// the find
lllyasviel/ControlNet-v1-1-nightly
Nightly release of ControlNet 1.1
Official nightly release of ControlNet 1.1 — a set of 14 conditioning models (depth, canny, openpose, segmentation, lineart, tile, etc.) that let you steer Stable Diffusion 1.5 generation with structural inputs instead of just text. Aimed at ML researchers experimenting with diffusion control, not end users — the README itself tells you to go use the A1111 extension instead.
The changelog per model is unusually honest about training data bugs from 1.0 (duplicated greyscale images, mislabeled prompts, JPEG artifacts) and what specifically was fixed, which is rare transparency for a research drop. The Tile model addresses a real, well-known problem — prompt bleed across tiled upscaling — with a documented mechanism rather than a hand-wavy claim. Each model ships a standalone gradio script so you can test a specific conditioning type without pulling in the full webui stack.
The repo actively discourages its own primary use case: it tells you not to install it directly and to use a separate third-party project (Mikubill/sd-webui-controlnet) for anything practical, which makes this repo mostly a model-weights changelog rather than usable software. It vendors a large chunk of detectron2 and OneFormer as committed subdirectories instead of pinned dependencies, so you inherit unpatched, uncredited copies of someone else's codebase. Despite a 2024 last push, the 'nightly' branch was never merged back into the main ControlNet repo as promised, and everything is hard-locked to SD 1.5 with no SDXL or newer-architecture path. Model weights aren't fetched or versioned by the repo — you manually download multi-GB checkpoints from HuggingFace and hope the filenames still match the code.