Companion repo link returns 404; usage snippet doesn't run standalone
Hi RockTalk, thanks for shipping the first MLX port of Lance and publishing the bf16 checkpoint — great work.
Two notes for users following the README:
The companion repo link returns HTTP 404. https://github.com/RockTalk/Lance-MLX (and the embedded reference to tools/lance_t2i.py) is unreachable. Is it private, moved, or pending a public release?
Without it, the Usage snippet doesn't run. The imports
from lance_mlx.lance import Lance, LanceConfig
from lance_mlx.vae_wan22 import Wan2_2_VAE
can't resolve (and there's no lance-mlx package on PyPI — pypi.org/pypi/lance-mlx/json 404s). The snippet also references undefined symbols (lance_cfg, text_ids), and a comment in the snippet itself says "see tools/lance_t2i.py in the companion repo for the full builder" — but that file lives in the 404'd repo. Net effect: anyone trying to use the checkpoint from this HF page alone hits an ImportError on line 2.
The same issue affects the image variant at RockTalk/Lance-3B-MLX (its README references the same companion repo).
For context: I'm porting Lance to MLX in parallel at mlx-community/Lance-3B-Video-bf16 (Apache-2.0). I verified your weights are numerically equivalent to ours — remapping your F32 keys into our layout + casting to bf16 produces byte-identical pixel output through our pipeline. So the checkpoint itself is verified ✅; it's just hard to load standalone right now.
Your published ocean-wave sample at 256² × 9f (24 steps, CFG=4) is also reproducible through our pipeline at the same config — looks comparable to your sample. Happy to share that grid if useful.
If the companion repo is intended to be open and just temporarily missing — no rush. If publishing it is on your roadmap, would you consider either:
- publishing a minimal
pip installable package (even a tarball release would help), or - dropping an end-to-end
inference.pydirectly in this HF repo?
Either would unblock community use. Thanks again!
Hey @xocialize — thanks for the careful, friendly report and for cross-verifying the weights are byte-identical through your pipeline. Really nice work on the parallel port.
The companion github repo was never made public — totally my oversight. Just fixed by going with your second option: bundled lance_mlx/ + a self-contained inference.py directly in each HF repo. The Usage snippet now actually runs:
hf download RockTalk/Lance-3B-Video-MLX --local-dir Lance-3B-Video-MLX && cd Lance-3B-Video-MLX
pip install -r requirements.txt
python inference.py --prompt "a calm ocean wave rolling onto a sandy beach"
(For T2I, use RockTalk/Lance-3B-MLX with the same flow — --size 512 --steps 30 recommended.)
The inference.py auto-fetches RockTalk/Wan2.2-VAE-MLX on first run (the bundled vae.safetensors in this repo uses an older "nested-conv" keying that doesn't strict-load with the current Wan2_2_VAE class — kept for archival). Verified end-to-end by pulling the code fresh from HF and generating a new image.
Yes, please share your ocean-wave grid — happy to compare side-by-side and link to your repo as a cross-validated peer port. I've added a "Related ports" pointer to mlx-community/Lance-3B-Video-bf16 in both READMEs already.
Thanks again — this is exactly the kind of friendly QA that keeps the MLX ecosystem honest.