Visual Tokenizers & VAEs
Discrete and continuous visual tokenizers, VAEs, and reconstruction modules, with decoder availability and latent conventions documented.
0.8B • UpdatedNote Apple AToken-So/D encoder, eight-codebook FSQ, semantic pooling, and pixel decoder. This is the upstream tokenizer used by InternLumina-U2, not a tokenizer newly trained by U2.
gatilin/bitdance_binary_tokenizer_d32c128
Image-to-Image • 0.4B • Updated • 48Note BitDance binary image tokenizer, D32/C128 variant, for image encoding and reconstruction.
gatilin/emu3_5_visual_codec
Image-to-Image • 0.5B • Updated • 30Note Emu3.5 VQ visual tokenizer and codec components. The language backbone is excluded.
gatilin/gear_ibq_tokenizer_16384
Image-to-Image • Updated • 40Note GEAR-IBQ image tokenizer and codec with a 16,384-entry codebook.
gatilin/hydra_tok_base_gan
Updated • 16Note HydraTok understanding/generation visual tokenizer and pixel reconstruction components, retaining native latent representations and branch interfaces.
gatilin/ibq_tokenizer_16384_gestalt
Updated • 31Note Gestalt-associated IBQ visual tokenizer with a 16,384-entry codebook, retaining native encoding, quantization, and pixel reconstruction interfaces.
gatilin/internvlu_qwenimage_vae
0.1B • Updated • 82Note Qwen-Image VAE used by InternVL-U, published separately from the understanding-side internvlu_vision_bridge.
gatilin/lance_wan22_vae
UpdatedNote Wan2.2 video VAE used by LANCE for pixel/video latent encoding and decoding. The LANCE understanding tower and generation backbone are excluded.
gatilin/llada2_0_uni_image_tokenizer
1B • Updated • 38Note LLaDA2.0-Uni vision encoder, VQ modules, and auxiliary conditioning embeddings. Outputs token IDs; the separate diffusion renderer and image VAE are excluded, so this package does not directly reconstruct pixels.
gatilin/longcat_next_visual_codec
Image-to-Image • 0.9B • UpdatedNote LongCat-Next visual codec with residual vector quantization and pixel encoding/decoding. The full language model is excluded.
gatilin/mammothmoda2_mammothtok_codec
Image-to-Image • 0.7B • UpdatedNote MammothModa2 MammothTok visual codec components. The complete unified-model backbone is excluded.
gatilin/mingtok_ming_univision_16b_a3b
0.7B • Updated • 15Note Continuous MingTok components from Ming-UniVision: encoding, semantic expansion, and pixel reconstruction. The language backbone and language-side latent prediction head are excluded.
gatilin/raev2_dinov3l_k7_general_codec
Image-to-Image • Updated • 1Note RAEv2 DINOv3-L semantic encoder and reconstruction decoder, K7 general variant. This is a representation autoencoder rather than a conventional KL-VAE.
gatilin/tokenflow_clipb_32k_visual_codec
Image-to-Image • 0.4B • UpdatedNote TokenFlow CLIP-B visual tokenizer with a 32K codebook and pixel encoder/decoder.
gatilin/uniar_rl_bsq_tokenizer
0.7B • Updated • 28Note UniAR-RL BSQ encoder, binary visual tokens, and decoded visual features. Pixel rendering still requires the separate SD3.5 decoder and VAE.
gatilin/unispace_pr_dinov2_visual_tokenizer
Updated • 71Note UniSpace PR-DINOv2 unified tokenizer with 768-dimensional semantic features, 128-dimensional reconstruction features, and a pixel decoder. Paired with the matching class-conditional DiT.
gatilin/unitok_large_visual_codec
Image-to-Image • 0.8B • Updated • 28Note UniTok Large visual tokenizer and reconstruction codec. The full language and generation backbones are excluded.
gatilin/vila_u_256_visual_codec
Image-to-Image • 1B • UpdatedNote VILA-U 256px visual codec with residual quantization. The language backbone is excluded.
-
gatilin/nld-uni-image-tokenizer
0.5B • Updated • 44