Seedance 2.5 Reference-Image Connections Disappear When Switching Workflow Tabs and Saving/Reloading
- Dominant language
- Python
- Stars
- 133k
- Forks
- 15.7k
- Avg merge
- 1d 7h
- Merged PRs (30d)
- 158
Description
### Custom Node Testing
- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)
### Expected Behavior
What happened?
I am using the official ByteDance Seedance Reference to Video node with the model set to Seedance 2.5.
Seedance 2.5 supports a large number of reference images, so I am using the node's dynamically expanding image_1, image_2, image_3, etc. inputs for character and environment reference images.
The workflow works correctly while the connections are present. I have successfully executed Seedance generations with approximately 17–20 reference images connected.
However, the reference-image connections are not persistent.
There are two reproducible failures.
Problem 1 — Switching workflow tabs removes Seedance reference connections
Create a Seedance 2.5 Reference to Video workflow.
Expand the Seedance node to many reference-image inputs.
Connect approximately 17–20 separate Load Image nodes to image_1, image_2, image_3, etc.
Run the workflow successfully.
Switch to another ComfyUI workflow tab.
Return to the Seedance workflow tab.
Actual result: Many of the Seedance reference-image connections have disappeared.
The Load Image nodes are still present, but their connections to the Seedance node are gone.
Problem 2 — Saving and reopening the workflow also loses connections
Connect the Seedance reference images.
Confirm the graph is working.
Save/export the workflow as JSON.
Reload the saved workflow.
Actual result: Many of the reference-image connections are missing after reload. The Seedance node may also return with fewer dynamically generated image_N inputs than it had when the workflow was saved.
JSON evidence
I inspected an affected workflow JSON after saving.
The Seedance node was serialized with image_1 through image_17, but multiple Load Image nodes that had been connected in the UI were saved with empty output link arrays:
"links": []
I also found stale/duplicate link IDs referencing Seedance input-slot indexes that no longer corresponded to valid image inputs in the serialized Seedance node.
In other words, this does not appear to be only a visual rendering problem in the graph. Some of the connections are actually missing or inconsistent in the saved workflow data.
Why this is especially serious for Seedance 2.5
Seedance 2.5 is specifically designed to work with large reference sets.
Having 15–30 reference-image inputs is therefore a normal use case for this node, not an unusual edge case.
Manually reconnecting potentially dozens of character, wardrobe, expression, prop, and environment references every time the user switches workflow tabs or reloads the workflow makes the multi-reference feature effectively unusable for production work.
What I expected
All dynamically created Seedance reference inputs and their connections should survive:
switching workflow tabs
saving the workflow
reopening the workflow
loading a workflow from JSON
The graph after reload should be identical to the graph that was saved.
Additional observations
The Seedance API itself appears to be functioning correctly.
When the reference connections are present, the workflow executes and generates successfully.
The failure happens specifically when ComfyUI changes workflow tabs or serializes/deserializes the workflow state.
This suggests the problem may involve the frontend handling of dynamic/autogrow inputs and their link serialization.
Environment
Windows
Chrome / Chromium
Local ComfyUI
Seedance model: Seedance 2.5
Node class in affected workflow: ByteDance2ReferenceNode
Approximately 17–20 reference images in the reproduced workflow
Audio reference also connected
Workflow successfully generates before the connections disappear
Related existing issues
This Seedance problem appears related to several existing reports, but I am opening a Seedance-specific issue because the new Seedance 2.5 workflow depends heavily on large dynamic reference sets and the bug directly prevents that advertised workflow from being used reliably.
Related issues:
ComfyUI_frontend #10417 — Flux Graph Inputs are being disconnected when changing tabs
ComfyUI_frontend #11176 — Autogrow input connections are dropped
ComfyUI #13045 — Missing links when saving
Attachments available
I can provide:
the original affected workflow JSON
screenshots showing the connected graph
screenshots showing the same graph after connections disappear
a screen recording of the tab-switch reproduction
examples of the inconsistent link data from the saved JSON
Please investigate this specifically with the Seedance 2.5 multi-reference workflow, because large dynamic reference sets appear to trigger or expose the problem very clearly.
[BandSeedance_ScreemTest_001.json](https://github.com/user-attachments/files/31116433/BandSeedance_ScreemTest_001.json)
### Actual Behavior
What happened?
I am using the official ByteDance Seedance Reference to Video node with the model set to Seedance 2.5.
Seedance 2.5 supports a large number of reference images, so I am using the node's dynamically expanding image_1, image_2, image_3, etc. inputs for character and environment reference images.
The workflow works correctly while the connections are present. I have successfully executed Seedance generations with approximately 17–20 reference images connected.
However, the reference-image connections are not persistent.
There are two reproducible failures.
Problem 1 — Switching workflow tabs removes Seedance reference connections
Create a Seedance 2.5 Reference to Video workflow.
Expand the Seedance node to many reference-image inputs.
Connect approximately 17–20 separate Load Image nodes to image_1, image_2, image_3, etc.
Run the workflow successfully.
Switch to another ComfyUI workflow tab.
Return to the Seedance workflow tab.
Actual result: Many of the Seedance reference-image connections have disappeared.
The Load Image nodes are still present, but their connections to the Seedance node are gone.
Problem 2 — Saving and reopening the workflow also loses connections
Connect the Seedance reference images.
Confirm the graph is working.
Save/export the workflow as JSON.
Reload the saved workflow.
Actual result: Many of the reference-image connections are missing after reload. The Seedance node may also return with fewer dynamically generated image_N inputs than it had when the workflow was saved.
JSON evidence
I inspected an affected workflow JSON after saving.
The Seedance node was serialized with image_1 through image_17, but multiple Load Image nodes that had been connected in the UI were saved with empty output link arrays:
"links": []
I also found stale/duplicate link IDs referencing Seedance input-slot indexes that no longer corresponded to valid image inputs in the serialized Seedance node.
In other words, this does not appear to be only a visual rendering problem in the graph. Some of the connections are actually missing or inconsistent in the saved workflow data.
Why this is especially serious for Seedance 2.5
Seedance 2.5 is specifically designed to work with large reference sets.
Having 15–30 reference-image inputs is therefore a normal use case for this node, not an unusual edge case.
Manually reconnecting potentially dozens of character, wardrobe, expression, prop, and environment references every time the user switches workflow tabs or reloads the workflow makes the multi-reference feature effectively unusable for production work.
What I expected
All dynamically created Seedance reference inputs and their connections should survive:
switching workflow tabs
saving the workflow
reopening the workflow
loading a workflow from JSON
The graph after reload should be identical to the graph that was saved.
Additional observations
The Seedance API itself appears to be functioning correctly.
When the reference connections are present, the workflow executes and generates successfully.
The failure happens specifically when ComfyUI changes workflow tabs or serializes/deserializes the workflow state.
This suggests the problem may involve the frontend handling of dynamic/autogrow inputs and their link serialization.
Environment
Windows
Chrome / Chromium
Local ComfyUI
Seedance model: Seedance 2.5
Node class in affected workflow: ByteDance2ReferenceNode
Approximately 17–20 reference images in the reproduced workflow
Audio reference also connected
Workflow successfully generates before the connections disappear
Related existing issues
This Seedance problem appears related to several existing reports, but I am opening a Seedance-specific issue because the new Seedance 2.5 workflow depends heavily on large dynamic reference sets and the bug directly prevents that advertised workflow from being used reliably.
Related issues:
ComfyUI_frontend #10417 — Flux Graph Inputs are being disconnected when changing tabs
ComfyUI_frontend #11176 — Autogrow input connections are dropped
ComfyUI #13045 — Missing links when saving
Attachments available
I can provide:
the original affected workflow JSON
screenshots showing the connected graph
screenshots showing the same graph after connections disappear
a screen recording of the tab-switch reproduction
examples of the inconsistent link data from the saved JSON
Please investigate this specifically with the Seedance 2.5 multi-reference workflow, because large dynamic reference sets appear to trigger or expose the problem very clearly.
### Steps to Reproduce
...
### Debug Logs
```powershell
[ERROR] Exception in callback _ProactorBasePipeTransport._call_connection_lost()
handle:
Traceback (most recent call last):
File "asyncio\events.py", line 89, in _run
File "asyncio\proactor_events.py", line 165, in _call_connection_lost
ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host
```
### Other
this bug is well documented since march but now that it affeects seedance 2.5 100% you should fix it fast
Contributor guide
Research direction
Start with the attached BandSeedance_ScreemTest_001.json and reproduce the tab-switch and save/reload failures using ByteDance2ReferenceNode with many image_N inputs. Compare the graph and serialized links before and after each action; done means all dynamic inputs and connections survive tab switching, saving, reopening, and JSON loading.
Written by the indexing model from the issue text.
Assessment
- Domain
- frontend
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 48/100