RTX 4090 Freezes in ComfyUI: New CK Attention Suspected of VRAM Spikes
Foreign_Fee_6036 · reddit · 2026-08-13
A developer using an RTX 4090 (with 128GB RAM) reported a severe VRAM leak issue when enabling the new CK attention in the latest version of ComfyUI while running the Minimax H3 I2V workflow.
Specifically, while the first generation is indeed faster than without CK attention, VRAM usage skyrockets from a normal 70-80% to 99% during the second generation, completely freezing the PC until the ComfyUI process is forcibly terminated. The developer confirmed the issue persists across a clean ComfyUI installation, CUDA 13, and the newest PyTorch.
More from Multimodal
- Video Generation Models Struggle with Complex Physical Interactions Like Handcuffing — flowersslop · 2026-08-13
- Testing LTX 2.5 Video Generation: A Children's Story Workflow — pfeifits · 2026-08-13
- GPT-4o Demonstrates Impressive 'Pixel Forensics' Capabilities — teortaxesTex · 2026-08-13
- Qualcomm VP on Image Gen Bottlenecks: Separating Scene Planning from Rendering — TWIML AI Podcast · 2026-08-13
- Runway Integrates xAI's Grok Imagine Image 2.0 Model — runwayml · 2026-08-13
- Video Diffusion Models Enable Single-Image Refocusing — CSProfKGD · 2026-08-13