github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 00:51:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #9767] How to input image to gemma3 #52897

New Issue
Closed
opened 2026-04-29 01:18:19 -05:00 by GiteaMirror · 6 comments
GiteaMirror commented 2026-04-29 01:18:19 -05:00
Owner
Copy Link

Originally created by @wrx33 on GitHub (Mar 14, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/9767

I have run gemma3 by "ollama run gemma3", and it can generate output when I input text.
I run it in the command line. How to input image to gemma3?

Originally created by @wrx33 on GitHub (Mar 14, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/9767 I have run gemma3 by "ollama run gemma3", and it can generate output when I input text. I run it in the command line. How to input image to gemma3?
GiteaMirror commented 2026-04-29 01:18:22 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Mar 14, 2025):

Add the path to the image.

$ ollama run gemma3
>>> describe this image: ./puppy.jpg
Added image './puppy.jpg'
Here’s a description of the image:

**Overall Impression:**

The image shows a very cute, fluffy, white puppy sitting on a stone step. It’s a heartwarming and adorable photo.
<!-- gh-comment-id:2724542775 --> @rick-github commented on GitHub (Mar 14, 2025): Add the path to the image. ```console $ ollama run gemma3 >>> describe this image: ./puppy.jpg Added image './puppy.jpg' Here’s a description of the image: **Overall Impression:** The image shows a very cute, fluffy, white puppy sitting on a stone step. It’s a heartwarming and adorable photo. ```
GiteaMirror commented 2026-04-29 01:18:22 -05:00
Author
Owner
Copy Link

@MichaelCombs28 commented on GitHub (Mar 14, 2025):

I've run into issues as well where the server responds with a 500.

Turns out the issue for me was GPU memory, I'm using gemma3:27b.

ggml_backend_cuda_buffer_type_alloc_buffer: allocating 53273.01 MiB on device 0: cudaMalloc failed: out of memory
Mar 14 09:42:36 my-pc ollama[482610]: ggml_gallocr_reserve_n: failed to allocate CUDA0 buffer of size 55860802560
Mar 14 09:42:36 my-pc ollama[482610]: SIGSEGV: segmentation violation
Mar 14 09:42:36 my-pc ollama[482610]: PC=0x5e7acac62fd0 m=60 sigcode=1 addr=0x58
Mar 14 09:42:36 my-pc ollama[482610]: signal arrived during cgo execution

I fetched these logs with journalctl -u ollama

It might be ideal to return the actual error message, I'll open a PR when I have time later today.

<!-- gh-comment-id:2725222728 --> @MichaelCombs28 commented on GitHub (Mar 14, 2025): I've run into issues as well where the server responds with a 500. Turns out the issue for me was GPU memory, I'm using `gemma3:27b`. ``` ggml_backend_cuda_buffer_type_alloc_buffer: allocating 53273.01 MiB on device 0: cudaMalloc failed: out of memory Mar 14 09:42:36 my-pc ollama[482610]: ggml_gallocr_reserve_n: failed to allocate CUDA0 buffer of size 55860802560 Mar 14 09:42:36 my-pc ollama[482610]: SIGSEGV: segmentation violation Mar 14 09:42:36 my-pc ollama[482610]: PC=0x5e7acac62fd0 m=60 sigcode=1 addr=0x58 Mar 14 09:42:36 my-pc ollama[482610]: signal arrived during cgo execution ``` I fetched these logs with `journalctl -u ollama` It might be ideal to return the actual error message, I'll open a PR when I have time later today.
GiteaMirror commented 2026-04-29 01:18:23 -05:00
Author
Owner
Copy Link

@pdevine commented on GitHub (Mar 14, 2025):

I'm going to go ahead and close this as answered (thanks @rick-github !).

@MichaelCombs28 0.6.1 is just about to come out which has some memory fixes (gemma3:27b should work better).

<!-- gh-comment-id:2725564539 --> @pdevine commented on GitHub (Mar 14, 2025): I'm going to go ahead and close this as answered (thanks @rick-github !). @MichaelCombs28 `0.6.1` is just about to come out which has some memory fixes (gemma3:27b should work better).
GiteaMirror commented 2026-04-29 01:18:24 -05:00
Author
Owner
Copy Link

@wrx33 commented on GitHub (Mar 17, 2025):

Add the path to the image.

$ ollama run gemma3

describe this image: ./puppy.jpg
Added image './puppy.jpg'
Here’s a description of the image:

Overall Impression:

The image shows a very cute, fluffy, white puppy sitting on a stone step. It’s a heartwarming and adorable photo.

Thanks for helping. But when I input the image path, I got the response: "I can not access the local file system, so I can not read the image".

<!-- gh-comment-id:2727968668 --> @wrx33 commented on GitHub (Mar 17, 2025): > Add the path to the image. > > $ ollama run gemma3 > >>> describe this image: ./puppy.jpg > Added image './puppy.jpg' > Here’s a description of the image: > > **Overall Impression:** > > The image shows a very cute, fluffy, white puppy sitting on a stone step. It’s a heartwarming and adorable photo. Thanks for helping. But when I input the image path, I got the response: "I can not access the local file system, so I can not read the image".
GiteaMirror commented 2026-04-29 01:18:26 -05:00
Author
Owner
Copy Link

@pdevine commented on GitHub (Mar 17, 2025):

@wangrx33 can you cut and paste the prompt? Also, what OS are you using?

<!-- gh-comment-id:2728104511 --> @pdevine commented on GitHub (Mar 17, 2025): @wangrx33 can you cut and paste the prompt? Also, what OS are you using?
GiteaMirror commented 2026-04-29 01:18:30 -05:00
Author
Owner
Copy Link

@user202729 commented on GitHub (May 5, 2025):

This is described in the blog.

https://ollama.com/blog/vision-models

However, there's some magic to it. The user input is parsed against a certain regex to extract file names to be loaded.

7e5c8eee5c/cmd/interactive.go (L510-L538)

In particular:

  • The file extension must be .jpg, .png or .jpeg.
  • The file must exists. Otherwise it will fail silently without any error message.
  • If it's successful, you should see a line that is exactly Added image '<file path>' (with the <file path> replaced with your file path)
  • If the model does not accept image, it will again silently fail.

The model otherwise cannot access the local file system by itself. (I think.)

There's also report that jpg works better than png https://github.com/ollama/ollama/issues/7596 , although it's not clear why.

<!-- gh-comment-id:2851551204 --> @user202729 commented on GitHub (May 5, 2025): This is described in the blog. https://ollama.com/blog/vision-models However, there's some magic to it. The user input is parsed against a certain regex to extract file names to be loaded. https://github.com/ollama/ollama/blob/7e5c8eee5c65bcf7c0d46d5fe3b084fd70d36015/cmd/interactive.go#L510-L538 In particular: * The file extension must be `.jpg`, `.png` or `.jpeg`. * The file must exists. Otherwise it will fail silently without any error message. * If it's successful, you should see a line that is exactly `Added image '<file path>'` (with the `<file path>` replaced with your file path) * If the model does not accept image, it will again silently fail. The model otherwise cannot access the local file system by itself. (I think.) There's also report that jpg works better than png https://github.com/ollama/ollama/issues/7596 , although it's not clear why.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#52897
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?