H3 Low VRAM加速-实验版

2026-09-30 update — H3 Accelerator 0.3.0rc5

Download the ZIP for your operating system from the download-variant selector:

• Linux: Linux package (12.14 MB) — RTX 5080 uses SM120.

• Windows: Windows auto package (51.19 MB) — one package for RTX 3060 (SM86) and RTX 4060 (SM89). GPU selection is automatic; no separate card package or switch is required.

Both ZIPs contain the current rc5 runtime, precompiled kernels, matching native source, licenses, installation/model notes, and two blank UI workflows with API companions. Models, LoRAs and reference images are not included.

INSTALL

Close ComfyUI, move the old node folder outside custom_nodes, then extract the complete ComfyUI-H3-Accelerator folder into ComfyUI/custom_nodes. Restart ComfyUI. Import a UI JSON from workflows/, select your installed models, and supply your own reference/prompt. Keep loader and sampling stages on auto. Required workflow plugins: H3 Latent Upscaler and DaSiWa Nodes; download links are included in the workflow notes.

PRESETS / EDITABLE CONTROLS

Original workflow: 960×544 → 1632×928, 1.7×, recommended starting point for the 5080.

15-second / 8 GB workflow: 640×352 → 960×512, 1.5×, reference long edge 512 with match mode, 362 frames at 24 fps, 8+3 steps. Start here on Windows 3060/4060. Coarse dimensions, refinement multiplier, reference size and duration remain editable in the UI. These are presets, not GPU resolution limits.

MODELS

Use the REF2VA Singularity v1.3 INT8 main model, Qwen3VL INT4 ConvRot text encoder, REF2V Turbo LoRA at 1.0, and optional Realism People LoRA at 0.8 (trigger r34l1sm). Install BOTH H3 VAEs: minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors, plus the H3 latent upscaler. Exact filenames, folders and author download links are in the ZIP/workflow notes.

FIXES / MEASUREMENTS

rc5 coordinates provider weights and forward workspace ownership, includes query partitioning, and fixes paired-VAE discovery and video decoder generator lifetime. Preserve DynamicVRAM, pinned memory, async offload and the default reserve; do not stack official Sparse Attention.

Linux 5080 paired 15-second regression: 363.727 → 342.844 s with identical decoded video/audio. Windows 3060 Laptop 6 GB completed the 15-second preset in 832.392 s including new-cache compilation. RTX 4060 full-video time is not measured; 400–600 s remains a tuning target. The Windows ZIP uses the supported native binaries from the recorded acceptance, with their exact source revisions.

中文简要

本次直接替换现有版本文件。Linux 包适配本机 5080；Windows 一个包包含 3060/4060 内核并自动选择。两个 ZIP 都内置原始高分辨率与 15 秒/8 GB 工作流，以及匹配源码、安装和模型说明。解压完整节点文件夹到 custom_nodes，重启后从 workflows 导入，填写自己的参考图和提示词，保持 auto。Windows 建议先用 640×352、1.5 倍到 960×512、参考长边 512/match、362 帧、8+3；一二采分辨率、参考尺寸、倍数和时长均可在 UI 调整。4060 的 400–600 秒是目标，尚无实机计时。