This is the official implementation for LERF.
LERF follows the integration guidelines described here for custom methods within Nerfstudio.
Follow these instructions up to and including "tinycudann" to install dependencies and create an environment
git clone https://github.com/kerrj/lerf
Navigate to this folder and run python -m pip install -e .
Run ns-train -h: you should see a list of "subcommands" with lerf, lerf-big, and lerf-lite included among them.
Now that LERF is installed you can play with it!
- Launch training with
ns-train lerf --data <data_folder>. This specifies a data folder to use. For more details, see Nerfstudio documentation. - Connect to the viewer by forwarding the viewer port (we use VSCode to do this), and click the link to
viewer.nerf.studioprovided in the output of the train script - Within the viewer, you can type text into the textbox, then select the
relevancy_0output type to visualize relevancy maps.
By default, the viewer shows raw relevancy scaled with the turbo colormap. As values lower than 0.5 correspond to irrelevant regions, we recommend setting the range parameter to (-1.0, 1.0). To match the visualization from the paper, check the Normalize tick-box, which stretches the values to use the full colormap.
The images below show the rgb, raw, centered, and normalized output views for the query "Lily".
The Nerfstudio viewer dynamically changes resolution to achieve a desired training throughput.
To increase resolution, pause training. Rendering at high resolution (512 or above) can take a second or two, so we recommend rendering at 256px
If your GPU is struggling on memory, we provide a lerf-lite implementation that reduces the LERF network capacity and number of samples along rays. If you find you still need to reduce memory footprint, the most impactful parameters for memory are num_lerf_samples, hashgrid levels, and hashgrid size.
lerf-big provides a larger model that uses ViT-L/14 instead of ViT-B/16 for those with large memory GPUs.
Be mindful that code for visualization will change as more features are integrated into Nerfstudio, so if you fork this repo and build off of it, check back regularly for extra changes.
Please open Github issues for any installation/usage problems you run into. We've tried to support as broad a range of GPUs as possible with lerf-lite, but it might be necessary to provide even more low-footprint versions. Thank you!
- Integrate into
ns-rendercommands to render videos from the command line with custom prompts
We've designed the code to modularly accept any image encoder that implements the interface in BaseImageEncoder (image_encoder.py). An example of different encoder implementations can be seen in clip_encoder.py vs openclip_encoder.py, which implement OpenAI's CLIP and OpenCLIP respectively.
(TODO expand this section)
The main file to look at for editing and building off LERF is lerf.py, which extends the Nerfacto model from Nerfstudio, adds an additional language field, losses, and visualization. The CLIP and DINO pre-processing are carried out by pyramid_interpolator.py and dino_dataloader.py.