Model does not support image input
EDIT: I finally figured out a solution. for now. I copied the mmproj-F32 gguf file from the main or working gemma 3 model into model folder of this one, and it seems to work so far for allowing image input.
EDIT2: okay and now i realized model page does in fact come with mmproj-gemma-3-27b-it-abliterated-f32.gguf which is needed for vision. i replaced other one i borrowed with this one, and working in lm studio so far without issue. so i guess LM Studio doesn't know how to retrieve or use the additional file on it's own, apart from the main models or something. good to know. Sorry i had to post this it seems to finally figure it out
Question. i been curious about what community does with Gemma 3 while still maintaining vision support. I use LM Studio for experimenting with new models since it's easy to use. This particular model is first i seen with the Vision Enabled tag indicating it can process image inputs. but when i download then load the model, there is no button to upload the image like the standard gemm-3-27b model, and if i drag an image it says Model does not support image input despite it clearly saying on model browser this model supports image inputs. Does anyone know what i am missing? I have had this same issue with community models of other models with vision support too, it always seems like only the main model works not the forks of that model.
How do I merge the vision into gemma3? I am using ollama so having these files separate is out of the question and vision does not work when I do 2 FROM statements in Modelfile.
llamacpp throws me error
.\llama-gguf-split.exe --merge ..\mlabonne_gemma-3-27b-it-abliterated-Q4_K_L.gguf ..\mmproj-mlabonne_gemma-3-27b-it-abliterated-f16.gguf gemma3a.guff
error: bad arguments
I don't think you can merge them like that, never seen anyone try though
Do you know how to do that or a some other way to get the file merge?
Do you know how to do that or a some other way to get the file merge?
You can't, you need to use llama.cpp I believe and in the shell command for starting inference, include both file paths (one for the fp16 main gemma3 model, one for the gemma3 vision mmproj model). Here's an example of how I use it that functions perfectly:
./build/bin/llama-gemma3-cli \
-m {large_main_model_path} \
--mmproj {mmproj_image_model_path} \
--image '{image_path_escaped}' \
-p "[Prompt]"