Skip to content

HAVC installer v2.4.0

Choose a tag to compare

@dan64 dan64 released this 09 Oct 17:03
· 156 commits to main since this release

HAVC installer v2.4.0 — qwen21-viggle colorizes about 2× faster, and colorize runs resume where they stopped.

Fast Pipeline on the qwen21-viggle family now splits colorization in two cached phases: each frame's text-encoder conditioning is computed once and cached (ref_te/), then the sampling runs with the UNet resident — ~3.0 s/frame against ~6.1 s/frame on a 708-frame field test. Frames are processed in chunks of 200 and a chunk's cache is deleted once its frames are colorized, so storage stays bounded and a cancelled run resumes from the first missing frame. The standard path is ~1 s/frame faster too (the unused negative text-encoder pass is skipped), and its runs now skip frames already colorized in ref_qwen/. All five qwen21-viggle configs move to the int8 ConvRot text encoder: faster and cleaner output in side-by-side tests. The GUI also keeps the B&W preview updated during the infer phase.

Start here:

  • HAVC-Setup-2.4.0.exe — HAVC Manager installer wizard (single-file, self-contained: no .NET runtime needed). Windows SmartScreen may warn (Windows protected your PC → More info → Run anyway); the exe is not code-signed yet. sha256: 8d05628e38649177412c28870c38ac312421b08e090dc5b970fc452aaf56b314
  • havc-2.4.0-py3-none-any.whl — HAVC server stack wheel: the content the manager installs (its version follows the installer). sha256: 6339d88eaa72bf52951b73ecee0c9d7758b1400bcb4497e017f1a7ffc6a2c3a1
  • release.json — the release manifest read by the installer/manager (sha256 for every artifact). Pinned packages and the ComfyUI runtime now live in the stable asset release HAVC Assets v1.0.2 (diffusers, spatial_correlation_sampler, comfy_bridge_v0.32.zip); tools.zip/NVEncC stay in v1.0.0; vscmnet2 comes from dan64/vs-cmnet2.

Requirements: Windows 10/11, NVIDIA RTX 30/40/50, CUDA 13.0+. With Fast Pipeline enabled, the qwen21-viggle pipeline now runs comfortably on GPUs with 12 GB of VRAM (peak usage around 11 GB); in standard mode the requirement is unchanged (14 GB+). Model weights are downloaded on first use and preserved across updates.

Notes: installer releases (and their wheel) stay on the 2.x.x line; the 1.x.x tags are reserved for asset buckets. Updates are incremental and keep the existing installation folder: the new comfy_bridge_v0.32.zip is re-extracted automatically (the runtime zip has a new file name; local model files are preserved), and the viggle configs ship with the wheel — if the manager reports customized configs, accept Update configs to pick up the int8 text encoder (a .bak of the previous file is kept). The new text-encoder weights are downloaded on first use. In case of problems, the update rolls back by itself (the previous wheel stays in cache\); Repair recovers a partially applied update.

Docs: docs/ in the repository.