Thanks for reporting this issue!

I've checked the implementation of Qwen 2.1 Sheet Encode. This node calls ComfyUI's built-in TextEncodeQwenImage21 and does not directly depend on TensorFlow.

The error:

AttributeError: module 'tensorflow' has no attribute 'Tensor'

may be caused by an incomplete or broken TensorFlow installation, or by another Python module conflicting with TensorFlow.

Could you please try the following checks?

1. Check the installed packages

Open PowerShell in your ComfyUI portable installation folder and run:

```powershell

.\python_embeded\python.exe -m pip show tensorflow tensorflow-intel tensorflow-cpu einops

```

If you use a virtual environment or another ComfyUI installation method, please use the Python executable that ComfyUI actually runs.

2. Check whether TensorFlow is working correctly

Run:

```powershell

.\python_embeded\python.exe -c "import sys; import tensorflow as tf; print('Python:', sys.executable); print('Version:', getattr(tf, '__version__', None)); print('Path:', getattr(tf, '__file__', None)); print('Tensor:', hasattr(tf, 'Tensor')); print('Variable:', hasattr(tf, 'Variable'))"

```

How to interpret the results:

* Tensor: True / Variable: True — The main TensorFlow types are available. This does not guarantee that the entire installation is healthy, but the reported error needs further investigation.

* Tensor: False or Variable: False — The imported TensorFlow module is missing expected attributes. This suggests an incomplete installation, a corrupted package, or a conflicting module.

* Version: None or Path: None — Python may be importing an incomplete package or a namespace package instead of a normal TensorFlow installation. This requires further investigation.

* Path points to an unexpected folder — A local file or another package may be shadowing the actual TensorFlow installation.

* ModuleNotFoundError: No module named 'tensorflow' — TensorFlow is not available in the Python environment being checked. Please verify that this is the same Python environment used by ComfyUI.

These results can help identify the problem, but they do not prove that TensorFlow is corrupted.

3. Test the standard ComfyUI encoder

Please also try replacing Qwen 2.1 Sheet Encode with ComfyUI's built-in Text Encode Qwen Image 2.1 node, using the same CLIP, VAE, reference image, and prompt.

If the same error occurs with the standard encoder, the issue is more likely related to the ComfyUI environment, its dependencies, or the Core encoding process rather than this custom node.

4. Please share the results

It would be very helpful if you could provide:

* The output of both commands above.

* Your ComfyUI version.

* Whether the standard Text Encode Qwen Image 2.1 node produces the same error.

The *complete traceback from the ComfyUI console**, especially the lines showing which Python file raised the exception.

Similar errors have occurred when einops encounters an incomplete or conflicting TensorFlow module.

However, I would prefer to confirm the actual traceback before recommending any package changes.

Please don't uninstall or reinstall TensorFlow yet, as other custom nodes may depend on it.

Thanks for your help! Once we have these results, we should be able to narrow down the cause and determine the safest fix.

@Jojo0Gojo Thanks for checking! The “tensorflow-cpu not found” warning alone doesn’t mean TensorFlow is broken or that you need to install the CPU package. The Qwen Character Sheet node itself doesn’t require TensorFlow.

When you say “ComfyUI easy,” do you mean ComfyUI Easy Install or the Easy Use custom nodes? A link to the version you use would help.

Could you also share the full traceback from the ComfyUI console when the Qwen workflow fails, along with the exact command you ran for the package check? The check needs to use the same Python environment as ComfyUI.

It’s helpful to know the H3 version works, but the Qwen workflow can use a different processing path. Let’s check the log first before installing or removing any packages.

@Jojo0Gojo Thanks for sharing the log! Could you try running the Qwen workflow once without any reference images connected and let me know whether it works? No package installation or removal is needed for this test.

@Jojo0Gojo Thank you for the detailed traceback, and thanks for your patience while I investigated. I built a separate CPU test environment based on Easy Install’s installation steps, using ComfyUI 0.39.2.

I don’t normally use the official ComfyUI Portable distribution and didn’t have an existing test installation available, so I haven’t tested the official Portable package itself. Setting up the Easy Install-based environment and investigating the error took some additional time.

Your exact environment is still difficult to reproduce. Package versions, custom nodes, and previous updates can all affect the shared Python environment, and I can’t account for every possible combination.

In controlled CPU tests, importing an incomplete TensorFlow module caused the same error in einops, even while processing PyTorch tensors. The normal test configuration did not produce the error. This identifies a possible cause, but the exact cause in your installation remains unconfirmed.

The traceback follows this path:

Qwen Character Sheet Encode → ComfyUI’s standard Qwen encoder → VAE processing → einops

Continuum is not directly involved in this reported execution path.

As a first step, please try this temporary workaround, which does not install, remove, or edit any packages:

Stop the affected ComfyUI instance.

Open a new PowerShell window in the installation folder containing python_embeded and ComfyUI.

Run:

$env:USE_TF = "0"
.\python_embeded\python.exe -s .\ComfyUI\main.py --windows-standalone-build

Append any additional launch options you normally need. Keep your character/reference images connected, then retry the same workflow.

This disables Transformers’ optional TensorFlow integration for that launch. It prevented the error in our controlled test with Transformers 4.57.6, although it may not help if another custom node imports TensorFlow directly.

To revert: stop ComfyUI, close this PowerShell window, and start ComfyUI using your usual launcher.

I’ve investigated this as far as I reasonably can with a separate test setup, but my ability to troubleshoot individual installations and dependency changes introduced by other custom nodes is limited. If the workaround doesn’t help, the next step would be to check TensorFlow’s import location and the installed nodes in your environment, with help from Easy Install’s troubleshooting resources where appropriate.

For now, please avoid removing packages or editing einops until the source of the problem is clearer, as those changes could affect other nodes.

@Jojo0Gojo Thanks for the update—glad updating resolved the error! I’ve recorded the Part prompts selector staying on “Head” in the H3 version’s Issue #1 for investigation. Thanks for reporting it!

Sorry for the confusing version numbers. I hadn’t kept the workflow names aligned with the package version. The workflow labeled V1.3 was the updated workflow, but its name was outdated.

I’ve now aligned the workflow names with v1.5.4. This is a naming correction only—no wiring or settings have changed, so there’s no need to download it again if you already have that workflow. The internal node ID still contains V1.3 for compatibility with existing workflows.

Thanks for asking! The current node doesn’t have a dedicated thigh-up option. “Left/Right 3/4” refers to portrait angles, not body framing. You can crop a full-body result for that framing. Could you point me to the specific preview image you mean?

@HarlequinArt ukr8b3g201

OP

24 minutes ago

No cropping or compositing was used. The original reference showed the character from around the thighs up, which may have influenced the result. There’s no dedicated thigh-up option in the current node. For the three full-body views, select Front, Left side, and Back.