Solo Dev Trains Open Pixel-Art DiT for 126 Hours on an RTX 4080
ApprehensiveTart3158 · reddit · 2026-11-20
Inspired by a trending HF dataset, an indie developer trained a DiT from scratch for the first time and released LiliPixel-V1-190M after six iterations: RGBA at 16×16 and 32×32, RGB at 64×64. The progression is instructive — v1 had no text encoder and only did 16×16; adding Voyage 4 nano embeddings as the text encoder greatly improved prompt adherence; from epoch 7 the data mix shifted toward more 64×64 images. Total: 126 hours on a single RTX 4080, 12 epochs (30M non-unique images seen), weights released under a fully open license in diffusers format.
More from Multimodal
- Chinese team open-sources real-time 3D scene reconstruction from monocular video at 20 FPS — dr_cintas · 2026-11-20
- Ideogram to skip open-sourcing 4.5 and release 4.6 open weights directly — multimodalart · 2026-11-20
- Fable 5.2 records and edits an entire game montage solo after 19 directed iterations — gdechichi · 2026-11-20
- 8-minute music film made 100% locally in ComfyUI on a single Mac, consistent faces throughout — Mardoslazar · 2026-11-20
- 30-second ghost train short made with Seedance 2.5, prompt shared — LudovicCreator · 2026-11-20
- fofr shares code-generated anime-style neon art of an indoor pool — fofrAI · 2026-11-20