Krea 2 Turbo in 3 Steps: The 3-Step Distill LoRA for ComfyUI (2.6x Faster)

Krea 2 Turbo already makes good images in 8 steps, but 8 steps on a 12B model still means a wait on a home GPU. If you just drop the step count to 3, the images go soft and lose their fine detail. The new Krea2-Turbo-Distill-3step-LoRA by lvladikov fixes that. It is a small LoRA that teaches Krea 2 Turbo to finish in 3 steps while keeping, and by the author’s numbers even slightly beating, the fine texture of the 8-step output. Denoising drops from 65.4 seconds to 24.9 seconds in the author’s test, about 2.6x faster. This guide covers the files you need, the exact ComfyUI folders, the settings that matter, what the trade-offs are and how it compares with the 2-step and 4-step versions.

What Is the Krea 2 3-Step Distill LoRA?

Krea 2 is Krea AI’s 12B open image model. It comes in two flavours: Raw, which needs many steps and real CFG, and Turbo, which has Krea’s own distillation baked in and runs in 8 steps with CFG off. Turbo is what most people use in ComfyUI.

Hugging Face user lvladikov has been pushing Turbo further with a series of step-distillation LoRAs. The 4-step LoRA came first, then a 2-step preview, and on October 28, 2026 the 3-step version. It sits between the two: one step fewer than the 4-step LoRA and one step more than the 2-step one, and the author says it beats both on quality.

It is a plain rank-64 LoRA over 228 modules. That matters because it loads with the stock LoraLoaderModelOnly node in ComfyUI, in diffusers, and in MLX on a Mac, with no custom nodes. Two weight files hold the same checkpoint, one with diffusers key names and one with ComfyUI key names.

The model page is lvladikov/Krea2-Turbo-Distill-3step-LoRA on Hugging Face. If you have not set up Krea 2 before, our Krea 2 BBOX guide walks through the base install.

Files You Need and Where They Go

Make sure you download the file with comfyui in its name. The diffusers version has different key names and will not apply correctly in ComfyUI.

Your folders should look like this:

ComfyUI/
  models/
    diffusion_models/krea2_turbo_bf16.safetensors
    text_encoders/qwen3vl_4b_bf16.safetensors
    vae/qwen_image_vae.safetensors
    loras/krea2_turbo_3step_rank_64_lora_comfyui.safetensors

If you already run Krea 2 Turbo, you only need the one LoRA file.

Set It Up in ComfyUI

The author provides a ready workflow, krea2_turbo_3step_lora_t2i.json, in the model repo. It uses stock nodes only. Download it and drag it onto the ComfyUI canvas.

To add the LoRA to your own Krea 2 Turbo workflow instead:

1. Update ComfyUI so it has native Krea 2 support (portable: run update/update_comfyui.bat; manual install: git pull in the ComfyUI folder, then pip install -r requirements.txt).
2. Open the Krea 2 Turbo text-to-image template from Workflow, Browse Templates, Image.
3. Add a LoraLoaderModelOnly node between Load Diffusion Model and the sampler.
4. Pick krea2_turbo_3step_rank_64_lora_comfyui.safetensors and set strength to 1.0.
5. In the sampler, set steps to 3, CFG to 1.0, sampler to euler and scheduler to simple.
6. Write your prompt and queue.

The chain looks like this:

Load Diffusion Model (krea2_turbo_bf16)
  -> LoraLoaderModelOnly (3-step LoRA, strength 1.0)
  -> KSampler (3 steps, CFG 1.0, euler, simple)
  -> VAE Decode (qwen_image_vae)

Other Krea 2 style LoRAs can go before or after the distill LoRA in the chain. Results vary, because those LoRAs were trained for the normal 8-step Turbo schedule.

Best Settings

The CFG point trips people up. In ComfyUI, CFG 1.0 skips the negative pass, which is what a distilled model wants. Setting 0.0 sounds like “less guidance” but actually produces a broken image.

The prompt works the same as normal Krea 2 Turbo. The author’s example is a good test: Astronaut in a jungle, cold color palette, muted colors, detailed, 8k.

Speed: How Much Faster?

The author measured denoising time at 1024×1024 in float16, leaving out text encoding and VAE decode:

That is about 2.6x faster, which matches the drop from 8 steps to 3. Text encoding and VAE decode add a fixed cost on top, so the end-to-end gain per image is a little smaller than 2.6x.

Estimate for a 16 GB card like the RTX 5080: Krea 2 is a 12B model, so on 16 GB you load it with fp8 weights or let ComfyUI offload part of it. Based on the 8-to-3 step reduction and typical fp8 speed for a 12B model at 1024×1024, expect roughly 5 to 10 seconds of denoising per image with this LoRA, compared with roughly 15 to 25 seconds for stock 8-step Turbo. This is an estimate, not a measured result.

Quality vs the 8-Step Output

Fewer steps normally means less detail. The author measured fine texture over 22 test prompts, with the 8-step Turbo image as the reference (1.0):

• Stock Turbo at 3 steps: 0.58x. Just lowering the step count loses almost half the fine texture.
• 4-step LoRA: 1.04x.
• 3-step LoRA: 1.11x. Slightly more fine texture than the teacher.

Detail at the 16-pixel and 8-pixel scales comes out at 1.10x and 1.12x, and colour matches the teacher. In short, the 3-step LoRA does not look like a blurry fast preview. It looks like a full Turbo image, sometimes a touch crisper.

Trade-offs and Known Issues

More texture is not free. The author is open about the costs:

• It drifts further from the 8-step image. Same prompt and seed give a similar but less identical picture than the 4-step LoRA. If you need the 3-step image to match an 8-step one closely, use the 4-step LoRA.
• Grain in flat areas. Skies, walls and smooth skin can show fine noise. Lower the strength to around 0.8 or 0.9 to soften it.
• Ghosting. Faint double edges show up mostly at 768×1024 and 1024×768. Square and larger sizes are cleaner.
• Small faces in crowds are still weak, as with most fast models.
• Quantized bases lose detail. A 4-bit nvfp4 base, as used on Comfy Cloud Turbo, measured about a fifth lower in fine texture than full bf16.
• Raw is not supported. Using it on Krea 2 Raw (13 steps, CFG 3.5) is untrained. The author says it adds detail but it is not a supported mode.

2-Step vs 3-Step vs 4-Step

My pick: use the 3-step LoRA by default. Switch to the 4-step LoRA when you are matching an existing 8-step image or the grain bothers you, and to the 2-step LoRA only for fast idea sketches.

Running It on a 16 GB GPU

The LoRA itself is small. The challenge is the 12B Krea 2 Turbo base. The bf16 file does not fit in 16 GB of VRAM on its own, so:

• Load in fp8. In the Load Diffusion Model node, set weight_dtype to fp8_e4m3fn. That stores the weights in 8-bit at about half the size, with a small quality cost.
• Let ComfyUI offload. ComfyUI moves part of the model to system RAM when VRAM runs out. 32 GB of system RAM is the minimum, and 64 GB is better.
• Avoid nvfp4 if you care about detail. It is the smallest option, but the author measured about 20% less fine texture with it.

For more on fitting big models into 16 GB, see our ComfyUI VRAM guide.

Mac and MLX

On Apple Silicon, fp8_scaled builds will not load, because the MPS backend does not support Float8_e4m3fn. Use the bf16 base, or the author’s MLX workflow, which uses the LVNodes Krea 2 MLX node. In diffusers, load the base with torch.bfloat16 and set the device to "mps":

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained("krea/Krea-2-Turbo", dtype=torch.bfloat16, device_map="mps")
pipe.load_lora_weights("lvladikov/Krea2-Turbo-Distill-3step-LoRA")
image = pipe("Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
             num_inference_steps=3, guidance_scale=1.0).images[0]
image.save("krea2_3step.png")

On an NVIDIA PC, change "mps" to "cuda".

How It Was Trained

• Starting point: the 4-step adapter’s weights.
• Data: about 30,000 samples of recorded teacher trajectories at 11 resolutions.
• Schedule: three calls on Turbo’s 3-step sigmas, [1.0, 0.8633, 0.6123], mu 1.15, no guidance.
• Objective: distribution matching, scored by eight critics on the frozen base model’s mid-network features.
• Hardware: a single H100/H200 with a bf16 base for about 91 hours.
• Checkpoint: chk00030000, with the learning rate annealed to zero over the last 600 samples. The author expects more checkpoints, so check the repo for newer files.

Distribution matching is why texture can exceed the teacher. The LoRA is not copying each 8-step image pixel for pixel. It is learning to produce images that look like they came from the same distribution, which is also why it drifts a bit further from the exact teacher image than the 4-step version.

FAQ

Do I need a custom node?

No. It is a standard LoRA. Use the core LoraLoaderModelOnly node.

Why does CFG 0 give a broken image?

In ComfyUI, CFG 1.0 is the “no guidance” setting for distilled models. 0.0 pushes the prediction the wrong way. Always use 1.0 with this LoRA.

Can I use it with Krea 2 BBOX or other fine-tunes?

It was trained on stock Krea 2 Turbo. It may work on Turbo-based fine-tunes, but results are not guaranteed.

Can I use the images commercially?

The LoRA uses the Krea 2 Community License, the same as the base model. Read its terms before commercial work.

Is it better than the 4-step LoRA?

It is faster and has more fine texture. The 4-step LoRA is more faithful to the 8-step image and has less grain and ghosting.

Does it work with img2img or edit?

It was trained for text-to-image. For editing, Krea 2 has separate edit models and LoRAs.

Final Thoughts

The Krea2-Turbo-Distill-3step-LoRA is one of the easiest speed upgrades you can make in ComfyUI right now. One 438 MB file, one stock node, three settings, and Krea 2 Turbo runs about 2.6x faster without the mushy look you get from simply lowering the steps. Watch for grain in flat areas and use square sizes to avoid ghosting. If you already use Krea 2 Turbo, add it to your workflow today.

More AI image guides:

• Krea 2 BBOX in ComfyUI: Comics With Panels and Speech Bubbles
• Best Open-Source Text-to-Image Models
• Best Qwen Image 2.1 Turbo Workflow for 16 GB
• Kromoss 8-Step Workflow for 16 GB GPUs

Sources

• Hugging Face: lvladikov/Krea2-Turbo-Distill-3step-LoRA (model card and files, read November 1, 2026)
• ComfyUI Wiki: Krea 2 3-step distill LoRA
• Hugging Face: 4-step LoRA and 2-step LoRA
• Hugging Face: Comfy-Org/Krea-2