PromptZone - AI Prompts, Guides and Tools for Builders

Cover image for FLUX.1 Krea dev: ComfyUI Configuration and Sampling Guide
Wiebke Chakraborty
Wiebke Chakraborty

Posted on Edited on

FLUX.1 Krea dev: ComfyUI Configuration and Sampling Guide

To configure FLUX.1 Krea dev in ComfyUI, load its diffusion weights, CLIP-L and T5 text encoders, and VAE in the official workflow. The text-to-image model was developed by Black Forest Labs and Krea and is compatible with the FLUX.1 dev architecture. 1, 2, 3

For acquisition and access conditions, see the Krea download guide.

What are the key facts about FLUX.1 Krea dev?

Fact Verified detail
Developer Black Forest Labs and Krea. 1
Released July 31, 2025. 1
Type Guidance-distilled rectified flow transformer for text-to-image generation. 2
Size or parameters 12 billion parameters. 2
License and access Downloadable weights under the FLUX dev non-commercial license; separate commercial licensing is available. 1, 2
Where it runs Local ComfyUI or Diffusers installations; Krea also provides a reference inference repository. 2, 3, 4

What does FLUX.1 Krea dev add to a ComfyUI workflow?

BFL describes Krea dev as a photographic model trained toward a particular aesthetic, with attention to reducing oversaturated textures. That is a useful reason to evaluate it for natural-looking scenes. 1

Architectural compatibility makes it relevant when you already have a FLUX.1 dev workflow. The model card describes Krea dev as a replacement for the original model in systems that support that architecture. 2

ComfyUI's official tutorial provides a complete workflow and identifies its separate diffusion model, text encoders, and VAE. This gives you a documented starting point before introducing additional components. 3

What limits affect Krea dev configuration?

The model card reports that prompt following can fail and depends on prompting style. It also says the model cannot supply reliable factual information and may reproduce social biases. 2

Consequently, inspect images that depict a specific object or situation against a reference. Attractive surface detail should not decide whether a product shape, written label, or requested spatial relationship is correct.

ComfyUI documents both original and reduced-precision model options and cautions that limited-memory devices may struggle. Its tutorial does not establish a universal generation time for a consumer GPU. 3

Record your own loading and generation times separately. Include your model file, precision, image dimensions, and workflow settings so that a later measurement describes the same job.

Krea's recommended resolution range is an inference recommendation, not proof of an absolute output limit. Likewise, its suggested guidance and step ranges describe a starting configuration rather than a guaranteed optimum. 4

The downloadable checkpoint has conditions on model use. Resolve those through the linked license and commercial access route before treating a working local setup as a production deployment. 1, 2

How do you configure FLUX.1 Krea dev in ComfyUI?

Start with an updated ComfyUI installation and import the workflow image or JSON linked in the official Krea tutorial. The ComfyUI pillar explains the surrounding workflow concepts. 3

  1. Put the original flux1-krea-dev.safetensors, or the tutorial's scaled FP8 version, in models/diffusion_models. Select that file in Load Diffusion Model. 3
  2. Put clip_l.safetensors and the chosen T5 encoder in models/text_encoders. Select both in DualCLIPLoader. 3
  3. Put ae.safetensors in models/vae and select it in Load VAE. Keep the imported workflow's connections intact for the first run. 3
  4. Enter your scene description and queue the workflow. If a required node is missing, check your ComfyUI version and startup import errors against the tutorial's troubleshooting notes. 3

Krea's reference repository recommends 28–32 inference steps, guidance between 3.5 and 5.0, and resolutions between 1024 and 1280 pixels. Use these as published starting ranges when evaluating your configuration. 4

For a separate Python baseline, the model card supports FluxPipeline and CPU offloading. Install the dependencies required by your PyTorch environment and authenticate for the gated model before running this example. 2

import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-Krea-dev", torch_dtype=torch.bfloat16
)
pipe.enable_model_cpu_offload()
pipe(
    "A ceramic bowl on folded linen, side window light, quiet colors",
    width=1024, height=1024, num_inference_steps=30, guidance_scale=4.5,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0].save("krea-baseline.png")
Enter fullscreen mode Exit fullscreen mode

This example uses the model card's pipeline and a setting inside Krea's recommended ranges. It is a baseline recipe, not a reported benchmark or a claim that this exact configuration fits every machine. 2, 4

Compare settings deliberately. Keep your prompt and image dimensions unchanged, then test another guidance value. Record whether the requested materials and composition improve before considering another adjustment.

Next, inspect another seed with the selected settings. A configuration chosen from one pleasing result still needs to suit the other scenes you intend to generate.

Save the workflow beside your outputs, using filenames that identify the experiment. A short note such as “better bowl texture, incorrect cloth color” is more useful than an unexplained quality score.

How do Krea dev sampling settings compare with dev and schnell?

Model Configuration distinction
FLUX.1 Krea dev Uses the FLUX.1 dev architecture with Krea's aesthetic training and published sampling recommendations. 1, 4
FLUX.1 dev Original guidance-distilled text-to-image model supported by FluxPipeline. 5
FLUX.1 schnell Distilled for generation in one to four steps; its card uses zero guidance. 6

Avoid treating identical step counts as a fair comparison when the models have different sampling recipes. Report the settings actually used and judge the usable outputs, including those that needed correction.

What should you check when configuring Krea dev?

Is FLUX.1 Krea dev a ComfyUI extension?

FLUX.1 Krea dev is a model checkpoint with a supported ComfyUI workflow. The official tutorial loads it through the diffusion-model loader alongside text encoders and a VAE. 2, 3

Which guidance value should I start with?

For FLUX.1 Krea dev, Krea recommends guidance of 3.5–5.0, and the model card demonstrates 4.5. Start there and judge whether another value better meets your scene requirements. 2, 4

Can I reuse FLUX.1 dev components?

FLUX.1 Krea dev is compatible with FLUX.1 dev, and ComfyUI identifies their shared text encoders and VAE. Check the filenames selected in your workflow instead of downloading another copy automatically. 2, 3

Does FP8 guarantee that it fits my GPU?

ComfyUI offers a scaled FP8 version of FLUX.1 Krea dev for lower-memory configurations. Its documentation still cautions about memory constraints, so test your complete workflow before relying on it. 3

Sources

Top comments (0)