Krea 2 LoRA training fits in 16GB VRAM, but the resulting LoRA bleeds into no-trigger prompts
Economy_Cucumber_702 · reddit · 2026-07-29
A detailed Krea 2 LoRA training report shows that 16GB of VRAM is enough for 768px training, contrary to a widely shared guidance thread.
Key results
- 1,152 steps completed in 67 minutes at 3.42 s/it
- Peak memory usage hit 15,284 MiB of 16,303 MiB VRAM (93.8%) and 17.5GB of 32GB system RAM
- After training, Turbo inference ran at about 13 seconds per 768×1024 image with 8 steps
Setup details
- GPU: RTX 5080 16GB
- CPU: Ryzen 9 9950X
- OS: Windows 11 Pro native, no WSL2
- Stack: torch 2.13.0+cu130, accelerate 1.6.0, transformers 4.57.6, diffusers 0.32.1, bitsandbytes 0.50.0, musubi-tuner 0.3.4
- Attention backend: plain SDPA; no Triton, flash-attn, xformers, or SageAttention
Model and file notes
- Krea 2 uses Qwen3-VL-4B-Instruct as text encoder and Qwen-Image VAE
- Training should use the RAW bf16 checkpoint; inference should use Turbo bf16
- Official Hugging Face repos are gated, but the Comfy-Org/Krea-2 mirror is ungated and matches the official RAW file byte-for-byte
- The pre-quantized fp8 Turbo file is a ComfyUI artifact and does not load in musubi because of extra keys
Caveat
The author says the resulting LoRA has a real flaw: it bleeds into no-trigger prompts, and neither an earlier checkpoint nor a lower multiplier fixed it. The post is a single-run report with no ablations, so it should be read as a practical benchmark rather than a definitive study.
More from Models
- Scobleizer says Grok 4.5 is the best coding model right now — Scobleizer · 2026-07-29
- Apple reportedly sues OpenAI over alleged trade-secret misuse tied to future hardware — emmanuelvivier · 2026-07-29
- Kimi K3 and Qwen3.8 show Chinese AI is now a sustained competitive force — emmanuelvivier · 2026-07-29
- User seeks a $1,300–$1,700 GPU with 48GB VRAM for local LLMs — Syosse-CH · 2026-07-29
- GPT-5.6 reportedly solves Feige’s 1/e conjecture in probability — MickeySteamboat · 2026-07-29
- BetterGPT-150M debuts as a 152M-parameter base model trained on 15B tokens — Rich-Title-3668 · 2026-07-29