[GH-ISSUE #14306] Image editing with Flux2 Klein models has not worked since around 0.15.6 #35070

Open
opened 2026-04-22 19:16:18 -05:00 by GiteaMirror · 1 comment
Owner

Originally created by @rjdg14 on GitHub (Feb 18, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14306

What is the issue?

On Ollama 0.15.4 (the last version I've tested that this works with) running on MacOS, giving the 4B or 9B Flux2 Klein models, for example, a .jpg image of a smiley face or a bridge over a river, and posting "Make the face frown filepath/emoji.jpg" or "Remove the bridge filepath/river.jpg" will modify the image. On 0.15.6 and later it still gives the "added image" response when you run it, but the output image has no resemblance to the image you requested it to edit, with it simply creating a brand new image rather than editing the existing one that the model was given.

Relevant log output


OS

macOS Sequoia 15.7.3

GPU

Apple M3

CPU

Apple M3

Ollama version

0.16.2; this issue has been in Ollama since about 0.15.6

Originally created by @rjdg14 on GitHub (Feb 18, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14306 ### What is the issue? On Ollama 0.15.4 (the last version I've tested that this works with) running on MacOS, giving the 4B or 9B Flux2 Klein models, for example, a .jpg image of a smiley face or a bridge over a river, and posting "Make the face frown filepath/emoji.jpg" or "Remove the bridge filepath/river.jpg" will modify the image. On 0.15.6 and later it still gives the "added image" response when you run it, but the output image has no resemblance to the image you requested it to edit, with it simply creating a brand new image rather than editing the existing one that the model was given. ### Relevant log output ```shell ``` ### OS macOS Sequoia 15.7.3 ### GPU Apple M3 ### CPU Apple M3 ### Ollama version 0.16.2; this issue has been in Ollama since about 0.15.6
GiteaMirror added the bug label 2026-04-22 19:16:18 -05:00
Author
Owner

@aotol commented on GitHub (Apr 4, 2026):

I can confirm this still happens on Ollama 0.20.2.

Environment:

  • Ollama: 0.20.2
  • macOS: Tahoe 26.3.1, build 25D2128

Command:
ollama run x/flux2-klein "./input.png turn this into anime style, keep facial features similar"

Expected:
The model should edit the provided image and preserve the original face / identity structure.

Actual:
Ollama accepts the image input, but the output ignores the source image and generates an unrelated new image instead.

So this issue is still reproducible on 0.20.2.

<!-- gh-comment-id:4187057962 --> @aotol commented on GitHub (Apr 4, 2026): I can confirm this still happens on Ollama 0.20.2. Environment: - Ollama: 0.20.2 - macOS: Tahoe 26.3.1, build 25D2128 Command: ollama run x/flux2-klein "./input.png turn this into anime style, keep facial features similar" Expected: The model should edit the provided image and preserve the original face / identity structure. Actual: Ollama accepts the image input, but the output ignores the source image and generates an unrelated new image instead. So this issue is still reproducible on 0.20.2.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#35070