Independent guide · Linux + CUDA · Last verified: July 12, 2026

How to run LingBot-Video locally

Everything you need to get LingBot-Video generating on your own hardware: exact version requirements, the model-size decision, and the ready-made inference scripts — with the memory caveats the README buries.

1 · Check the requirements

Note the CUDA 13.0 nightly PyTorch build — this is a bleeding-edge stack. The root requirements.txt pins the recommended build for you.

2 · Pick your model size honestly

• Dense 1.3B (single GPU). The realistic choice for one consumer/prosumer GPU. Official single-GPU scripts target it directly (base generation, no refiner). Supports T2I, T2V, TI2V.
• MoE 30B-A3B + Refiner (multi-GPU). Only 3B params activate per step, but the whole 30B checkpoint loads. Official scripts shard it with FSDP + context parallelism (CP8) — and each rank first builds the transformer in host memory, so a large-RAM machine is mandatory.

No official VRAM numbers yet. Robbyant hasn’t published per-GPU VRAM requirements. Anything you read elsewhere quoting exact figures is guesswork — the guidance above reflects what the official scripts and README actually imply.

3 · Install

git clone https://github.com/Robbyant/lingbot-video
cd lingbot-video

python -m venv .venv
source .venv/bin/activate
python -m pip install -U pip

# Base requirements: direct DiT inference + rewriter (transformers backend)
pip install -r requirements.txt
pip install -e .

# Optional: SGLang Diffusion / fused & FP8 MoE runtime
python -m pip install --no-deps -r requirements-sglang.txt

4 · Run your first generation

Set the model paths, rewrite your prompt (see the prompt guide — this step is not optional), then generate:

# 1. Point to your downloaded weights
export MODEL_DIR="<path_to_lingbot-video-model>"
export REWRITER_BASE_MODEL="<path_to_rewriter_base_vlm>"
export REWRITER_ADAPTER="<path_to_rewriter_lora>"

# 2. Plain prompt → structured JSON caption
python rewriter/inference.py --backend transformers --mode t2v \
  --prompt "<plain_user_prompt>" --duration 5 --output prompt.json

# 3. Generate (reference parameters from the official README)
python scripts/inference.py \
  --backend diffusers \
  --model_dir "$MODEL_DIR" \
  --mode t2v \
  --prompt_json prompt.json \
  --output out/base.mp4 \
  --height 480 --width 832 --fps 24 \
  --steps 40 --guidance_scale 3 --shift 3 \
  --transformer_dtype bf16 --text_encoder_dtype bf16 --vae_dtype fp32

Or use the ready-to-run scripts:

export DENSE_MODEL_DIR="<path_to_lingbot-video-dense>"
export MOE_MODEL_DIR="<path_to_lingbot-video-moe>"

# Single GPU (Dense; base generation, batched CFG)
MODEL_DIR="$DENSE_MODEL_DIR" ./scripts/single-gpu/run_dense_t2v.sh

# Multi-GPU (FSDP + CP8; MoE T2V with refiner)
MODEL_DIR="$MOE_MODEL_DIR" ./scripts/multi-gpus/run_moe_t2v_refiner_fsdp_cp8.sh

All scripts accept environment overrides like PROMPT_JSON, HEIGHT, WIDTH, STEPS, GUIDANCE_SCALE, SEED, FPS, and BACKEND. MoE scripts default to grouped expert execution (LINGBOT_MOE_EXPERT_BACKEND=grouped_mm).

5 · Choose a backend

• diffusers — the direct reference path; works with the base install.
• sglang — SGLang Diffusion runtime with optional fused / FP8 MoE kernels; needs requirements-sglang.txt. Falls back to diffusers automatically if SGLang isn’t installed.
• --enable_fsdp_inference — shards base and refiner DiT across GPUs to relieve VRAM pressure after load.

About ComfyUI

There is no official ComfyUI integration as of our last check. Community interest is high, so this may change quickly — we update this page when it does.

Frequently asked questions

Stuck on output quality rather than setup? It’s almost certainly the prompt format — read the prompt guide. Back to the LingBot-Video model guide.