Skip to main content
Back

RealRebelAI Shrinks Alibaba's Wan2.2-Animate to Run on 8 GB GPUs

Community quantizations by RealRebelAI shrink Alibaba's 30 GB character animation model into GGUF files that fit on 8-12 GB consumer cards.

  • RealRebelAI published GGUF quantizations of Alibaba's Wan2.2-Animate-2-14B character animation model.
  • Files range from Q2_K (~5.9 GB) up to Q8_0 (~17.4 GB), with Q4_K_M as the recommended tier.
  • Enables the 30 GB bf16 model to run on 8-12 GB consumer GPUs via ComfyUI.
  • Requires the custom Rebels W3A8 Loader to avoid a silent failure that loads it as Wan 2.1.
  • Precision-sensitive layers like timestep and conditioning projections are preserved during quantization.
  • Companion encoder, VAE, and CLIP vision files come from Comfy-Org's repackaged repo.

Alibaba's Wan2.2-Animate-2-14B is one of the more capable open character animation models around, but the full bf16 weights weigh in near 30 GB, which puts it out of reach for most consumer GPUs. A community contributor going by RealRebelAI has published a set of GGUF quantizations of that model on Hugging Face, along with a custom ComfyUI loader node that works around a subtle metadata problem that would otherwise break the model silently.

Wan2.2-Animate-14B is a unified model for character animation and replacement with holistic movement and expression replication, released with weights and inference code by Alibaba's Wan-AI team. You upload a still photo of a character and a driving video with the desired motions, and the system extracts poses and masks, then animates the character in one of two modes: an animation mode that applies the movements and expressions from the drive video onto the static photo, producing a video where the character mimics the same gestures and angles, plus a replacement mode that swaps a character into an existing scene.

Why the quantizations matter

The original checkpoints are simply too large for most people to run locally. The RealRebelAI repo exists so the same model runs on 8 to 12 GB cards instead of requiring the full 30 GB bf16 version, with all original licensing terms carrying over from the base model under Apache 2.0.