lvladikov/Krea2Turbo-PiD-SuperFast-HighRes

Krea 2 Turbo × PiD — Super-Fast High-Res (ComfyUI)

ComfyUI workflows that render high-resolution Krea 2 Turbo pictures fast, in two short stages:

1. Krea 2 Turbo with my 4-step LoRA renders a small image in 4 steps (the 3-step LoRA is wired beside it, and Krea 2 Turbo also runs without a LoRA at its own 8 steps, see Steps and LoRAs).
2. NVIDIA's PiD (Pixel Diffusion Decoder) turns that small latent straight into an image 4× its width and height, in 4 more steps. There is no VAE decode and no separate upscaler: PiD draws the full-size image itself, guided by the small latent and the prompt.

The default is 512×896 → 2048×3584; any size works (see Sizes).

Why it is fast: Krea 2 renders at a quarter of the output's width and height, so it handles 1/16 of the pixels (at the default, 1,792 image tokens instead of 28,672); PiD, a much smaller decoder, does the full-size work.

Files

File	Use
Krea2Turbo-PiD-SuperFast-HighRes-ComfyUI-local.json	your own ComfyUI
Krea2Turbo-PiD-SuperFast-HighRes-ComfyUI-cloud.json	Comfy Cloud
krea2-pid-generate.py	the same two stages in Python (diffusers + NVIDIA's PiD code), see Python
requirements.txt	the Python packages for krea2-pid-generate.py and PiD's code: pip install -r requirements.txt
examples/	100 example images and their prompts, see Examples

Drag a workflow into ComfyUI, set the size and steps in 0 · Size and steps (default 512×896, 4 steps), write a prompt in 1 · Prompt, press Run. The draft preview shows the Krea 2 render; the saved image is PiD's 4× output. Both workflows use core ComfyUI nodes only.

Steps and LoRAs

Three ways to render; the default is the 4-step LoRA at 4 steps. Krea 2 renders at a quarter of the output's width and height, so each of its steps is quick: a step at 512×896 works on 1,792 image tokens, 1/16 of the 28,672 of a 2048×3584 render. A step or two more adds little time: with the Python script on an RTX 5090, Krea 2 Turbo's own 8 steps took 0.4 s longer than the 4-step LoRA's 4 (most of the Krea 2 stage is loading the model onto the GPU), and PiD's decode to 2048×3584 took 3.3 s.

Option	In 2 · Krea 2 Turbo + LoRA	Advised steps
4-step LoRA (default)	4-step LoRA on, 3-step LoRA bypassed	4-5
3-step LoRA	3-step LoRA on, 4-step LoRA bypassed	3-4
Krea 2 Turbo, no LoRA	both LoRAs bypassed	8

Select a node and press Ctrl+B to toggle its bypass, and set Steps in 0 · Size and steps. Use one LoRA at a time. The 3-step LoRA is still in training: if a picture is not what you want, try 4 steps, or go back to the 4-step LoRA. Past the advised steps a step LoRA starts to over-sharpen skin into a crackled look.

Local: requirements and models

A ComfyUI recent enough to include Krea 2 and PiD support (update ComfyUI first). Download these files into your ComfyUI models/ folders:

File	Folder	Size	Source
krea2_turbo_bf16.safetensors	diffusion_models/	26.3 GB	Comfy-Org/Krea-2
qwen3vl_4b_bf16.safetensors	text_encoders/	8.9 GB	Comfy-Org/Krea-2
qwen_image_vae.safetensors	vae/	0.3 GB	Comfy-Org/Krea-2
krea2_turbo_4step_rank_64_lora_comfyui.safetensors	loras/	0.4 GB	lvladikov/Krea2-Turbo-Distill-4step-LoRA
krea2_turbo_3step_rank_64_lora_comfyui.safetensors (only if you use it)	loras/	0.4 GB	lvladikov/Krea2-Turbo-Distill-3step-LoRA
pid_1.5_qwenimage_1024_to_4096_4step_bf16.safetensors	diffusion_models/	2.8 GB	Comfy-Org/PixelDiT
gemma_2_2b_it_elm_bf16.safetensors	text_encoders/	5.2 GB	Comfy-Org/PixelDiT

Smaller alternatives from the same repositories also load: krea2_turbo_fp8_scaled (13.1 GB), pid_1.5_qwenimage_1024_to_4096_4step_int8_convrot (1.6 GB), gemma_2_2b_it_elm_fp8_scaled (2.6 GB). The nvfp4 Krea 2 build softens the LoRAs' fine texture; prefer bf16 or fp8. PiD needs no VAE file: it uses ComfyUI's built-in pixel_space VAE. The Qwen-Image VAE above only decodes the small draft preview (Krea 2 uses the Qwen-Image VAE).

Comfy Cloud

Comfy Cloud hosts Krea 2 Turbo, its text encoder, the Qwen-Image VAE and PiD's Gemma text encoder. Import these files into your Comfy Cloud model library, once, from their URLs:

1. The 4-step LoRA: https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA/resolve/main/krea2_turbo_4step_rank_64_lora_comfyui.safetensors
2. The PiD decoder: https://huggingface.co/Comfy-Org/PixelDiT/resolve/main/diffusion_models/pid_1.5_qwenimage_1024_to_4096_4step_bf16.safetensors
3. Only if you use it, the 3-step LoRA: https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA/resolve/main/krea2_turbo_3step_rank_64_lora_comfyui.safetensors

Then open Krea2Turbo-PiD-SuperFast-HighRes-ComfyUI-cloud.json. Comfy Cloud names an import <user>__<repo>__<file>, and the workflow uses those names (lvladikov__Krea2-Turbo-Distill-4step-LoRA__krea2_turbo_4step_rank_64_lora_comfyui.safetensors in the 4-step LoRA loader, Comfy-Org__PixelDiT__pid_1.5_qwenimage_1024_to_4096_4step_bf16.safetensors in Load PiD); if yours differ, pick them there.

On Comfy Cloud's RTX PRO 6000 an image at the defaults took about 1.4 credits when 10 prompts share a job (the 100 examples: 142 credits), and about 5.8 credits for a single run that loads the models (21.1 GPU-seconds).

Settings

Stage	Steps	CFG	Sampler / scheduler	Other
Krea 2 Turbo + 4-step LoRA (on)	4 (advised 4-5)	1	euler / simple	no negative prompt
Krea 2 Turbo + 3-step LoRA (bypassed)	3 (advised 3-4)	1	euler / simple	no negative prompt
Krea 2 Turbo, no LoRA (both bypassed)	8	1	euler / simple	no negative prompt
PiD	4	1	lcm / simple	latent_format = qwenimage (Krea 2 uses the Qwen-Image VAE), degrade_sigma = 0

Sizes

Set Width and Height in 0 · Size and steps: they are the Krea 2 render, in multiples of 16. PiD's output is computed from them (4× each side), so nothing else needs changing.

Krea 2 render	PiD output
512×896 (default)	2048×3584
896×512	3584×2048
768×768	3072×3072
576×1024	2304×4096
1024×576	4096×2304
512×512	2048×2048

The decoder, PiD's 2K→4K v1.5, is trained for outputs from 2048 to 4096 px, so keep the Krea 2 render at 512 px or more on its long side; smaller renders decode soft.

Prompts

Krea 2 reads full scene captions: long, specific prompts (subject, setting, light, materials, style) work better than short tags. The same prompt goes to both Krea 2 and PiD.

Python (diffusers)

krea2-pid-generate.py runs the same two stages without ComfyUI, on an NVIDIA (CUDA) GPU (PiD's code targets CUDA; on a Mac, use the workflows). Krea 2 Turbo's repository is gated: accept its license on its page and log in with hf auth login before the first run.

pip install -r requirements.txt
hf auth login
git clone https://github.com/nv-tlabs/PiD
cd PiD && hf download nvidia/PiD --local-dir . \
    --include "checkpoints/PiD_v1pt5_res2kto4k_sr4x_official_qwenimage_distill_4step/*" \
    --include "checkpoints/QwenImage_VAE_2d.pth" && cd ..

python krea2-pid-generate.py --prompt "A close-up portrait of ..." --out image.png
python krea2-pid-generate.py --prompt "..." --width 896 --height 512 --steps 5      # -> 3584x2048
python krea2-pid-generate.py --prompt "..." --lora 3step --steps 3                 # the 3-step LoRA
python krea2-pid-generate.py --prompt "..." --lora none                            # Krea 2 Turbo itself, 8 steps

requirements.txt lists every package the script and PiD's code import. Krea 2 Turbo and the LoRA download from their repositories on the first run. --width / --height set the Krea 2 render (default 512×896, multiples of 16); PiD outputs 4× each side. --lora picks the 4-step (default) or the 3-step LoRA, or none for Krea 2 Turbo itself (8 steps); --steps overrides the step count; --offload moves Krea 2 between the CPU and the GPU for GPUs with less memory. On an RTX 5090 (32 GB) with --offload, a run took about 75 s from start to saved image, model loading included: Krea 2's stage 34 s (the same at 4 and 8 steps), PiD's decode 3.3 s.

Examples

examples/ holds 100 images made with the Comfy Cloud workflow at its defaults, from the same 100 prompts as the MOSS-Image × PiD examples: 50 portraits and 50 prompts drawn at random from Lakonik/t2i-prompts-3m, each rendered by Krea 2 Turbo + the 4-step LoRA at 512×896 (4 steps, seed 4242) and decoded by PiD to 2048×3584 (4 steps). Its README shows all 100 as thumbnails that open the full images; examples/prompts.json lists each image's prompt and seed, and a JSON file beside each image holds its settings.

Licenses

Component	License
Krea 2 Turbo (Krea) and the Comfy-Org repack	Krea 2 Community License Agreement
4-step LoRA and 3-step LoRA	Krea 2 Community License Agreement (Derivatives of Krea 2 Turbo)
PiD by NVIDIA — the Comfy-Org/PixelDiT files used here	NVIDIA Source Code License v1 (nsclv1), not Apache-2.0: read its terms before any use beyond them
These workflows, the script and this README	Apache-2.0

The models are not included here: each workflow downloads them from their own repositories, under their own licenses. Krea 2 Turbo's agreement and its Acceptable Use Policy apply to everything rendered with it.

Credits

Krea for Krea 2, NVIDIA for PiD, and Comfy-Org for ComfyUI's Krea 2 and PiD support and the repacked files.