Fizgig 8.6 ships MOSS-Image LoRA fixes, shares hard-won training tips

shootthesound · reddit · 2026-10-19

Fizgig 8.6 is out with MOSS-Image LoRA training improvements plus practical tips from extensive testing.

• LoRAs now train block 22 (previously skipped, as AI Toolkit does); side-by-side runs show better likeness from early epochs and more stable training
• Recommend training on Base and applying the LoRA to SFT in ComfyUI; if training on SFT, use the new attention-only toggle or anatomy breaks within epochs
• The Fast preset is back (0.25 MP, rank 16) and works well on face-cropped datasets; the author admits his earlier dismissal of 0.25 MP was wrong
• MOSS converges deceptively slowly, so presets now run 70 epochs; also fixes MOSS preview crashes with block swap on smaller GPUs

Original post →

More from Multimodal

• Pika demos Motion Transfer: one photo plus one clip moves any character — lmoroney · 2026-10-19
• Midjourney sref combo turns Toy Story characters into retro anime style — sergeantsref · 2026-10-19
• Indie dev builds Utsushi, an open-source Higgsfield clone in Go/HTMX — _playlogic_ · 2026-10-19
• AI-Generated Fantasy Short Film 'Skirmish at the Ruins' Released — QuiteOak · 2026-10-19
• Testing Krea2 Art Style Prompts on Moss Image Yields Moody Results — Large_Election_2640 · 2026-10-19
• 'The Glitch': A Robot Folk Horror Short With a Cracked-Doll Android — BotLordBotflix · 2026-10-19

Multimodal channel →