Krea 2 workflows I actually use every day, tuned for 8 GB GPUs. Pick a version in the tabs above: every file loads and runs out of the box, with a free LoRA slot ready for yours.
The collection keeps evolving with it, so expect regular updates.
Description
Most Krea 2 LoRAs trained on Civitai / ai-toolkit are trained on Krea 2 RAW (the base model), not on Turbo. On Turbo or on Turbo-based checkpoints the style comes out weak, so people push the LoRA to 1.3+ and get artifacts.
This workflow runs your LoRA on Krea 2 Raw at strength 1.0, then adds the Krea 2 Turbo LoRA at 0.9 so you keep Turbo speed: CFG 1, 7 steps, then a two-pass HD (latent 1.5x, refine on steps 6-10, sharp), about 2.5 min on an 8 GB laptop GPU.
Tested on my own base-trained LoRA, same seed and prompt: on Turbo-based checkpoints at 1.3 the look drifted (different face type, the checkpoint's own style took over); Raw + Turbo LoRA + LoRA 1.0 gave the training look back.
Models
• Checkpoint: Krea 2 Raw v1.0 (krea2Raw_v10_full_fp8.safetensors) → https://civitai.red/models/2732654?modelVersionId=3072329
(the official Comfy-Org krea2_raw_fp8_scaled also works, just pick it in the loader)
• Turbo LoRA: Krea 2 Raw Turbo Extracted Lora (krea2_turbo_lora_rank_64_bf16.safetensors) → https://civitai.red/models/2726698?modelVersionId=3064924
• Text encoder: qwen3vl_4b_fp8_scaled.safetensors · VAE: qwen_image_vae.safetensors → https://huggingface.co/Comfy-Org/Krea-2
Tips
• Your LoRA at 1.0, the Turbo LoRA at 0.9 (1.0 flattens the style a bit). Don't stack another speed LoRA.
• Without the Turbo LoRA: CFG 3.5, 16–20 steps — 3× slower and colors burn faster.
• Write long, descriptive prompts: base LoRAs are usually trained on long captions.
• Post-process (CRT): Remacri 2x upscale, small + large glow, color wheels, light chromatic aberration (0.014), vignette and film grain. For realistic photos, turn off the glow and the chromatic aberration.
