Low-Quality Photo LoRA Test: Flux1-dev Still Beats Krea 2, Z-Image and Flux2
wreck_of_u · reddit · 2026-09-25
Testing character LoRAs trained on 15 blurry 1990s photos, the author finds Flux1-dev delivers the best likeness — beating Krea 2, Z-Image and Flux2. Newer models apparently require high-quality datasets, while Flux1 handles overfitted low-quality LoRAs well; upscaling the source images only produced "GAN-like" slop. Workflow: ai-toolkit, 512x512, manual captions, no enhancement.
More from Models
- Micius-Flash: 27B Multimodal Model for Hardware and Semiconductor Agents — zhiman-ai · 2026-11-11
- Baidu says RSI auto-optimized ERNIE inference, 6.24x decode speedup on Kunxin P800 — xiaohu · 2026-11-11
- Baidu's ECHO: multi-token prediction reusing the last layer as a drafter — xiaohu · 2026-11-11
- ERNIE 5.5 Flash Preview: 280B-param multi-router MoE activating only 14B per token — xiaohu · 2026-11-11
- ERNIE 5.5 recreates the Pelican test: 117-second animation drawn frame-by-frame in Python — xiaohu · 2026-11-11
- 35,000+ Grok-built apps exposed patient data, plaintext passwords, and sales pipeline — The Decoder · 2026-11-11