ULBDVC 超低码率确定性视觉通信

Ultra-Low Bitrate Deterministic Visual Communication — live demo · Inventor: Jingsong Deng · Patent pending (provisional)

渐进式码本 Progressive Codebooks合成分析闭环 Analysis-by-Synthesis非对称边缘算力 Asymmetric Edge Compute确定性·零幻觉 Deterministic · No Hallucination
📷 图像 Image→编码端:查共享码本 Encoder: shared codebook lookup→只传索引 Indices only (193 B → 1409 B)→解码端:查表重建 Decoder: table lookup→逐比特一致 Bit-exact output
或选择样例 or pick a sample:
编码中 Encoding…
工作原理 How it works (v2). 编码端与解码端预先共享固定的乘积量化(PQ)码本(约 0.24 MB,从不传输)。① 合成分析:先用最小二乘求出“经固定双线性插值放大后最接近原图”的缩略图(亮度 32×32、色度 16×16)。② 乘积量化:每个 4×4 亮度块拆成两个子向量,各查 256 项子码本;色度单独码本。③ 渐进残差:每一阶段只编码上一阶段剩下的误差。④ 领域码本:编码端试遍“自然 / 图形文档”两套码本,选误差最小者,用 1 字节头告知解码端。
阶段 1:193 B(0.024 bpp) · 阶段 2:385 B(0.047 bpp,<0.05) · 阶段 3:897 B(0.109 bpp) · 阶段 4:1409 B(0.172 bpp)。解码只是查表 + 固定矩阵乘法——相同字节 = 逐比特相同图像,无生成模型、无随机、无“脑补”。
Encoder and decoder share fixed product-quantized codebooks (~0.24 MB, never sent). (1) Analysis-by-synthesis: solve for the thumbnail that, after a fixed bilinear upsampler, best matches the original. (2) Product quantization: each 4×4 luma block is split into two sub-vectors, each indexed in its own 256-entry codebook; chroma has its own codebook. (3) Progressive residuals: each stage codes only what the previous one missed. (4) Domain codebooks: the encoder tries natural vs graphic/document sets and keeps the lower-error one, signalled in a 1-byte header. Decoding is table lookup plus fixed matrix products — same bytes ⇒ bit-identical image.