Suddenly SeedVR2 gives me OOM errors where it didn't before

A significant regression has emerged in the SeedVR2 video generation ecosystem, where users with 8GB VRAM GPUs are suddenly facing Out-Of-Memory (OOM) errors on workflows that previously executed successfully. The issue is highlighted by a user running ComfyUI 0.14.1 on a Windows machine with an NVIDIA RTX 5060 Laptop GPU (8GB). Despite using the same portable ComfyUI version, workflow, and the memory-efficient Q6_K GGUF variant of the 7B parameter SeedVR2 model, the system now fails where it worked just days prior.

Technical logs show the pipeline attempting to initialize with the DiT (Diffusion Transformer) and VAE models set to use CUDA, with CPU offload enabled for both components. The system reports ample free VRAM (6.69GB) before model loading, but crashes occur during the inference phase. Crucially, the user confirmed the problem persists even after rolling back to older SeedVR2 releases and trying an ultra-lightweight 3B parameter model, which should theoretically guarantee functionality on 8GB hardware. This points to a fundamental change in how recent SeedVR2 builds manage memory allocation or model loading, rather than a simple increase in model size.

The context makes this particularly disruptive. SeedVR2, as a state-of-the-art video generation model, is often used by creators and researchers on consumer-grade hardware. The GGUF format (via llama.cpp) was specifically designed to enable large models to run on limited VRAM by intelligently offloading layers to system RAM. The fact that this fails indicates a breakdown in this offloading mechanism or an unexpected VRAM overhead in the new build. For the AI tooling community, this serves as a warning about the instability of nightly builds and the challenges of maintaining compatibility in fast-moving open-source projects, potentially halting creative work until a fix is deployed.