Realism LoRAs break under fp8 quantization on SD Forge: an 8GB AMD user's plea
Automatic-Passion-37 · reddit · 2026-10-28
An RX 6650 XT (8GB VRAM) user on SD Forge (zluda AMD fork) reports that realism LoRAs perform terribly when the main checkpoint (cyberrealistic pony merge) runs in fp8 quantization—plasticky skin, lost details, sometimes no effect at all—while character LoRAs work fine. He suspects fp16-trained LoRAs don't blend properly with an fp8-quantized UNet.
Attempts so far: no stable toggle to force LoRAs/text encoder to fp16 while keeping the UNet in fp8; switching fp8 modes changes nothing; raising LoRA weights to 2.5–3 helps slightly but artifacts kill the effect. He's asking for launch arguments, Forge settings, or LoRA block weight configs that fix this without blowing VRAM.
More from Multimodal
- Google open-sources EmbeddingGemma 2, a unified multimodal embedding model for text, image, audio, video — google · 2026-10-28
- NVIDIA open-sources Nemotron 3.5 Transcribe: 30B MoE ASR model with ~4B active params — NVIDIAAI · 2026-10-28
- Agentic video editing is 'solved': hundreds of videos/day, timeline editing obsolete — ZeroStateReflex · 2026-10-28
- Seedance 2.0 Mini demo: bubble-trick penguin AI wildlife clip — Burak_Antakya · 2026-10-28
- MOSS meets the ComfyUI team as MOSS-Image lands ComfyUI support — Physical_Hat4022 · 2026-10-28
- Blind test of 152 listeners: Suno matches human composer, but almost no one can tell which is AI — teropa · 2026-10-28