ALXOPENSOURCE/lingbot-video-1.3b-fp8

LingBot-Video Dense 1.3B — Selective Fused FP8 for ComfyUI

Current integration release: v1.0.2. The FP8 checkpoint is byte-for-byte unchanged from v1.0.1; v1.0.2 fixes Qwen structured-prompt normalization in the bundled ComfyUI node pack.

This repository contains a community-made selective FP8 transformer checkpoint and the matching ComfyUI integration for LingBot-Video Dense 1.3B.

It supports text-to-video, first-frame image-to-video, and an experimental first/last-frame workflow. This is an independent project and is not an official Robbyant release.

Important: this is the quantized transformer component, not a standalone pipeline. The processor, scheduler, text encoder, VAE, and other base-model components must be downloaded from the official model repository.

What is quantized

Native CUDA E4M3 W8A8 scaled matrix multiplication is used selectively for:

• fused attention Q/K/V projections
• fused MLP gate/up projections

Attention output, MLP down projection, conditioning, normalization, input, and output paths remain BF16. The original BF16 checkpoint is not modified.

Measured performance

On one RTX 5080, a matched cached 256×144×5-frame, one-step sampler smoke test measured:

Checkpoint	Sampler time
BF16	4.651 seconds
Selective FP8	3.651 seconds

That is 1.27× faster or approximately 21.5% less sampler time in this specific smoke test. This is not an end-to-end render benchmark or a universal GPU-family claim. See the complete benchmark notes.

Requirements

• ComfyUI
• NVIDIA CUDA GPU/runtime capable of PyTorch torch._scaled_mm with E4M3 tensors
• Matching CUDA-enabled PyTorch build
• Official robbyant/lingbot-video-dense-1.3b base components
• The included comfyui custom-node pack

Validated on an RTX 5080 with Windows/WDDM, Python 3.11.9, PyTorch 2.12.1+cu130, CUDA 13.0, Diffusers 0.38.0, Transformers 5.3.0, and SageAttention 2.2.0. Other configurations may work but have not been validated.

Installation

The full installation guide, dependency list, troubleshooting notes, and workflow documentation are in comfyui/README.md. The canonical source repository is ALX-CODE/lingbot-video-1.3b-fp8.

1. Install the custom node

Copy or clone the contents of the comfyui directory into:

ComfyUI/custom_nodes/ComfyUI-LingBotVideo/

Then install comfyui/requirements.txt using the Python environment that launches ComfyUI.

2. Download the official base components

Download the official base repository into:

ComfyUI/models/lingbot_video/lingbot-video-dense-1.3b/

You may exclude transformer/* when using only this FP8 checkpoint. The tested official revision is f9789a7d9b4772a47aba62d4eb5282ddefd1da21.

3. Download this FP8 transformer

hf download ALXOPENSOURCE/lingbot-video-1.3b-fp8 `
  --include config.json diffusion_pytorch_model.safetensors `
  --local-dir ComfyUI/models/lingbot_video/lingbot-video-dense-1.3b/transformer_fp8_dense

The expected SHA-256 digests are recorded in SHA256SUMS.txt.

Included supporting material

• config.json and diffusion_pytorch_model.safetensors: FP8 transformer
• comfyui/: custom nodes, runtime code, downloader/conversion tools, tests, and documentation
• comfyui/workflows/: compact distributable T2V, TI2V, and experimental FLF workflows
• comfyui/example_workflows/: editable ComfyUI workflow variants
• LICENSE, NOTICE.md, and SHA256SUMS.txt: licensing, attribution, and integrity metadata

FlashVSR and RIFE are optional and disabled by default in the public workflows. When enabled, FlashVSR runs before RIFE.

Limitations

• The Hub Inference API is disabled because this repository requires custom ComfyUI runtime code and external base-model components.
• First/last-frame conditioning is experimental.
• The preview demonstrates successful generation; it is not a BF16/FP8 quality-equivalence comparison.
• Quality and speed depend strongly on resolution, duration, step count, attention backend, GPU, and memory pressure.

License and attribution

Apache-2.0. This checkpoint is a quantized derivative of the official Dense 1.3B weights. The integration vendors and adapts portions of Robbyant/LingBot-Video. See NOTICE.md, LICENSE, and the upstream model repository for attribution.

Please cite the original LingBot-Video paper:

@article{ma2026lingbotvideo,
  title={Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence},
  author={Ma, Shuailei and others},
  year={2026},
  doi={10.48550/arXiv.2607.07675}
}