diffuseR

• Quantized artifacts stay locally built. Prebuilt weights are not hosted for any model: only flux2 and zimage could be redistributed (Apache-2.0 and ungated), while LTX-2.3 is under the LTX-2 Community License and FLUX.1-schnell is gated, so hosting would cover half the catalog and leave two models on a different workflow.
• Model residency: resident_load(), resident_activate(), resident_deactivate(), resident_generate(), resident_status() and resident_unload() keep a pipeline's weights page-locked on the host and treat the GPU copy as disposable, so handing a small card between models is a DMA transfer rather than a full reload. Same contract as whisper and chatterbox, with no gpu.ctl dependency. This sits above the per-generation phase offloading in the txt2img_* functions: those swap one component at a time within a render, residency decides who owns the card between renders. For a phase-offloading pipeline (the default) activation is the ownership claim and the transfers stay per-phase; only a pipeline loaded with phase_offload = FALSE is copied to the card wholesale, and that path is checked against free VRAM first. resident_status() reports components_on_gpu alongside state, because the two legitimately disagree: a render returns every component to pinned host memory as its phase ends, so an active handle can hold nothing. Verified on an RTX 5060 Ti (16 GB) against local artifacts: flux2 (11.22 GB pinned, 9.1 s render), flux1 (15.73 GB, 40.1 s), zimage (13.45 GB, 19.5 s) and ltx (18.41 GB across 5 components). All three image models reproduce bit-for-bit across a deactivate/activate cycle. FLUX.1 and LTX both have pinned sets larger than the card, so they exercise the refusal path rather than bulk onload.

Addressing the CRAN review of the 0.2.0 submission:

• Every exported .Rd with a \usage block now documents its return value: 50 @return tags added, chiefly to the nn_module generators for the FLUX, FLUX.2, Z-Image, LTX-2.3 and Gemma3 ports.
• Examples: 14 of the 23 \dontrun{} blocks now run during check, and were rewritten to be self-contained instead of referencing undefined objects. The 9 that remain need model weights on disk and are itemised in cran-comments.md.
• ddim_scheduler_create() was uncallable at its documented defaults: beta_schedule was never passed through match.arg(), so switch() errored on the length-3 default, and the device default was a length-2 vector that torch_tensor() rejects. ddim_scheduler_step() had the same missing match.arg() on prediction_type. Every internal caller passed these explicitly, so the broken defaults went unnoticed. device now defaults to torch_device("cpu").
• DESCRIPTION: software names single-quoted ('Python', 'Stable Diffusion', 'Hugging Face' with its URL) and the trailing whitespace that had been folding into double spaces since the first commit removed.
• save_video()'s mp4 example is no longer live: the encoder inherits the session's stdin, which R CMD check --as-cran uses to feed the example script to R, so it consumed part of the script.