GTX 970 (Maxwell) + ComfyUI in 2026: Stable Working Setup, Safe Upgrade Path, and Rollback Guide

Desktop Feedback Windows
,
1

Yes, ComfyUI can still work on an NVIDIA GTX 970 (4 GB VRAM) in 2026.

After a careful migration and validation process, this setup was confirmed working with:

  • GPU: NVIDIA GeForce GTX 970

  • VRAM: 4 GB

  • OS: Windows

  • Python: 3.10.10

  • PyTorch: 2.5.1+cu121

  • Torchvision: 0.20.1+cu121

  • Torchaudio: 2.5.1+cu121

  • NumPy: 2.2.6

  • ComfyUI core: v0.19.1-1-g1de83f91

  • Frontend package: 1.43.1

  • Launch mode: --lowvram

This is not a “latest everything” guide.
It is a stability-first guide for people with older GPUs who want a safe and reproducible path.

The Problem

Many GTX 970 users run into one or more of these issues when trying to update ComfyUI:

  • New ComfyUI versions launch, but the interface looks old

  • PyTorch/CUDA combinations break older stable setups

  • Updating Python creates unnecessary incompatibilities

  • Custom nodes fail and make it look like the whole installation is broken

  • Users assume the GTX 970 is no longer viable

The good news:

The GTX 970 is still usable with ComfyUI

You just need a conservative stack and a safe update workflow.

What Actually Worked

This setup was tested successfully:

  • Python 3.10.10
  • torch 2.5.1+cu121
  • torchvision 0.20.1+cu121
  • torchaudio 2.5.1+cu121
  • numpy 2.2.6
  • ComfyUI core v0.19.1-1-g1de83f91
  • comfyui_frontend_package 1.43.1
  • Launch option: --lowvram

Three workflows were tested successfully after the update.

Important Notes Before You Start

1. Do not rush to upgrade Python

If your goal is stability, Python 3.10.10 is a good choice for this hardware and this stack.

2. Do not chase the newest CUDA stack blindly

A GTX 970 is an older card. The objective is working reliably, not chasing every new optimization.

3. Separate “core works” from “custom nodes are broken”

Some custom nodes may fail to load, and ComfyUI itself can still work perfectly fine.

Safe Upgrade Strategy

The safest approach is:

  1. Keep your old working branch

  2. Create a test branch

  3. Update ComfyUI there

  4. Fix the frontend if necessary

  5. Validate real workflows

  6. Promote the tested branch to your new stable branch

That is exactly what was done here.

Step-by-Step Guide

Step 1 — Check your current branch

In Windows Command Prompt:

cd /d C:\AI\ComfyUI
git branch --show-current
git rev-parse HEAD

This tells you:

  • which branch you are on

  • which exact commit you are using

Step 2 — Create a test branch before touching anything

Example:

cd /d C:\AI\ComfyUI
git checkout -b test-comfyui-update

This gives you a safe place to test updates without destroying your stable branch.

Step 3 — Update dependencies carefully

If needed:

C:\AI\ComfyUI\.venv\Scripts\python.exe -m pip install -r manager_requirements.txt

Then launch ComfyUI:

cd /d C:\AI\ComfyUI
C:\AI\ComfyUI\.venv\Scripts\python.exe main.py --lowvram

Step 4 — Check if the frontend is outdated

This was the key issue in this case.

ComfyUI launched, but the UI still looked old.
The log showed this warning:

Installed frontend version 1.28.7 is lower than the recommended version 1.42.11.

That meant:

  • backend/core was updated

  • frontend package was still old

Check your installed frontend package:

C:\AI\ComfyUI\.venv\Scripts\python.exe -m pip show comfyui-frontend-package

If it is outdated, update it:

C:\AI\ComfyUI\.venv\Scripts\python.exe -m pip install --upgrade --force-reinstall comfyui-frontend-package

Check again:

C:\AI\ComfyUI\.venv\Scripts\python.exe -m pip show comfyui-frontend-package

In this tested case, it moved from:

1.28.7

to

1.43.1

After that, the UI updated correctly.

Step 5 — Ignore non-critical custom node failures at first

The following custom node issues appeared during startup:

  • APZmedia-comfyui-fast-image-save → turbojpeg library missing

  • comfyui_face_parsing → opencv-contrib-python-headless install error / permissions issue

  • ComfyUI_LayerStyle → missing blend_modes

Despite that, ComfyUI launched and tested workflows still worked.

So the rule is:

Do not assume startup warnings mean the core installation failed

First confirm whether your real workflows still function.

Step 6 — Validate with real workflows

Do not stop at “it opens”.

Test:

  • your usual workflow

  • a second workflow

  • a third workflow

That is what was done in this case, and all three worked.

That is the point where a test update becomes trustworthy.

Step 7 — Consolidate the tested branch into a stable branch

Once validated, create a clearly named stable branch:

cd /d C:\AI\ComfyUI
git checkout -b stable-gtx970-comfyui-0191

Confirmed working branch:

stable-gtx970-comfyui-0191

Confirmed commit:

1de83f91c34a3396c667d764224054ba87027e82

Step 8 — Create rollback tags

This is highly recommended.

Example:

git tag backup-pre-consolidation-stable-gtx970 92d97380bd02d9883295aeb2d29365cecd9a765e
git tag stable-validated-0191 1de83f91c34a3396c667d764224054ba87027e82

This makes future recovery much easier.

Step 9 — Save a freeze of the environment

Create a package snapshot:

C:\AI\ComfyUI\.venv\Scripts\python.exe -m pip freeze > C:\AI\ComfyUI\requirements-final-stable-comfyui-0191.txt

This is your insurance policy.

Step 10 — Save a local environment record

Create a text file with your known good state. Example:

Branch: stable-gtx970-comfyui-0191
Commit: 1de83f91c34a3396c667d764224054ba87027e82
ComfyUI core: v0.19.1-1-g1de83f91
Frontend: 1.43.1
Python: 3.10.10
torch: 2.5.1+cu121
torchvision: 0.20.1+cu121
torchaudio: 2.5.1+cu121
numpy: 2.2.6
GPU: NVIDIA GeForce GTX 970
Launch mode: --lowvram

This helps enormously later.

Final Working Configuration

GPU: NVIDIA GeForce GTX 970
Python: 3.10.10
torch: 2.5.1+cu121
torchvision: 0.20.1+cu121
torchaudio: 2.5.1+cu121
numpy: 2.2.6
ComfyUI core: v0.19.1-1-g1de83f91
Frontend: 1.43.1
Launch mode: --lowvram
Stable branch: stable-gtx970-comfyui-0191
Commit: 1de83f91c34a3396c667d764224054ba87027e82

What I Do Not Recommend

For GTX 970 users trying to keep a reliable setup, I do not recommend doing all of this at once:

  • upgrading Python unnecessarily

  • switching Torch versions randomly

  • changing CUDA stack and ComfyUI version at the same time

  • assuming every custom node must work immediately

  • deleting your old stable branch too early

That path usually leads to chaos wearing a lab coat.

Practical Advice

If your GTX 970 setup already works, the best mindset is:

stability first, upgrades second

Use a branch-based workflow:

  • one stable branch

  • one test branch per experiment

  • one rollback point

  • one frozen requirements file

That turns “I hope it still works” into “I know how to recover”.

Conclusion

Yes, ComfyUI still works on the GTX 970.

Not with every shiny optimization, not with infinite VRAM fantasies, and not by updating everything blindly — but with the right stack and a careful process, it is absolutely possible to keep this hardware productive.

If you are on older NVIDIA hardware, do not assume you are locked out.
Test methodically, keep rollback points, and treat frontend updates separately from core updates.

That made all the difference here.

1 Like