Feeding Real Photos to the Distillation Critic: Krea2 LoRA Stays Sharp at Just 2 Steps
TimeTruth2490 · reddit · 2026-08-28
A community trainer shares mid-training progress on a Krea 2 Turbo distillation LoRA: on top of the existing Progressive Distillation, he adds a GAN critic in latent space (LADD-style Latent Adversarial Diffusion Distillation). The key twist: the critic's "real" reference isn't limited to teacher outputs—it's partly fed real photographs, which is where the surprising robustness comes from. Cost: 2-3x slower training. The critic exists only at training time; the released LoRA stays a plain drop-in file.
Extreme test: running the LoRA far outside its 4-step training spec—2 steps, strength 2—against native Krea 2 Turbo: native output is ghosted, smeared and half-formed, while the LoRA delivers coherent, sharp compositions across all 15 test prompts. Not for production, but useful as a fast preview: 2 steps give a reliable read on composition at a quarter of the 8-step cost; works best on close subjects/portraits, degrades on distant scenes. Based on these results, the author is considering a properly trained 2-step LoRA. Comparison images are on Hugging Face.
More from Multimodal
- Creator turns self into mecha anime character using AI — JBOOGZEE · 2026-08-28
- User tests Minimax H3: struggles with motion control adherence — cal_01 · 2026-08-28
- Saudi Photographer Switches Entire Workflow to Google Veo — aziz4ai · 2026-08-28
- Dev shares progress on Blender-to-Unity scene pipeline — kevinkern · 2026-08-28
- Seedance 2.5 tops motion video leaderboard, beating Gemini and Grok — yuwen_lu_ · 2026-08-28
- MiniMax H3 Community Roundup: Reverse Painting Video, PDD Acceleration, Audio Fix — optimisticalish · 2026-08-28