realrebelai's picture
Update README.md
79be5d5 verified
|
Raw
History Blame Contribute Delete
1.86 kB
metadata
license: cc-by-nc-sa-4.0
base_model: robbyant/lingbot-world-v2-14b-causal-fast
pipeline_tag: image-to-video
tags:
  - world-model
  - video
  - gguf
  - comfyui
  - wan
  - camera-control
  - low-vram

LingBot-World-v2 14B causal-fast — GGUF (8GB-VRAM ready)

GGUF quantizations of robbyant/lingbot-world-v2-14b-causal-fast for ComfyUI_Rebels_LingBotWorld — an action-controlled world model running on consumer GPUs (tested: RTX 3070 8GB / 16GB RAM).

You author a camera/movement track (or type one into the Action Builder node); the model renders the video that track produces from your start image, with genuine action following. Offline generation, chunked causal sampling, 4 distilled steps per chunk.

Files

File Notes
LingBot-World-14B-Q4_K_S.gguf (~11.7GB) tested tier; Q5-bumped v-projections keep effective bpw ~6.7
other tiers (Q4_K_M…Q8_0) quality ladder; RAM-streamed, VRAM use is unchanged

Also required: Wan2.1_VAE.pth, a UMT5-XXL GGUF encoder (loads via ComfyUI-GGUF CLIPLoader, type wan)

Settings that matter (8GB)

  • Resolution preset 256×448 (default) or 320×544; world-memory window 6+2
  • The KV cache is the world memory: it scales with window × resolution (upstream 18+6 @ 480×832 ≈ 21GB — does not fit consumer cards; the sampler pre-checks and refuses instead of hanging)
  • frame_num 4n+1; start at 21; export 16 fps

License

CC BY-NC-SA 4.0 (inherited from upstream): non-commercial, attribution, share-alike. Quantization is a format conversion only. Quants + nodes by RealRebelAI.

https://cdn-uploads.huggingface.co/production/uploads/68761990332d15464ccc8dee/OBwm3rX-WJ5s8a-EePAbK.mp4