Generating Fluid Simulations

This documentation details how to generate fire and water simulations like those shown in the Infinigen launch video. It assumes you have already completed [Installation.md], [HelloWorld.md]. Fluid simulations require significant computational resources.

Setup

To generate fluids, you run install.sh with the optional FLIP-Fluids setup step (bash install.sh flip_fluids), or, please run bash worldgen/tools/install/compile_flip_fluids.sh to install flip fluids now.

Example Commands

Generate a video of a single scene with simulated fire generated on the fly

python -m tools.manage_datagen_jobs --specific_seed 3930249d --output_folder outputs/fire --num_scenes 1 --pipeline_config local_256GB.gin monocular_video.gin --cleanup none --config plain.gin fast_terrain_assets.gin use_on_the_fly_fire.gin

Because fluid simulation takes a long time, the fire resolution can be reduced in use_on_the_fly_fire.gin, by setting set_obj_on_fire.resolution = {resolution}. This will reduce the fire quality but speed up the simulation.

Generate a video of a single valley scene with simulated river

python -m tools.manage_datagen_jobs --specific_seed 1c95bb72 --output_folder outputs/river --num_scenes 1 --pipeline_config local_256GB.gin monocular_video.gin opengl_gt.gin cuda_terrain.gin --pipeline_overrides iterate_scene_tasks.frame_range=[150,200] --config river.gin simulated_river.gin no_assets.gin no_creatures.gin fast_terrain_assets.gin --cleanup none 

Similar to fire, the simulation can be sped up by reducing the resolution. In simulated_river.gin, the resolution can be modified by setting make_river.resolution = {resolution}. The simulation can also be sped up by reducing the simulation duration in simulated_river.gin by setting make_river.simulation_duration = {duration}. For instance, before running the command above, the duration can be reduced to a number greater than 200 since that is the last frame of the video.

Also, note that this command will produce a scene without assets to speed up the process. However, the liquids generally interact with the objects by splashing on them, and a scene like this can be produced by removing the no_assets.gin option in the above command.

Generate videos of random scene types, with simulated fire generated on the fly when needed

python -m tools.manage_datagen_jobs --output_folder outputs/onthefly  --num_scenes 10 \
    --pipeline_config slurm_high_memory.gin monocular_video.gin \
    --config fast_terrain_assets.gin use_on_the_fly_fire.gin \
    --cleanup none --warmup_sec 12000 

Generate videos of valley scenes with simulated rivers

python -m tools.manage_datagen_jobs --output_folder /n/fs/pvl-renders/kkayan/river --num_scenes 10 \
    --pipeline_config slurm_high_memory.gin monocular_video.gin opengl_gt.gin cuda_terrain.gin \
    --pipeline_overrides iterate_scene_tasks.frame_range=[100,244] \ 
    --config simulated_river.gin no_assets.gin no_creatures.gin fast_terrain_assets.gin \
    --cleanup none --warmup_sec 12000

Using Pre-Generated Fire

High-resolution fluid simulations take a long time, so assets on fire can pre-generated and imported in the scene instead of being baked on-the-fly. This allows for fire to be simulated once, instead of every time a scene is generated. To pre-generate fire assets of all types (bush, tree, creature, cactus, boulder), run:

python tools/submit_asset_cache.py -b $BLENDER -f {fire_asset_folder} -n 1 -s -40 -d 184

where fire_asset_folder is where you want to save the fire. The number of assets per type, start frame and duration can be adjusted.

If you only want to pre-generate one type of asset once, run:

$BLENDER --background -noaudio --python fluid/run_asset_cache.py -- -f {fire_asset_folder} -a {asset} -s {start_frame} -d {simulation_duration}

where fire_asset_folder is where you want to save the fire. asset can be one of CachedBushFactory, CachedTreeFactory, CachedCactusFactory, CachedCreatureFactory, CachedBoulderFactory.

Import pre-generated fire when generating a scene

After fire is pre-generated with one of the previous commands, edit config/use_cached_fire.gin and set the FireCachingSystem.asset_folder variable to fire_asset_folder you used when pre-generating fire. After this use_cached_fire.gin can be used instead of use_on_the_fly_fire.gin when generating a scene. This will import the fire from the folder it is saved instead of simulating it on-the-fly.

Example Command

python -m tools.manage_datagen_jobs --specific_seed 3930249d --output_folder outputs/fire --num_scenes 1 --pipeline_config local_256GB.gin monocular_video.gin --wandb_mode online --cleanup none --config plain.gin fast_terrain_assets.gin use_cached_fire.gin