FlashVSR+ for ComfyUI · T8

用于 ComfyUI-FlashVSR-Plus-T8 的模型文件,包含 Turbo、Base、V1 官方权重与 INT8 ConvRot 派生权重。

Official checkpoints and INT8 ConvRot derivatives for the companion ComfyUI nodes. Original floating-point files are preserved byte-for-byte; SHA-256 checksums and source revisions are recorded in model_files.json.

快速使用 / Quick start

在 ComfyUI Manager 安装 FlashVSR+ T8,重启后使用仓库中的 VIDEO 或 IMAGE 示例工作流。Model Loader 可按所选模式与解码器自动下载文件。

Install FlashVSR+ T8 through ComfyUI Manager, restart, and import a VIDEO or IMAGE example workflow. The model loader downloads the components required by the selected preset.

手动下载时保持以下结构,放入 ComfyUI/models/flashvsr_plus,或在 Model Loader 指定模型根目录:

flashvsr_plus/
  TCDecoder.ckpt
  Wan2.1_VAE.pth
  turbo/
    dit.safetensors
    dit_int8_convrot.safetensors
    lq_proj_in.safetensors
  base/
    dit.safetensors
    dit_int8_convrot.safetensors
    lq_proj_in.safetensors
  v1/
    dit.safetensors
    dit_int8_convrot.safetensors
    LQ_proj_in.ckpt
配置 / Configuration 所需文件 / Required files
Turbo / Base · BF16 or FP16 Corresponding dit.safetensors, lq_proj_in.safetensors, and decoder
V1 · BF16 or FP16 v1/dit.safetensors, v1/LQ_proj_in.ckpt, and decoder
INT8 ConvRot Corresponding dit_int8_convrot.safetensors, LQ projection, and decoder; original DiT is optional
Tiny decoder TCDecoder.ckpt
Wan VAE Wan2.1_VAE.pth

每个 INT8 DiT 约 1.45 GB。仅 DiT block 的 Linear 采用 Comfy Kitchen W8A8 ConvRot;注意力、LQ 投影及解码器保留浮点计算。INT8 节省权重显存,速度与画质取决于内容和硬件。

Each INT8 DiT is approximately 1.45 GB. Only transformer-block linear layers use Comfy Kitchen W8A8 ConvRot; attention, LQ projection, and decoders remain floating point. Performance and quality vary with hardware and content.

验证 / Validation

配套节点已完成 45 项回归测试;Turbo + Tiny + Triton 的 BF16、INT8 均有 1920×1080、5 秒、24 FPS、120 帧真实 ComfyUI 验证,逐帧解码、时间戳及音频检查通过。transformers==4.57.1 另有隔离环境完整工作流验证。本节点不直接依赖 transformers。详见 验证说明。

The companion nodes passed 45 regression tests. Real 1080p, five-second ComfyUI workflows were checked for all 120 frames, timestamps, and audio; the tested configuration is Turbo / Tiny / Triton with BF16 or INT8. See the linked validation notes for scope and platform limits.

来源与许可 / Sources and license

  • FlashVSR+ Turbo / Base, Tiny decoder, Wan VAE: OpenImagingLab/FlashVSR_Plus, official checkpoints, revision 46b79520946ac67dec94d1566490690953c7debc.
  • FlashVSR V1.1: official checkpoints, revision 27561b186ded3402d7c975f4fd722e2885b6135f.
  • INT8 ConvRot derivatives and ComfyUI integration: T8star. All model files are provided under Apache-2.0, consistent with the source model cards. See LICENSE.

社区与资源 / Community

B站 · YouTube · API · 免费画廊 · 在线 AI 应用 · ComfyUI 整合包 · Hugging Face

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for t8star/Flashvsr-Plus-Comfy

Finetuned
(5)
this model