Workflow Compiler error when binding Active Learning image name to input image filename

[pasko15]

Search before asking

• I have searched the Inference issues and found no similar bug report.

Bug

I am running a workflow for blueprint processing where preserving the original file name is important for traceability.

Each blueprint has a unique name / identifier, and when I upload images directly, the filenames are preserved correctly. However, when images are sent to my dataset through the Active Learning / Roboflow Dataset Upload block, Roboflow assigns its own UUID-based name instead. That makes the uploaded samples difficult to track back to the original blueprint source.

I tried to solve this by setting the Image Name field in the Dataset Upload block to use the input image name, but I get the following compiler error:

Workflow Compiler expected node: $inputs.image.name to be present in execution graph, but condition failed to be met. This is most likely the bug. Contact Roboflow team through github issues (https://github.com/roboflow/inference/issues) providing full context of the problem - including workflow definition you use.

In short:
I want the image uploaded by the Dataset Upload / Active Learning block to keep the same name as the original input image, instead of receiving a generated UUID.

Environment

• Inference: 1.0

The issue persists both on the serverless API and on batch processing.

Minimal Reproducible Example

Goal

Preserve the original filename of input images when they are uploaded to a Roboflow dataset through the Roboflow Dataset Upload / Active Learning block in a Workflow.

The dataset contains blueprint images with unique identifiers in their filenames, so preserving the original name is important for traceability.

Workflow

Simplified workflow structure:

Input (InferenceImage)
→ processing blocks
→ Roboflow Dataset Upload (Active Learning)

Attempted configuration

I attempted to set the Image Name field in the Dataset Upload block to: steps.name_extract.name

Additional

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!

[eskayML]

it looks like the issue is related to the dynamic binding of the image name in the dataset upload block. a possible fix could be to add a static binding for the image name in the workflow, or to modify the dataset upload block to preserve the original filename. should i open a PR for this?

[pasko15]

it looks like the issue is related to the dynamic binding of the image name in the dataset upload block. a possible fix could be to add a static binding for the image name in the workflow, or to modify the dataset upload block to preserve the original filename. should i open a PR for this?

Yes, I need the it to keep the original file name. I have tried both within the block and by running a seperate python script. I keep getting an error that this might be a bug and I should post the issue on Git.

And, yes, please do open a PR. Thanks in advance.

[furai-ai]

Hi! I've been digging into the compiler source.

The root cause appears to be that $inputs.image.name isn't registered as a valid selector in the execution graph — the compiler validates bindings against known output kinds, but .name isn't declared as a property of InferenceImage inputs.
I'm planning to trace this through the execution engine type definitions and submit a PR. Is there a preferred approach —
exposing .name as a first-class image input property, or handling it inside the Dataset Upload block manifest?

[PawelPeczek-Roboflow]

Hi :)

Sorry for realising so late that this issue was submitted. Great to see people figuring out different ways of using Workflows.

Let me help on that. It looks like name property apart from being linked would need to be actually delivered to the block executing upload. Maybe an easier way to solve this problem would be extending the definition of WorkflowImageData entity which is a representation of image in runtime to have additional attribute which can be specified by caller submitting data (something like image_id) which then could be used by any block using that image and from the caller perspective would require adding image_id to the dict specifying input.

[furai-ai]

Hi :) Thanks for the detailed response! The `image_id` approach makes sense — extending `WorkflowImageData` with a caller-supplied identifier is cleaner than trying to thread `name` through the binding system. Happy to help implement it if you'd like a PR. I've been digging into the inference codebase quite a bit lately and find the Workflows architecture genuinely interesting. I'm also currently building my own AI-powered project, so runtime data flow is something I think about a lot. Are there any open positions or contributor opportunities on the team? Dmitrii Mingaziev AI Engineer focused on memory, identity, and evolving AI systems. E-mail: ***@***.***

…

On 24 Mar 2026 at 01:06 +0700, Paweł Pęczek ***@***.***>, wrote: PawelPeczek-Roboflow left a comment (roboflow/inference#2113) Hi :) Sorry for realising so late that this issue was submitted. Great to see people figuring out different ways of using Workflows. Let me help on that. It looks like name property apart from being linked would need to be actually delivered to the block executing upload. Maybe an easier way to solve this problem would be extending the definition of WorkflowImageData entity which is a representation of image in runtime to have additional attribute which can be specified by caller submitting data (something like image_id) which then could be used by any block using that image and from the caller perspective would require adding image_id to the dict specifying input. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[PawelPeczek-Roboflow]

Open roles listed here: https://roboflow.com/careers
Feel free to apply if anything looks interesting

[furai-ai]

Hi Paweł, Thanks for sharing — I’ll check out the roles and apply. Really enjoyed digging into the issue and working with your system, would love to contribute. Best, Dmitrii Dmitrii Mingaziev AI Engineer focused on memory, identity, and evolving AI systems. + 84 77 241 0573 | WhatsApp

…

On 24 Mar 2026 at 17:25 +0700, Paweł Pęczek ***@***.***>, wrote: PawelPeczek-Roboflow left a comment (roboflow/inference#2113) Open roles listed here: https://roboflow.com/careers Feel free to apply if anything looks interesting — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[eskayML]

Beat me to the PR @furai-ai 💪😂

Great work