開源生成式影像工具
InvokeAI 6.14.1 Patches the Trust Boundary for External Image Downloads and Corrects VAE and FP8 Memory Reservations
InvokeAI’s patch release tightens trust checks for Alibaba Cloud image downloads and injected sessions while correcting memory estimates for FP8, FLUX.2, and VAE workloads. The new version also supports Krea-2 LoRAs with DoRA magnitude, but the project has not provided a vulnerability identifier or quantifiable VRAM comparisons.
InvokeAI released version 6.14.1 on September 6. Although it appears to be a minor patch in the 6.14 series, its changes focus on external data boundaries and GPU memory management, making it particularly important for self-hosted, multi-user deployments and installations connected to cloud image providers. PyPI provides source distributions and wheels uploaded via Trusted Publishing, with support for Python 3.11 through 3.12.
On the security front, the project says it has closed a vulnerability in Alibaba Cloud image download handling. The full changelog describes the fix as adding safeguards for external downloads and requiring injected sessions to have trusted status. This continues the tightening introduced in 6.14.0 around non-public URLs and download paths: generative-image workflows often combine remote URLs, provider-returned objects, and local assets into a single data flow. If the backend retrieves arbitrary locations on a user’s behalf, it can create risks of server-side requests or cross-tenant data access. Operators should therefore treat this upgrade as a security patch, not merely a UI update.
For memory management, version 6.14.1 fixes an FP8 loader that did not fully honor the skip list, excessive RAM reservation, and inaccurate working-memory estimates for FLUX.2 denoising and VAE encoding and decoding. The goal is not to reduce the model weights themselves, but to allow the cache to free space more accurately ahead of peak operations, reducing cases where a model loads successfully only to encounter an out-of-memory error during the VAE stage. On shutdown, the service now evicts cache records instead of only releasing shared weights. Video uploads are also streamed to a single on-disk copy, reducing intermediate duplication for large assets.
For model compatibility, the release can load Krea-2 LoRAs generated by ai-toolkit that include DoRA magnitude, validating their structure by restoring flattened Kohya keys. Workflows can also be exported directly as PNG screenshots, making node graphs easier to review and share. However, the project has not published a CVE, an affected-version matrix, or attack reproduction steps, nor has it provided before-and-after RAM or VRAM benchmarks. Users should still conduct regression testing with their own models, quantization formats, and batch settings.