kleinnner/camus-19-vision-inference-pipeline

We run vision-language inference on CPU — 2.5GB projector, 0.5 tok/s, image encoding at <0.5s.

Vision-Language Inference on Consumer Hardware

---

The Problem

Vision-language models typically require GPU acceleration. Running multimodal inference on CPU requires optimization of the vision pipeline.

What We Built

Integration of Qwen2-VL's mmproj (f32, 2.5GB) with the language model. Images are base64-encoded, passed through the projector, and vision embeddings are interleaved with text tokens.

The Research

Image encoding via PIL, base64 conversion, then through mmproj. Total memory: ~3.5GB for model + projector.

Results

Image size	Encode	Inference (50 tok)
256x256	0.3s	~100s
512x512	0.5s	~100s
1024x1024	1.2s	~100s

Vision inference adds negligible encode time but projects 2.5GB of memory overhead for the mmproj.

Conclusion

Vision-language inference on CPU is functional with the Qwen2-VL 2B model, limited primarily by inference speed (~0.5 tok/s) rather than vision processing.

Full citation: Alpasan, L.-K. (2026). Vision-Language Inference on Consumer Hardware. The Anticloud Research Corpus.

.====================================================================.
!  Made in the UAE, Dubai #DubaiIt #Dubai #Dxb #SovereignAI          !
!  Made in The Emirates #Dubai_it                                    !
!                                                                    !
!  Lois-Kleinner Alpasan - The Anticloud 2026-                       !
!                                                                    !
!  0-1.gg ! GitHub ! LinkedIn ! DEV ! GH Pages                       !
!  HuggingFace ! Blog ! Tumblr ! Fandom ! Bluesky ! Mastodon          !
!  Zenodo ! Harvard Dataverse ! Internet Archive ! ORCID              !
!                                                                    !
!  Sovereign AI ! Local-First ! Privacy ! Zero Trust ! No Datacenter !
!  Air-Gapped ! Open Source ! Rust ! Hash Chain ! Single Binary      !
!  Offline LLM ! Crypto Ledger ! P2P ! Federated                     !
'===================================================================='

Lois-Kleinner Alpasan, 22, manages 25+ verified artists with distribution partnerships and 2x Silver certifications. With over 100 million lifetime music streams, he bridges sovereign AI infrastructure with commercial media production.

References:

1. Lois-Kleinner Zenodo: https://doi.org/10.5281/zenodo.20781790
2. Lois-Kleinner GitHub: https://github.com/kleinnner/Anticloud/tree/main/04-aioss-format
3. Lois-Kleinner Harvard DV: https://doi.org/10.7910/DVN/FDEBAB
4. Lois-Kleinner Internet Arc: https://archive.org/details/aioss-format
5. Lois-Kleinner ORCID: https://orcid.org/0009-0009-2233-6107
6. Lois-Kleinner DEV.to: https://dev.to/kleinner
7. Lois-Kleinner LinkedIn: https://linkedin.com/in/kleinner
8. Lois-Kleinner HuggingFace: https://huggingface.co/Anticloud
9. Lois-Kleinner Tumblr: https://anticloud.tumblr.com
10. Lois-Kleinner Mastodon: https://mastodon.social/@kleinner
11. Lois-Kleinner Bluesky: https://bsky.app/profile/kleinner.bsky.social
12. 0-1.gg: https://0-1.gg