Local diffusion workflow: krea-turbo + Qwen3-VL encoder generates 7680x7680 photoreal portraits in 8 steps
Taugamori · reddit · 2026-11-21
A Reddit user shares their two-day-old portrait generation workflow:
- Stack: diffusion model krea-turbo-int8convrot, text encoder qwen3vl4bint8convrot, wan2.1vae, and seedraskinkrea2 skin LoRA
- Pipeline: generate at 1536×1536 → VAE decode → bicubic upscale to 7680×7680 → save PNG
- Settings: seed 123456, 8 steps, CFG 1.0, Euler sampler, Beta scheduler, shift 4.0, denoise 1.0, empty negative prompt
The sample output is a photorealistic portrait with convincing skin texture and lighting, demonstrating what this low-step local combo can produce.
More from Multimodal
- Indie dev open-sources dflow_2b, a 2B text-to-image model trained on a single GPU — quantiler · 2026-11-21
- Ark: an AI-made musical horror tale set in a world where the sun never rises — SoullessGamesDev · 2026-11-21
- AutoCut + FrameAtlas: open-source local-first video editing driven by Claude Code via MCP — moheshmohan · 2026-11-21
- 10 ChatGPT image prompts that generate full brand design kits in 2 minutes — anthara_ai · 2026-11-21
- Redditor shares fully AI-generated "Sold Y - Gang" music video — Awkward-General2552 · 2026-11-21
- Hunting for an open-source video interpolation alternative to Topaz Aion — Glittering-Cold-2981 · 2026-11-21