Krea 2 Turbo 2/4-step distill LoRAs get Torch, Apple MLX and Comfy Cloud workflows with auto step switching
TimeTruth2490 · reddit · 2026-10-04
The author of the 2-step and 4-step Krea 2 Turbo distill LoRAs released follow-up ComfyUI workflows and apps (Apache-2.0) covering three setups: Comfy Cloud, local ComfyUI, and Apple MLX. Highlights:
- MLX route runs Krea 2 from ComfyUI's bf16 files in its own sub-environment: 1 min per 1MP image, 41s for subsequent batch images.
- Auto step switching: the graph picks the 2-step LoRA (2-3), 4-step LoRA (4-7), or native model (7+) automatically; default is 4 steps.
- Memory trick: on Apple Silicon via MPS, the step LoRA shares weights with Krea instead of a second copy, halving memory from 48GB to 24GB.
- Random prompts: seeded scene lists of 14k dictionary words or live rows from a 3M+ row Hugging Face dataset, turned photorealistic by a prompt LLM with refusal retries and fallback.
- Extras: style LoRAs with trigger words; upscaling via SeedVR2 7B (3.5MP × 4 → 8864×6624); live local progress; stage-by-stage model loading with optional keep-loaded.
- LVNodes pack: depends only on the Python standard library (MLX nodes auto-install mflux/MLX/transformers 5, 1.3GB, in their own environment).
More from Multimodal
- AI video: Back to the Future DeLorean flies through 100 years of history — mark_k · 2026-10-05
- Open workflow extends static video shots without degradation using Ref2v, masks and bridge resampling — roychodraws · 2026-10-05
- AI Short Puts Two Michael Jackson Personas on One Mic for a Rap Duet — SimplyAnnisa · 2026-10-04
- Kling 4.0 Fast shows surprisingly strong acting: expressions, body language and timing — LudovicCreator · 2026-10-04
- A 14-agent pipeline on one RTX 5090 produced a 10-minute animated short end-to-end — developervkmp · 2026-10-04
- MiniMax H3 ComfyUI workflow generates consistent character sheets, beats Qwen on quality — cgpixel23 · 2026-10-04