You can pick which model you perfer.
3b, 7b, 7b_sharp. All of them have fp8, fp16 and Q4/8 gguf version.
To make everytime simple, 3b is the smaller size model and 7b_sharp is the sharper version of 7b. The default model for video upscale is 3b_fp8.
The model will be download automatically when you first run the workflow.
Note that the new Version 2.5.0 that just release today has make lots of change to the perviouse version. Old workflow won’t work on current version without modification.
They have split the node to 4 nodes.
(DiT model, VAE model, torch.compile settings, and main upscaler for granular control)
The new Version 2.5 support torch.compile:
According to official claim, 20-40% DiT speedup and 15-25% VAE speedup with full graph compilation
But Triton is required for torch.compile with inductor backend
If you do not have Trition installed, you will get warning.
━━━━━━━━ Phase 1: VAE encoding ━━━━━━━━
[15:13:47.522] ❌ [ERROR] Cannot use torch.compile with 'inductor' backend: Triton is not installed.
Triton is required for the inductor backend which performs kernel fusion and optimization.
To fix this issue:
1. Install Triton: pip install triton
2. OR change backend to 'cudagraphs' (lightweight, no Triton needed)
3. OR disable torch.compile
For more info: https://github.com/triton-lang/triton
[15:13:47.523] ⚠️ [WARNING] torch.compile failed for VAE submodules: torch.compile with inductor backend requires Triton. Install with: pip install triton
[15:13:47.523] ⚠️ [WARNING] Falling back to uncompiled VAE
If you do not have Trition installed, you will get warning.
━━━━━━━━ Phase 2: DiT upscaling ━━━━━━━━
[15:13:49.856] ❌ [ERROR] Cannot use torch.compile with 'inductor' backend: Triton is not installed.
Triton is required for the inductor backend which performs kernel fusion and optimization.
To fix this issue:
1. Install Triton: pip install triton
2. OR change backend to 'cudagraphs' (lightweight, no Triton needed)
3. OR disable torch.compile
For more info: https://github.com/triton-lang/triton
[15:13:49.857] ⚠️ [WARNING] torch.compile failed for DiT: torch.compile with inductor backend requires Triton. Install with: pip install triton
[15:13:49.857] ⚠️ [WARNING] Falling back to uncompiled model
Note:
The embeded python in ComfyUI Portable does not have all the file required, which can casue file not found error when enable torch.compile.
cuda_utils.c:14: error: include file 'Python.h' not found
Fix:
Download Python 3.13 64-bit from python.org.
After install, copy all the files in C:\Users\your_name\AppData\Local\Programs\Python\Python313\include\ folder to D:\ComfyUI\python_embeded\include
This repository was created to address a common pain point for AI enthusiasts and developers on the Windows platform: building complex Python packages from source. Libraries like flash-attention, xformers are essential for high-performance AI tasks but often lack official pre-built wheels for Windows, forcing users into a complicated and error-prone compilation process.
The goal here is to provide a centralized, up-to-date collection of direct links to pre-compiled .whl files for these libraries, primarily for the ComfyUI community and other PyTorch users on Windows. This saves you time and lets you focus on what’s important: creating amazing things with AI.
Pick one that match your CUDA, Python, and Torch Version
For example:
Matches cu128 (CUDA 12.8), cp312 (Python 3.12), and base Torch 2.7.x.
