QuantStack's Wan2.2 Brings Alibaba's Image-to-Video AI to 8GB GPUs
QuantStack's GGUF conversion of Wan2.2-I2V-A14B shrinks a 27B Mixture-of-Experts video model into quantizations that fit on consumer GPUs through ComfyUI.
- QuantStack released GGUF quantizations of Wan2.2-I2V-A14B, hitting 831K+ downloads
- Converts a 27B Mixture-of-Experts video diffusion model into 2-bit through 8-bit formats
- Sizes range from 5.3GB (Q2_K) to 15.4GB (Q8_0), fitting consumer GPUs
- Runs through ComfyUI via the ComfyUI-GGUF custom node by city96
- Base Wan2.2-I2V-A14B generates 5-second 480P or 720P clips from a source image
- Apache 2.0 license, supports English and Chinese prompts, bundled with a MoE denoising architecture
Hugging Face showed more than 831,000 downloads for the QuantStack conversion of Alibaba’s Wan2.2-I2V-A14B when checked. The community release packages the 14-billion-parameter image-to-video model as quantized GGUF files, making local inference possible on consumer GPUs through reduced precision and CPU offloading. Hugging Face’s counter records file requests and can include repeat downloads.
Full BF16 weights for a 27B model occupy roughly 54GB before runtime allocations for attention, the VAE, text encoder, latents, and ComfyUI. QuantStack’s release compresses the model to between 2-bit and 8-bit formats. Users can trade fidelity and speed for a smaller memory footprint, with practical options for 8GB, 12GB, 16GB, and 24GB cards.
Two experts change the download math
The original model card describes a Mixture-of-Experts architecture with separate high-noise and low-noise experts. QuantStack provides a GGUF file for each expert at every quantization level, so a complete setup requires a matching pair. The derivative retains the original Apache 2.0 license.
| Quantization | Per expert | Expert pair | Practical trade-off |
|---|---|---|---|
| Q2_K | 5.3GB | 10.6GB | Smallest footprint, with visible losses in texture and motion detail |
| Q3_K_M | 7.18GB | 14.4GB | Suited to constrained hardware with offloading |
| Q4_K_M | 9.65GB | 19.3GB | Common balance between memory use and output quality |
| Q5_K_M | 10.8GB | 21.6GB | Retains more detail while increasing memory pressure |
| Q6_K | 12GB | 24GB | Higher fidelity with substantial offloading on many cards |
| Q8_0 | 15.4GB | 30.8GB | Largest option and closest to the source weights |
The pair column estimates disk space for both experts. Peak VRAM depends on ComfyUI’s loading strategy, resolution, frame count, attention implementation, system RAM, and whether inactive components move to the CPU. A quantized file’s size therefore provides a starting point rather than a complete hardware requirement.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
