Running image generation in the browser on local hardware: 10s pixel art on an RTX 3060
Bartholomheow · reddit · 2026-09-26
A Reddit user demonstrates running small image generation models directly in the browser using local hardware, enabling dynamic asset creation without a server.
- First attempt: converting the tiny Supra2 model to ONNX worked in Chrome but was slow with poor output
- Second attempt: converting aziibpixelmix, a 4-step pixel art model, to ONNX — generates 512x512 pixel art in 10s on an RTX 3060; model and demos published on Hugging Face
- Lessons: browser-side inference suffers from obscure unmaintained bugs, precision loss even with standard setups, WASM CPU mode runs out of memory, and integrated GPU works as fallback; expect "worked on my machine" issues across GPUs and drivers
The author asks for help generalizing step count and resolution, since ONNX conversion locked the model to fixed settings.
More from Infra
- WRX90E BIOS Lacks PCIe Bifurcation, Needs Custom BIOS for Multi-GPU — TheZachMueller · 2026-09-26
- DIY '2400cc Inference Racer': Dual RTX 3090s, ThinkPad ECU and a VW Golf Radiator Serve 27B Models — TooManyPascals · 2026-09-26
- A100 SXM4 connector production ends, forcing new tooling investment — EAccelerate_42 · 2026-09-26
- Two local AIs talk to each other with no cloud: hands-on with Braid — Scobleizer · 2026-09-26
- The handiest GPU this dev ever bought is a ~$300 Intel Arc A310, not NVIDIA or AMD — TheZachMueller · 2026-09-26
- Alibaba's T-Head unveils Zhenwu V900 chip: 216GB per card, sales in Q1 2027 — shashib · 2026-09-26