficult to score inherently indeterministic mining results, 404-gen found a creative solution in using image references to anchor the evaluation process of miner outputs. More on this here.

ALGORITHMS

"Give me a place to stand, and I will move the earth."

Each algorithm we used followed a unique philosophical principle that we translated into point manipulation. Each object will go through all 4 algorithms.

Color Distribution

Algorithm 1 – Anarchy

Algorithm 1 embraces chaos and erasure. It represents anarchy in its purest form - destruction without reconstruction. Points are removed from the original object and are never redistributed. No new order arises.

Algorithm 1 heavily downsamples the original 3D asset based on a density filter, pruning the most crowded regions and leaving only .5% of the original points. Edges are added by connecting a given point to all points whose euclidean distance is less than a given radius parameter.

Step	Details
Downsample	Keep 0.5% of original points
Density Filter	Remove top 5% densest points (r = 0.1)
Structural Filtering	Keep 90% of largest component
Graph Reconnection	Radius-based (r = 0.08)

Algorithm 2 – World-Building

Algorithm 2 is more constructive and represents using existing points as a reference to create new points. These points are introduced either as mirrored counterparts or as generative extrapolations. It is a philosophy of creation from entropy.

Algorithm 2 has two branches of constructing new points in different areas of the space: There's a 70% chance of routing towards extrapolating points through duplication, rotation, and transformation, and there's a 30% chance of routing towards symmetrical mirroring logic, which enforces partial or full symmetry across the X, Y, or Z axis. Both routes connect each point to its nearest neighbors in a k-dimensional tree.

Step	Details
Downsample	Keep .3-.5% of original points
Density Filter	Redistribute high-density points to low (r = 0.1)
Mirroring / Duplication	30% chance of symmetry (mirror + rotation)
Transformation	70% chance of applying geometry perturbation
Structural Filtering	Keep 90% of largest component
Graph Reconnection	k-NN (k=30)

Algorithm 3 – Anger

Algorithm 3 is the most emotional. Rather than following rules of philosophy or form, this algorithm lashes out. It represents unfiltered rage, distorting shape and edges and becoming raw and fractured. This algorithm does not seek beauty, but catharsis.

Algorithm 3 has a 40% chance of applying a harsh and unstable downsampling, producing larger gaps but retaining more detail of the original object. The other 60% probability applies voxel downsampling with a voxel size between .01 and .15, causing severe fragmentation. This process is done by creating small voxels (3D cubes) and selecting a single point from each voxel to represent its contents.

Step	Details
Downsample	Keep 0.5–1% of original points
Transformation	Either point sampling distortions or voxel jittering
Graph Reconnection	i→i+1 chain

Algorithm 4 – Will to Power

Algorithm 4 is inspired by Nietzsche’s Will to Power. It operates as if each cluster of points is competing with itself, twisting and bending the form to its will. Rather than aiming for symmetry or chaos, this algorithm aims for influence.

This algorithm is much lighter on the downsampling compared to the other algorithms, retaining 20% of its original points. Random rotation angles between 10 and 180 degrees are selected, and each sine term is scaled by a random factor from .1, .5, and 1, causing asymmetric distortion. Each point is connected to its 6 nearest neighbors.

Step	Details
Downsample	Keep 10–20% of original points
Mirroring / Duplication	Half mirrored (scale range −0.8× to −1.2×)
Transformation	Sinusoidal warps, 10–180° rotations
Graph Reconnection	k-NN (k = 6)