image editor crashes

[Lluviosa]

got prompt
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Requested to load WanVAE
0 models unloaded.
loaded partially 128.0 127.9998779296875 0
Using scaled fp8: fp8 matrix mult: False, scale input: False
Requested to load QwenImageTEModel_
loaded completely 9.5367431640625e+25 7909.737449645996 True
CLIP/text encoder model load device: cpu, offload device: cpu, current: cpu, dtype: torch.float16
0 models unloaded.
loaded partially 128.0 127.9998779296875 0
0 models unloaded.
loaded partially 128.0 127.9998779296875 0
Crashes

[dosubot]

Hi @Lluviosa! I'm Dosu and I’m helping the comfy-cli team.

Crashes during model loading in ComfyUI via comfy-cli can often be traced to compatibility issues between PyTorch and XFormers, especially on Windows. If XFormers is installed after PyTorch, it may downgrade PyTorch to an incompatible version, causing runtime errors or missing modules like qwen_vl_utils and gguf—this can lead to crashes when loading models or assigning devices and dtypes. Workarounds include manually installing compatible versions, using pip install xformers --no-deps, or switching to Python 3.9 for better compatibility. Forcing incompatible versions can also cause runtime failures, so it's important to confirm library compatibility before installation. See more details here.

If you can share the specific error message or stack trace, your comfy-cli and ComfyUI versions, and your operating system, I can help narrow down the cause and suggest more targeted troubleshooting steps.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[Lluviosa]

@dosu it only happens to image editor music model works perfectly fine: Checkpoint files will always be loaded safely.
Total VRAM 8192 MB, total RAM 24513 MB
pytorch version: 2.6.0+cu124
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 3060 Ti : cudaMallocAsync
Using pytorch attention
Python version: 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
ComfyUI version: 0.3.66
ComfyUI frontend version: 1.28.7
[Prompt Server] web root: H:\Comfy\ComfyUI\venv\lib\site-packages\comfyui_frontend_package\static

Import times for custom nodes:
0.0 seconds: H:\Comfy\ComfyUI\custom_nodes\websocket_image_save.py

Context impl SQLiteImpl.
Will assume non-transactional DDL.
No target revision found.
Starting server

To see the GUI go to: http://127.0.0.1:8188
got prompt
model weight dtype torch.bfloat16, manual cast: None
model_type FLOW
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load AceT5Model
loaded completely 5706.8 537.60791015625 True
Requested to load ACEStep
loaded partially 5654.67480545044 5654.6748046875 0
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 50/50 [00:16<00:00, 2.99it/s]
Requested to load MusicDCAE
loaded completely 9.5367431640625e+25 495.50055503845215 True
Prompt executed in 28.45 seconds

[dosubot]

The crash only affecting the image editor (while the music model works fine) suggests a workflow-specific issue, not a general environment or installation problem. There’s a known bug in comfy-cli where workflows using loop mechanisms or certain custom nodes can cause crashes due to composite or unexpected node IDs (like '406.0.0.428'), which aren’t present as top-level keys in the workflow JSON. This results in a KeyError and a crash, even if everything runs fine in the ComfyUI web interface. The only workaround is to avoid such node configurations or preprocess the workflow JSON to ensure all referenced node IDs are present, as comfy-cli currently lacks robust error handling for this scenario. If your image editor workflow uses loops or custom nodes, this could be the root cause. More details here.

If you can share the exact error message or stack trace, I can help confirm if this is the issue or suggest more targeted troubleshooting steps.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[Lluviosa]

@dosu ComfyUI crashes or produces nonsensical model-load values when attempting to load QwenImageTEModel_.

Observed console output:

Requested to load QwenImageTEModel_
loaded completely 9.5367431640625e+25 7909.737449645996 True
CLIP/text encoder model load device: cpu, offload device: cpu, current: cpu, dtype: torch.float16

Environment:

GPU: NVIDIA RTX 3060 Ti

CUDA: 12.4 (driver 13.0 compatible)

PyTorch: 2.9.0 + cu124

ComfyUI (latest main branch, launched via python main.py)

Windows 10/11 x64

Model: QwenImageTEModel_ (exact source uncertain)

Expected behavior:
Model should load normally and print a reasonable tensor size (e.g. 512 MB or similar).

Actual behavior:

Reports an impossible tensor/memory size (9.5367431640625e+25)

Suggests broken model metadata or mismatched loader code

May prevent the workflow from executing further

Preliminary diagnosis:

The QwenImageTEModel_ checkpoint is likely corrupted, incomplete, or not compatible with the loader used by ComfyUI.

The node that calls QwenImageTEModel_ may be using an outdated or experimental interface expecting a different model schema.

The “loaded completely … True” message comes from a failed tensor deserialization (NaN or overflow in size field).

Steps to reproduce:

Launch ComfyUI (python main.py --gpu-only)

Load a workflow or node referencing QwenImageTEModel_.

Observe console output (see above).

Suggested verification for maintainer:

Add checksum or schema validation before reporting “loaded completely.”

Check that the model file’s header (e.g., .safetensors or .bin) matches the expected Qwen image-text encoder architecture.

Confirm that the custom node invoking this model aligns with the current ComfyUI API for model loading.

Temporary workaround for users:

Remove or rename the Qwen node folder from ComfyUI/custom_nodes/.

Restart ComfyUI.

Re-download or update the Qwen model and node once an updated loader is available.

[Lluviosa]

I need that access to comfy cloud beta gang 💀

[dosubot]

The crash and nonsensical tensor size when loading QwenImageTEModel_ almost certainly points to a corrupted, incomplete, or incompatible checkpoint file, or a mismatch between the custom node's loader code and the expected model schema in ComfyUI. This kind of error—reporting an impossible tensor/memory size—typically comes from a failed tensor deserialization, such as a NaN or overflow in the size field.

For maintainers, the best verification steps are: add checksum or schema validation before reporting “loaded completely,” check that the model file’s header matches the expected Qwen image-text encoder architecture, and confirm that the custom node invoking this model aligns with the current ComfyUI API for model loading.

For users, the most reliable workaround is to remove or rename the Qwen node folder from ComfyUI/custom_nodes/, restart ComfyUI, and re-download or update the Qwen model and node once a compatible loader is available.

No comfy-cli code or recent changes directly address this issue; model loading and tensor deserialization are handled by ComfyUI itself, not comfy-cli. Existing comfy-cli issues and PRs do not discuss this crash or nonsensical tensor size values.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[Lluviosa]

@dosu what do i do then, step by step i just wanna turn my roblox avatar into real life

[dosubot]

Here’s a practical step-by-step guide to turn your Roblox avatar into a real-life image using ComfyUI and comfy-cli:

1. Export your Roblox avatar: Save a clear screenshot or render of your avatar as a PNG or JPG.

2. Choose a compatible image-to-image model: Avoid models/nodes with loader issues (like QwenImageTEModel_). Use a well-supported model such as Stable Diffusion, RealisticVision, or similar. Download the model checkpoint and place it in your ComfyUI models folder.

3. Create a workflow in ComfyUI:

   • Use the ComfyUI graphical interface to build an image-to-image workflow: start with an image input node, connect it to your chosen model, and add a prompt node describing the desired transformation (e.g., “make this avatar look like a realistic human photo”).
   • Save/export the workflow as a JSON file.

4. Run the workflow with comfy-cli:

   • Make sure ComfyUI is running.
   • Use comfy-cli to execute your workflow:

     comfy-cli run path/to/your_workflow.json
     

   • Monitor the output for errors. If you see model loading issues, double-check model compatibility and file integrity.

5. Troubleshooting:

   • If you hit crashes or nonsensical tensor sizes, remove problematic custom nodes (like Qwen) from custom_nodes/, restart ComfyUI, and try again.
   • Stick to mainstream models and avoid experimental or poorly documented nodes.

There’s no official Roblox-to-real-life workflow, so you’ll need to experiment with prompts and models. If you want a template, start with a basic image-to-image workflow and tweak the prompt for realism. If you need help with prompt engineering or model selection, let me know what you have and I can suggest specifics!

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}