realrebelai/LingBot_ComfyUI

LingBot-Video-Dense-1.3B — ComfyUI repack (single-file, 8GB-VRAM ready)

Repackaged weights of robbyant/lingbot-video-dense-1.3b for use with the ComfyUI_Rebels_LingBot node pack. No weight modification — the original diffusers shards are merged/renamed into single files for ComfyUI's model folders. All configs ship inside the node pack.

Runs T2V and TI2V on an RTX 3070 8GB / 16GB RAM: the Qwen3-VL encoder is loaded on CPU, used, and freed before the 1.3B DiT (2.79GB bf16, fully VRAM-resident) starts.

Files

File	Put in	Size
LingBot_1.3b_DiT.safetensors	ComfyUI/models/diffusion_models/	~2.8 GB
LingBot_text-encoder.safetensors	ComfyUI/models/text_encoders/	~8 GB
LingBot_vae.safetensors	ComfyUI/models/vae/	~0.5 GB

The text encoder is the repo's Qwen3-VL, shards merged into one file (bit-identical tensors). lm_head.weight is intentionally absent — it is tied to embed_tokens; the node pack re-ties it at load.

Recommended settings

• 832×480, up to 81 frames, 40 steps, shift 3.0
• guidance 3.0 (t2v) / ~3.0 (TI2V — higher burns color on image runs)
• export at 24 fps
• VAE tiling off unless you must (temporal color/exposure drift on long clips)

~14 s/it on a 3070 at the settings above (sequential CFG); ≈10 minutes per 81-frame clip.

Notes & limitations

• The trained window is 81 frames; longer generations collapse.
• Square/other aspect ratios are out-of-distribution (letterboxing artifacts) — stay at 832×480.
• 1.3B-class video: modest motion complexity; rendered text (signs/labels) is not legible.
• TI2V requires the image to be fed to both the encode node (vision conditioning) and the sampler (first-frame latent) in the node pack's workflow.

License

Weights inherit the upstream LingBot-Video license (see license_link). This repo is a format repack only; review the upstream terms before commercial use or redistribution.

Repack + nodes by RealRebelAI.