High school grad builds RGBA VAE for FLUX.2, enabling true transparent image output
Superb_Composer_3389 · reddit · 2026-09-30
A developer fresh out of high school spent months making FLUX.2 Klein output natively transparent images, releasing weights alongside a first-ever paper.
Components:
- RGBA VAE: the FLUX.2 VAE widened to 4 channels and fine-tuned so latents carry alpha.
- Extract-9B: a LoRA for Klein Base 9B that takes a photo plus the empty background and returns the object as a transparent PNG, keeping soft shadows semi-transparent.
- Remove-9B: an object-removal LoRA that chains with Extract, requiring only a brush over the object.
Numbers (small test set): Extract-9B hits IoU 0.934 on 20 held-out objects, identical scores at 6 vs 25 steps (11 s/image, 10 GB VRAM in fp8). The RGBA VAE costs 10–18% colour error vs stock — the price of fitting alpha.
Limitations: Remove-9B struggles with flat graphics; no tests above 512². An extractor trained on Qwen-Image-2.1 matches accuracy, runs faster, but needs two GPUs; the paper compares them (author favors Qwen).
Weights are on Hugging Face under a non-commercial licence, with a Gradio demo. Credits AlphaVAE and ostris' ai-toolkit.
More from Multimodal
- AssemblyAI launches Universal-3.6 Pro Realtime, tops two streaming speech benchmarks — AssemblyAI · 2026-09-30
- AI studio delivers Odyssey pilot in 15 days with 2,070 images and 1,558 videos — chaitu · 2026-09-30
- Viggle Launches PINOC, a 3D Engine Turning One Photo Into an Animatable Gaussian Character in a Minute — HeyAmit_ · 2026-09-30
- Fast Node Refresh mod syncs ComfyUI LoRA lists via file deltas, ending 30s+ freezes — Incognit0ErgoSum · 2026-09-30
- Can MiniMax H3 run on a 4090? Users weigh RAM spillover needs — 1thatonedude1 · 2026-09-30
- ComfyUI Refmod Beats LORAs: Seconds to Build, No Reference Scanning — Free_Pressure8623 · 2026-09-30