LoginSign Up
← Back to all sparks
S

sd2r

ANALYTICS
Velocity0.0

R package sd2r by zabis13 — release notes from GitHub.

github.com/zabis13/sd2r ↗

Local Stable Diffusion inference lands in R, shipped as Rcpp bindings over stable-diffusion.cpp

r-packageimage-generationlocal-inferencercpp-bindingstokenizers
◆Current state
sd2r is a young project wrapping stable-diffusion.cpp for R via Rcpp. The 0.1.0 release established the package structure and the core call surface — sd_ctx(), sd_txt2img(), sd_save_image() — with Vulkan GPU support behind a configure flag and a worked SD 1.5 example at 512x512. The two releases since are not code but asset bundles: precompiled tokenizer vocabularies and BPE merge tables shipped as header files, growing from four tokenizers to twelve.
◆Where it's heading
The asset releases are the more revealing half of this history. The first bundle covered CLIP, Mistral, Qwen and UMT5 — enough for SD 1.x through Flux. The second adds T5, Gemma, Gemma2 and GPT-OSS merges and splits UMT5 out as the Wan video encoder, so the tokenizer surface now reaches well beyond the image models the package currently exposes. Vocabulary support is being staged ahead of the inference paths that would use it.
◆Prediction
Given that tokenizers for Flux, SD3 and the Wan video encoder are already bundled while the documented API stops at txt2img, the next step is most likely exposing those model families through the R interface.

◆Recent moves

  1. 2mo ago

    Tokenizer bundle triples to twelve, adding Gemma, GPT-OSS and Wan

    Expands the shipped tokenizer assets from four to twelve header files, adding T5 for Flux and SD3, Gemma and Gemma2, GPT-OSS, Qwen2 merges, and splitting out UMT5 as the Wan video encoder. The vocabulary coverage now runs ahead of the package's documented inference surface, which points at where the bindings are headed.

    View source ↗
  2. 5mo ago

    First tokenizer asset bundle: CLIP, Mistral, Qwen, UMT5

    Publishes the initial four tokenizer vocabularies as downloadable header files, separating large binary assets from the source release. Packaging infrastructure rather than a change to what the package does.

    View source ↗
  3. 6mo ago

    First release: stable-diffusion.cpp bindings with Vulkan support

    ⚡ SPARK

    The founding release, wiring stable-diffusion.cpp into R through Rcpp with sd_ctx(), sd_txt2img() and sd_save_image(), plus optional Vulkan GPU acceleration. Everything after this has been tokenizer assets, making this the release that defines the package's entire capability surface.

    View source ↗