Azuma–Hoeffding inequality Bellman optimality equation

Markov decision process Algorithm 1: Q-learning

exploration–exploitation trade-off optimism in the face of uncertainty

upper confidence bound by Hoeffding's inequality,

Thompson sampling regret minimization

Lai–Robbins lower bound w.p. at least 1 − δ,

minimax regret contraction mapping theorem

temporal-difference learning by the contraction mapping theorem,

policy gradient theorem actor–critic methods

trust-region policy optimization Lemma 1 (performance difference).

maximum-entropy reinforcement learning Boltzmann exploration

eligibility traces telescoping over t = 1, …, T:

stochastic approximation martingale difference sequence

sub-Gaussian concentration taking expectation over s′ ∼ P(·|s,a),

union bound Chernoff bound

importance sampling exploration vs. exploitation:

off-policy evaluation occupancy measure

contextual bandits Q.E.D. □

adversarial bandits online convex optimization

mirror descent Algorithm 1: Q-learning

follow-the-regularized-leader Lyapunov stability

sample complexity by Hoeffding's inequality,

value iteration · policy iteration dynamic programming principle

Bellman operator w.p. at least 1 − δ,

Robbins–Monro conditions Bellman backup · bootstrapping

generalized advantage estimation by the contraction mapping theorem,

natural policy gradient Fisher information matrix

Kullback–Leibler divergence Lemma 1 (performance difference).

successive elimination doubling trick

best-arm identification telescoping over t = 1, …, T:

pure exploration confidence radius

Bernstein inequality taking expectation over s′ ∼ P(·|s,a),

McDiarmid's inequality optional stopping theorem

← Projects / 02Field manual · 说明书

A character-to-skin pipeline

BLOCK.

From character concepts to Minecraft-ready skins.

The system first synthesizes structured preview images and then converts them into valid skin atlases with a deterministic decoder.

Specification Technical record

Method: Bi-stage MLLM pipeline
Output: Minecraft-ready skin atlases
Decoder: Deterministic
Paper: arXiv 2603.03964
Model release: BLOCKv0.5 / Hugging Face
Project dated: March 2026

BLOCK is an open-source bi-stage MLLM pipeline for translating character concepts into Minecraft-ready skins. The system first synthesizes structured preview images and then converts them into valid skin atlases with a deterministic decoder.

I’m iterating and working on generating Minecraft skins end-to-end. If you are also interested, feel free to contact me.

Current Results

Hugging Face BLOCK example 1 preview — Example 1. This is the current effect on a relatively simple character concept.

Hugging Face BLOCK example 1 generated skin — Example 1. This is the current effect on a relatively simple character concept.

Hugging Face BLOCK example 2 preview — Example 2. This is the current effect on a more detailed fantasy-style character with fine-grained clothing and hair cues.

Hugging Face BLOCK example 2 generated skin — Example 2. This is the current effect on a more detailed fantasy-style character with fine-grained clothing and hair cues.

Hugging Face BLOCK example 3 preview — Example 3. This is the current effect on a harder sports-jersey example where some preview details are still compressed.

Hugging Face BLOCK example 3 generated skin — Example 3. This is the current effect on a harder sports-jersey example where some preview details are still compressed.