Krea 2 Smartphone Photography LoRA: Concept Sliders & Color Correction
sktksm · reddit · 2026-07-22
The author shares their complete experiment process and findings from training a 'Smartphone Photography' concept slider LoRA for the Krea 2 model.
Key Findings & Workflow:
- Prompting over Direct Training: The author found that directly training LoRAs often flattened skin details, but the base model already possessed the capability for realistic amateur photography via prompting.
- Concept Sliders: A slider LoRA based on prompting insights can consistently push the model toward the desired look without reproducing complex prompts every time.
- Parameters: The LoRA is extremely strong; recommended strength is between 1.0 and 2.0 (sweet spot around 1.5).
- Post-Processing: Because training causes heavy oversaturation, using a ComfyUI Color Correct node post-VAE Decode is strongly recommended (saturation between -15 and -20).
- Sampler Setup: Uses a two-pass setup: 1st pass with Euler/Beta (12 steps), 2nd pass with res4smunthe-kass/kloptimal (3 steps).
More from Multimodal
- Fable users demo a procedural dog simulation built with three.js agent skills — willdepue · 2026-07-22
- NVIDIA Cosmos 3 tops 8+ open leaderboards across world reasoning and generation — MonaJalal_ · 2026-07-22
- vLLM-Omni v0.22 adds Cosmos 3, robot serving, and broader quantization support — MonaJalal_ · 2026-07-22
- Instruct-Particulate infers articulated 3D structures from static meshes — MonaJalal_ · 2026-07-22
- Wan2GP users test negative prompt and NAG settings on LTX 2.3 — ffzero58 · 2026-07-22
- StreamDiffusion + BiRefNet Powers Real-Time Infinite Canvas Generation — PurzBeats · 2026-07-22