A refined and enhanced workflow by Experimental_dAIver.

This workflow brings the impressive quality, prompt adherence and efficiency of the new Krea 2 Turbo model family into your local ComfyUI setup. It is specifically optimized for low-VRAM GPUs (tested on RTX 4050 6 GB) and features the powerful NdSuperLoraLoader for seamless multi-LoRA stacking with automatic trigger-word handling, CacheDiT_Model_Optimizer + PathchSageAttentionKJ for noticeable speed gains, the intuitive selectLatentSizePlus aspect-ratio and resolution selector, ModelSamplingAuraFlow with shift control, and a complete integrated SEEDVR2 Video Upscaler Subgraph for outstanding 4K+ stills and short video sequences. The workflow uses the dedicated krea2 CLIP type together with qwen_image_vae for maximum compatibility and quality.

Version 1.0 is the initial public release and delivers a clean, fast and highly flexible base specifically tuned for Krea 2 Turbo (and Base) checkpoints.

To use Krea2, you need the latest ComfyUI versions 0.25 and later.

Key features in v1.0:

• Full native support for Krea2_Turbo_fp8mixed.safetensors and all other quantized variants (FP8 / MXFP8 / NVFP4 / INT8 / ConvRot INT8) from the Winnougan collection

• NdSuperLoraLoader (nd-super-nodes) with full multi-LoRA support, auto-fetch trigger words and clean strength control

• CacheDiT_Model_Optimizer + PathchSageAttentionKJ – both optional but enabled by default for significant speed improvements on low-VRAM hardware with almost no quality loss

• selectLatentSizePlus – convenient aspect-ratio and resolution selector with photography presets (including golden-ratio-friendly options) and one-click orientation swap

• Advanced sampling via KSamplerAdvanced (ER_SDE or Euler) + ModelSamplingAuraFlow (shift 4 recommended) with Beta or Simple scheduler – excellent results already at 8–10 steps

• High-quality decoding with qwen_image_vae

• Complete SEEDVR2 Video Upscaler Subgraph – powerful DiT-based upscaler that delivers stunning high-resolution results on still images and short video clips, with Lab color correction, temporal settings and intelligent resolution handling

• Professional output options with SaveImageExtended and ChronoSaveForCivit (perfect for direct CivitAI uploads with embedded metadata and workflow)

• Self-documenting workflow containing a detailed MarkdownNote with all download links, recommended settings and folder guidance

Required Custom Nodes

• nd-super-nodes (HenkDz) – NdSuperLoraLoader and modern UI components https://github.com/HenkDz/nd-super-nodes

• ComfyUI-CacheDiT – CacheDiT_Model_Optimizer

• comfyui-kjnodes – PathchSageAttentionKJ

• ComfyUi-MzMaXaM – selectLatentSizePlus https://github.com/MzMaXaM/ComfyUi-MzMaXaM

• save-image-extended-comfyui – SaveImageExtended

• ComfyUI-SeedVR2_VideoUpscaler – full SEEDVR2 Video Upscaler subgraph https://github.com/ainvfx/ComfyUI-SeedVR2_VideoUpscaler

Important: Update ComfyUI to the latest version before installing the custom nodes. Krea 2 support and the involved nodes benefit greatly from recent core updates.

Models & Downloads (exact paths)

The following list explains exactly which files you need and where to place them.

1. Main Diffusion Model (Krea 2 Turbo FP8):

File: Krea2_Turbo_fp8mixed.safetensors (or any other quant variant you prefer) Download: All quantized models (FP8 / MXFP8 / NVFP4 / INT8 / ConvRot INT8) for both Base and Turbo: 👉 Winnougan/Krea-2-Base-Turbo-NVFP4-FP8-INT8 Target folder: ComfyUI/models/diffusion_models/Krea2/ (Create the Krea2 subfolder if it does not exist. The UNETLoader references this exact path structure.)

2. Text Encoder:

File: qwen3vl_4b_fp8_scaled.safetensors Download: https://huggingface.co/Comfy-Org/Qwen3-VL/blob/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors Target folder: ComfyUI/models/text_encoders/

Important: In the CLIPLoader node set the type widget to krea2.

3. VAE:

File: qwen_image_vae.safetensors Download: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/blob/main/split_files/vae/qwen_image_vae.safetensors Target folder: ComfyUI/models/vae/

4. LoRAs (for NdSuperLoraLoader – full multi-LoRA support):

Any compatible LoRAs (.safetensors). Target folder: ComfyUI/models/loras/

5. SEEDVR2 Models (for the integrated upscaler subgraph – optional but highly recommended):

• DiT Model: seedvr2_ema_3b-Q8_0.gguf

• VAE: ema_vae_fp16.safetensors

Download from the official ComfyUI-SeedVR2_VideoUpscaler repository or Hugging Face and follow the node’s installation instructions.

Folder Structure Overview

text

📂 ComfyUI/
├── 📂 models/
│   ├── 📂 diffusion_models/
│   │   └── 📂 Krea2/
│   │       └── Krea2_Turbo_fp8mixed.safetensors
│   ├── 📂 text_encoders/
│   │   └── qwen3vl_4b_fp8_scaled.safetensors
│   ├── 📂 vae/
│   │   └── qwen_image_vae.safetensors
│   └── 📂 loras/
│       └── (your LoRAs here)

Recommended Settings

• Sampler: er_sde (recommended) or euler

• Scheduler: beta (recommended) or simple

• Steps: 8–10 (Turbo model delivers excellent quality at very low step counts)

• CFG Scale: 1.0

• Shift (ModelSamplingAuraFlow): 4–7 (4 is the sweet spot in most cases)

• Resolution & Aspect Ratio: Freely selectable via selectLatentSizePlus – includes “2:3 Standard Photography”, many other standard ratios and resolutions up to 2048 px. Use swap_orientation for instant portrait/landscape flip. Batch size supported.

• LoRA Strengths: Start with 0.6–1.0 per LoRA when stacking multiple models.

A compact settings table is also embedded directly in the workflow’s MarkdownNote node for quick reference.

How to use the workflow

1. Load the JSON in ComfyUI.

2. Select your desired aspect ratio and resolution in the selectLatentSizePlus node (enable swap_orientation if needed).

3. Enter your prompt in the Positive Prompt node (the NdSuperLoraLoader automatically handles trigger words from your LoRAs).

4. Load one or more compatible LoRAs into the NdSuperLoraLoader.

5. Generate. The workflow produces a high-quality base image extremely efficiently.

6. (Optional) Run the integrated SEEDVR2 upscaler subgraph for beautiful high-resolution results on still images or short video sequences.

7. Use SaveImageExtended and/or ChronoSaveForCivit for full filename control, metadata embedding and CivitAI-ready exports.

You can bypass or mute the upscaler group completely for fast testing. The workflow also works with other Krea 2 checkpoints – simply change the filename in the UNETLoader. CacheDiT and SageAttention can be disabled individually if you ever encounter compatibility issues (they are optional speed boosters).

This is a clean, fast and powerful base specifically tuned for the new Krea 2 Turbo FP8 model — perfect for quick iterations, creative multi-LoRA compositions and high-quality final outputs even on modest hardware.

Huge thanks to Winnougan for creating and quantizing the excellent Krea 2 models, and to HenkDz (nd-super-nodes), the CacheDiT team, kjnodes, MzMaXaM, ainvfx (SeedVR2), as well as the authors of SaveImageExtended and ChronoSaveForCivit for making these advanced, low-VRAM workflows possible.

If you have questions or want to share your results — I’m happy to hear from you in the comments. Enjoy the workflow!