Use GPT Image 2.5 in Dify: check the plugin and model
Find the Dify OpenAI image tool, check Flare and Sunburst support, and stop a workflow from silently keeping GPT Image 2 as its selected default model.
Ofox 
To use GPT Image 2.5 in Dify, check the installed OpenAI Tools plugin and explicitly select Flare or Sunburst in the image tool’s model parameter. The updated source keeps gpt-image-2 as the default. Upgrading a plugin and choosing the new model are separate steps.
Check availability in your installation: the official PR #3860 was merged on September 9, 2026, and its manifest declares version 0.2.0. We have not independently confirmed that this version is distributed through every Marketplace installation. These are source-verified setup checks, not a claim that a full Dify deployment and paid generation were tested here.
Identify the correct plugin and tool
This change belongs to the OpenAI Tools plugin. It is not merely a new language-model entry in the OpenAI model-provider settings. Inspect the image tool node in the workflow you actually run.
The official tool source retains the tool identifiers:
| Task | Tool ID | What to inspect |
|---|---|---|
| Generate from text | gpt_image_2_generate | Prompt, Model, size and quality |
| Edit a reference | gpt_image_2_edit | Reference input, prompt and Model |
In the updated source, the English labels are GPT Image 2 / 2.5 Generate and GPT Image 2 / 2.5 Edit. The 2 in those tool IDs does not by itself mean the tool cannot use 2.5. The selected model is the important field. Conversely, seeing a new model in a different node does not prove that this image tool uses it.
Check the installed version before copying settings
Open the installed plugin’s details and inspect its version. Then open the generation or editing tool’s parameter form and look for these model choices:
gpt-image-2gpt-image-2.5-flaregpt-image-2.5-sunburst
If your installation does not offer the new choices, check whether an official update is available for that plugin. If it is not, record the installed version and the missing field. Do not describe a source merge as proof that your hosted workspace has already updated.
For a self-hosted installation, follow the plugin’s current packaging and dependency requirements rather than changing only the version label. The merged change updates dependencies as well as tool parameters. A version string edited by hand does not add those capabilities.
Select Flare or Sunburst explicitly
Choose the exact new model in the tool node; do not leave the model at its default. Use Flare as a starting point for fast iteration, or Sunburst when detailed editing is the main task. These are intended uses, not measured results for your workflow; see Flare vs Sunburst.
For the first text-to-image check, keep the workflow small: one prompt input, one image generation tool and an output that exposes the returned image. Start with 1024x1024 and a simple subject. That makes it easier to distinguish configuration failures from complex editing requests.
The updated source accepts xhigh and max for Image 2.5, along with the other documented quality values. It validates parameters before sending the request. A validation error in the tool therefore does not necessarily mean the OpenAI service received or rejected a request.
Verify the model from output, not the node’s name
When the upstream response includes usage, the updated implementation emits JSON containing the selected model alongside that usage. Keep it with the run parameters when comparing costs or reporting a failure. If no usage JSON is returned, inspect the saved node parameters and a sanitized request trace; missing usage alone does not prove the wrong model ran.
| Result | Next action |
|---|---|
Usage still identifies gpt-image-2 | Recheck the node’s explicit model value and whether you ran the saved workflow version |
| Model choices are missing | Check the installed Tools plugin version and available update |
| Custom dimensions are rejected locally | Use the size-rule checker |
| A returned PNG looks opaque | Check the file with the alpha inspection guide |
| The upstream service rejects access | Record the request details and follow the API access checks |
Do not infer the model from the image’s appearance. Also check the actual image output, rather than assuming that a successful workflow step produced a downloadable file.
Add editing only after the simple path works
For editing, connect the reference file required by gpt_image_2_edit, then state what should change and what should remain. Verify the file reaches the tool as an image input rather than merely inserting its name into the prompt.
Keep the original file, output file, selected model and usage together. That gives you a reproducible record if the product label changes or the background is not removed. For transparent output, inspect both the requested format and the actual alpha values; a .png suffix is not enough.
Start with a single deliberate run, then evaluate its output and cost before putting the workflow into a loop. The practical success criteria are the right tool, an explicitly selected 2.5 model, an accepted request and a usable saved image.
Frequently Asked Questions
- Does Dify support GPT Image 2.5?
- The official OpenAI Tools source added Flare and Sunburst in merged PR 3860 with plugin version 0.2.0. Marketplace distribution was not independently confirmed on September 9. Check the version and model choices in your installed tool.
- Why does my Dify workflow still call GPT Image 2?
- The updated tools retain gpt-image-2 as the default model. An update alone does not select Flare or Sunburst; inspect the Model parameter in the image tool node and, when usage is returned, the selected model reported alongside it.
- Should I select an LLM node to generate the image?
- This guide concerns the OpenAI Tools image generation and editing tools. Changing a text-model provider or an LLM node does not establish which model the separate image tool uses.
