H3VM v0.21.1 LAB-PZ1 Predictor Zoo

H3VM v0.21.1 LAB-PZ1 Predictor Zoo

Experimental comparison package for H3VM predictor research.

This release contains 9 predictor / scheduling algorithm variants prepared for side-by-side testing under the same H3VM v0.21.1 baseline.

Purpose

The goal of this LAB build is to compare different predictor strategies under consistent conditions and evaluate their impact on:

• generation speed
• predictor drift
• boundary stability
• temporal consistency
• image quality
• GPU utilization
• failure / fallback behavior

Package

H3VM_v0.21.1_LAB-PZ1_Predictor-Zoo.zip

Includes 9 experimental algorithm variants for comparative testing.

Status

⚠️ LAB / Experimental

This package is intended for benchmarking and research only.
It is not a stable production release and does not replace the official H3VM v0.21.1 release.

Test results and selected candidates will be documented separately after validation.

H3VM v0.21.1

H3VM v0.21.1 Release Notes

v0.21.1 is a product-surface cleanup and compatibility release: fewer modes, stronger isolation, a focused Core engine, new Fast-mode prediction choices, and adjustable secondary-GPU participation.

Main changes

1. Improved compatibility with newer ComfyUI multi-GPU startup behavior. Dual-GPU users should launch with --cuda-device all; H3VM does not require additional plugin-specific launch flags as long as both target CUDA devices remain visible.
2. Independent / lazy runtime isolation. Installing H3VM no longer means unrelated workflows should inherit its runtime behavior. H3VM activates when its execution path is actually used.
3. Product modes reduced to three practical dual-GPU modes: Fast, Capacity and Background. Single GPU remains available by disabling multi-GPU support.
4. Fast mode adds multiple predictors. SPECTRAL is the new default fast predictor; LINEAR remains available for content-dependent compatibility and A/B testing.
5. Adjustable secondary-GPU participation. 100% / 75% / 50% / 25% / custom 1–100% supports equal and heterogeneous GPU combinations.
6. Removed the compatibility-heavy all-in-one Master Loader from the public surface. H3VM is now a focused Loader-replacement execution layer: existing Model Loader -> H3VM Core -> existing workflow.

Under the hood

• Mode4 run boundaries are owned by ComfyUI OUTER_SAMPLE; expected_steps_hint is scheduling guidance only.
• Predictor startup avoids committing temporary step-index coordinate history before a real timestep is available.
• Dual Video VAE adds best-effort same-weight identity validation.
• Bundled MiniMax H3 reference workflow is standardized at 20 Steps / Hint 20.
• Newer ComfyUI heterogeneous-GPU visibility and startup preflight fixes are retained.
• Public node surface remains two nodes: H3VM Core + optional H3VM Video VAE.

Recommended first run

• Start ComfyUI with --cuda-device all for dual-GPU use.
• Insert H3VM Core after the existing Model Loader.
• Use Multi-GPU Fast + SPECTRAL + 100% participation first.
• Use Capacity for VRAM-limited workloads and Background when the PC also needs to remain responsive for browsing, editing or other foreground work.