ComfyUI VRAM Leak Issue Sparks Community Investigation into Memory Management

ComfyUI VRAM Leak Issue Sparks Community Investigation into Memory Management

Users of ComfyUI, the node-based interface for Stable Diffusion, are raising alarms over a persistent VRAM retention issue that forces frequent software restarts when using large language models (LLMs) for prompt generation. A Reddit user with a 24GB GPU reported that while their workflow—combining LLM Party to load a GGUF model with image generation nodes—functions perfectly on the first run, subsequent executions fail due to insufficient memory. Despite using ComfyUI’s built-in unload model node and manual cache-clearing buttons, VRAM usage remains unchanged, only fully releasing upon full application restart. This behavior has ignited a broader discussion within the AI art community about memory handling in complex, multi-model workflows.

According to Zhihu’s comprehensive ComfyUI configuration guides, the platform is designed for efficiency and low VRAM consumption, even on GPUs under 3GB, thanks to its modular, node-based architecture. Unlike WebUI, which is noted for its high memory footprint, ComfyUI’s strength lies in its ability to isolate and manage individual computational tasks through discrete nodes. However, this granular control also exposes potential vulnerabilities when third-party plugins, such as LLM Party, interact with the core engine. The Zhihu articles emphasize proper environment configuration, proxy removal, and manual model management as best practices—but do not explicitly address LLM memory persistence, suggesting a gap in documentation for hybrid LLM+diffusion workflows.

Experts within the AI art community suspect the issue stems not from ComfyUI’s core codebase, but from how external LLM integration nodes handle PyTorch tensor caching and GPU context cleanup. When an LLM is loaded via GGUF through LLM Party, the model’s weights and intermediate activations may be allocated to VRAM using non-standard memory pointers that bypass ComfyUI’s internal memory tracker. Consequently, the ‘Unload Model’ node, designed to clear Stable Diffusion checkpoints, fails to recognize or release these foreign allocations. This results in a cumulative VRAM drain that only a full application reset can resolve.

Further analysis reveals that similar issues have been reported in other node-based AI frameworks, such as Nuke and DaVinci Resolve, where third-party plugins introduce memory leaks by holding onto GPU contexts beyond their lifecycle. In ComfyUI’s case, the lack of a unified memory manager for non-Stable Diffusion models—such as LLMs, control nets, or upscalers—creates an architectural blind spot. While ComfyUI excels at orchestrating image generation pipelines, its memory management system was primarily optimized for diffusion models, not multi-modal LLM integrations.

Community members have proposed temporary workarounds: manually forcing garbage collection via Python scripts after each workflow run, or using separate GPU instances for LLM and diffusion tasks. Some developers are also exploring patches to extend the unload node’s functionality to detect and clear GGUF model tensors. Meanwhile, the ComfyUI core team has not yet issued an official statement, though GitHub issue trackers are beginning to see increased reports.

For users, the lesson is clear: while ComfyUI offers unparalleled flexibility, its power comes with responsibility. Complex workflows involving external models require deeper awareness of memory allocation patterns. Until official support for LLM memory cleanup is integrated, users should treat LLM loading as a high-cost operation and plan workflows accordingly—perhaps by batching prompts or using lightweight models to reduce VRAM strain.

As AI workflows grow increasingly hybrid—blending language, vision, and audio models—the need for robust, cross-model memory management becomes critical. ComfyUI’s architecture is ahead of its time, but its memory subsystem must evolve to match. For now, the community’s vigilance and technical ingenuity remain the best defense against invisible VRAM leaks.