A 480K-parameter latent model cleans up GPT Image 2 speckle artifacts
Parking_Baby_57 · reddit · 2026-07-27
A tiny latent residual removes GPT Image 2’s speckle artifacts
The author built a 0.48M-parameter residual model to clean up a consistent failure mode in GPT Image 2: over-sharpening, bright specks, hard edges, and scale-like texture artifacts.
How it works
- Encode the image with the FLUX.2 VAE
- Add a scaled residual predicted from the latent: z + α·R(z)
- Decode back through the VAE
The model runs in the VAE latent space because the artifacts behave like a texture statistic layered over the content. The author says this is cheaper than pixel-space denoising and easier to separate from real detail.
Cost and limits
- About 0.7s per 1.5MP image on a recent GPU
- Around 3GB VRAM
- Best α varies by image
- It can only dim artifacts; in some cases it also softens genuine detail
- Macro structure is unchanged, and the whole image still round-trips through the VAE
The post includes a demo site, a Hugging Face Space, and GitHub code plus weights.
More from Multimodal
- Do Danbooru tags for Anima and Illustrious need to be real searchable labels? — Manicarus · 2026-07-27
- LTX Crossview IC-LoRA adds gizmo-based camera control for AI video — TomLikesRobots · 2026-07-27
- How do beginners usually set LoRA strengths and order in ComfyUI? — Murky_Room4447 · 2026-07-27
- Anima keeps backgrounds but loses custom character consistency, user asks if LoRA fixes it — Capital-Caregiver818 · 2026-07-27
- Two RTX 5080s failed to beat one RTX 5090 in ComfyUI single-render tests — Geekdomo · 2026-07-27
- InVideo Agent One ranks #1 on an independent AI video-agent benchmark — aziz4ai · 2026-07-27