Change8

v0.18.0

📦 comfyuiView on GitHub →
9 features🐛 23 fixes1 deprecations🔧 9 symbols

Summary

This release focuses heavily on VRAM optimization, especially around LTX VAE decoding and tiled operations, alongside introducing support for mxfp8 precision and new command-line flags like --fp16-intermediates.

Migration Steps

  1. If using text encoders, be aware that the fix to run them on MPS GPU for Apple Silicon was reverted; check performance implications.
  2. If using seedream-3-0-t2i or seedance-1-0-lite nodes, plan to replace them due to deprecation.

✨ New Features

  • Support for mxfp8 precision.
  • Added --fp16-intermediates command line argument to use fp16 for intermediate values.
  • LTXV: Accumulate VAE decode results on intermediate_device.
  • Added essentials_category to nodes and blueprints for the Essentials tab.
  • Aligned local API with cloud specification for assets.
  • Added --enable-dynamic-vram option to force enable dynamic VRAM.
  • [Trainer] Enabled FP4, 8, 16 training via native dtype support and quant linear autograd function.
  • Added slice_cond and per-model context window cond resizing.
  • Added Quiver SVG nodes via API nodes.

🐛 Bug Fixes

  • Fixed Tencent TextToModel and ImageToModel nodes in api-nodes.
  • Fixed stale frontend chunks by using no-store cache headers.
  • Fixed issue where EmptyLatentImage did not follow intermediate dtype.
  • Enabled Pytorch Attention for AMD gfx1150 (Strix Point).
  • Skipped running model finalizers at exit.
  • Fixed potential issue related to data handling.
  • Implemented atomic writes for userdata to prevent data loss on crash.
  • Disabled SageAttention for Hunyuan3D v2.1 DiT.
  • Marked weight_dtype as advanced input in Load Diffusion Model node.
  • Reduced LTX VAE VRAM usage and saved use cases from OOMs/Tiler.
  • Reduced WAN VAE VRAM, saved use cases for OOM/Tiler.
  • Implemented inplace VAE output processing to reduce peak RAM consumption.
  • Fixed case where pixel space VAE could cause issues.
  • Removed dead weight init code in cascade.
  • Fixed VRAM leak in tiler fallback in video VAEs.
  • Fixed missing init variable in ltx vae.
  • Fixed regression issue.
  • Addressed issue where fp16 intermediates didn't work for some text encoder models.
  • Implemented chunked encoder + CPU IO chunking for LTX VAE to reduce Big VRAM.
  • Added more exclusion criteria to pinned read to fix corrupt outputs (rare cases).
  • Reduced tiled decode peak memory.
  • Fixed missing init variable in ltx vae.
  • Fixed issue where EmptyImage node did not follow intermediate device/dtype.

Affected Symbols

⚡ Deprecations

  • API Nodes: seedream-3-0-t2i and seedance-1-0-lite models are marked as deprecated.