Sign In

Krea 2 Ultimate 12GB — Krea 2 Turbo + Kroma 0.3.1 w4a8, Text to Image + Style Reference + style LoRAs

Download

1 variant available

Config Other

The_frizzy1_krea-2-ultimate-12gb_v2.1.0.json

99.69 KB

Verified:

Type
Workflows
Stats

11

Reviews

No reviews yet

Published

Nov 18, 2026

Base Model
Hash
AutoV27344FCA349

Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.

02_KU_kroma_t2i_20261013.png

☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1


Both official Krea 2 Turbo templates in one graph: text to image and image style reference, with the official style LoRAs and the optional prompt enhancer. Since v2.1.0 a Model switch adds Kroma, lodestones' Krea 2 fine-tune — same text encoder, VAE and Turbo settings, 8 steps, CFG 1. Only the model you pick is loaded.

What you need

✔ Custom nodes — None. Core ComfyUI only.
✔ ComfyUI version — 0.39 or newer. v2.0.0 ran on 0.37.4; the Kroma path was only tested on 0.39.2.
✔ VRAM — 12 GB — tested, not estimated. Peak 11.6 GB on Kroma, 11.3 GB on Krea 2 Turbo
✔ System RAM — 10.4 GB peak, on the style-reference run. On 16 GB it will swap.
✔ Tested on — RTX 3060 12 GB · 32 GB DDR4 · ComfyUI 0.39.2
✔ Licence — Workflow MIT · Krea 2 and Kroma: Krea 2 Community Licence

Models to download

Click a file name to download it. Put it in the folder shown, inside ComfyUI/models/.

Required — 3 files, 19.0 GB

diffusion_models/

File: krea2_turbo_int8_convrot.safetensors — 13.5 GB (12 GB (tested))
Download:
https://huggingface.co/Comfy-Org/Krea-2/resolve/main/diffusion_models/krea2_turbo_int8_convrot.safetensors

Put it here:
ComfyUI/models/diffusion_models/

text_encoders/

File: qwen3vl_4b_fp8_scaled.safetensors — 5.2 GB
Download:
https://huggingface.co/Comfy-Org/Krea-2/resolve/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors

Put it here:
ComfyUI/models/text_encoders/

vae/

File: qwen_image_vae.safetensors — 0.3 GB
Download:
https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors

Put it here:
ComfyUI/models/vae/

Style reference and style LoRAs — optional. All go in loras/.

Other official style LoRAs, same folder, about 0.5 GB each: krea2_dotmatrix · krea2_kidsdrawing · krea2_neondrip · krea2_rainywindow · krea2_retroanime · krea2_softwatercolor · krea2_sunsetblur · krea2_vintagetarot

Kroma — only if you set Model to Kroma. The text encoder and VAE above are shared, so this is the one new file.

diffusion_models/

File: krea_kroma_v0.3.1_turbo_opd_w4a8.safetensors — 8.8 GB (Kroma, 12 GB (tested))
Download:
https://huggingface.co/jakethessmake/Krea-Kroma-v0.3.1-Turbo-OPD-ComfyUI-Quants/resolve/main/krea_kroma_v0.3.1_turbo_opd_w4a8.safetensors

Put it here:
ComfyUI/models/diffusion_models/

File: krea_kroma_v0.3.1_turbo_opd_int8_convrot.safetensors — 14.1 GB
Download:
https://huggingface.co/jakethessmake/Krea-Kroma-v0.3.1-Turbo-OPD-ComfyUI-Quants/resolve/main/krea_kroma_v0.3.1_turbo_opd_int8_convrot.safetensors

Put it here:
ComfyUI/models/diffusion_models/

The original Kroma is bf16 at 26.3 GB, lodestones/Kroma. I did not try it. Kroma is a Krea 2 fine-tune, so the Krea 2 Community Licence applies.

Speed on my RTX 3060

Wall is the median, queue to finished, with the text encoder already loaded. Sampling is time in the sampler. Same prompts and seeds for both models, ComfyUI 0.39.2.

  • Kroma · text to image — 1 MP, 18 s wall, 16 s sampling, 11.6 GB peak vram, 5.5 GB peak ram, 3 runs

  • Krea 2 Turbo · text to image — 1 MP, 19 s wall, 16 s sampling, 11.3 GB peak vram, 6.8 GB peak ram, 4 runs

  • Kroma · text to image — 2K (4 MP), 1 min 23 s wall, 1 min 19 s sampling, 11.6 GB peak vram, 7.7 GB peak ram, 2 runs

  • Krea 2 Turbo · text to image — 2K (4 MP), 1 min 22 s wall, 1 min 18 s sampling, 11.3 GB peak vram, 9.4 GB peak ram, 2 runs

  • Kroma · style reference — 1 MP, 45 s wall, 40 s sampling, 11.3 GB peak vram, 10.4 GB peak ram, 1 runs

Kroma and Krea 2 Turbo come out within a second of each other at both sizes. The Model switch buys you a different look, not speed.

What did not work, and what to watch

  • The style-reference row is one run, not a median. Treat that 45 s as a single data point. There is no Krea 2 Turbo style-reference timing at all.

  • Style reference is the hungry one: 10.4 GB of system RAM on top of 11.3 GB of VRAM. Fine on 32 GB, it will swap on 16 GB.

  • 2K eats system RAM too — Krea 2 Turbo peaked at 9.4 GB there.

  • The tested Kroma file is the w4a8 quant at 8.8 GB. The bf16 original was not tried, so it has no numbers here.

  • If your models live in sub-folders like diffusion_models/Kroma/, the loaders show red until you re-pick them once.

Get the workflow

More videos on this stuff: https://www.youtube.com/playlist?list=PLG_da2oqTx7k