You want an image. ComfyUI connects the pieces that create it; your coding agent helps inspect, configure, and run those pieces. Start with one working model and one simple workflow. Add a style adapter or structural guidance after you know the baseline works.
For this learning path, start with Krea 2 Turbo, then try a FLUX model from Civitai using the exact version's recommended workflow. The point is to understand what you are connecting, not to collect the largest folder of models.
Family, variant, and file are different things
The family identifies a model architecture and ecosystem. A variant changes its purpose or behavior. A file is the particular set of weights you download. A community fine-tune can have its own name while still belonging to an underlying family.
Krea 2 Turbo is the distilled generation variant; Krea 2 RAW is the base intended for training and adaptation. The official ComfyUI guide provides the matching template and companion-model instructions. Follow that template's settings rather than copying settings from a different family. Krea 2 in ComfyUI.
FLUX.1 Krea Dev is a separate model from Krea 2. It belongs to the FLUX.1 Dev architecture. The shared Krea name does not make its workflow interchangeable with Krea 2. FLUX.1 Krea Dev.
Vibes for this guide
Use the materials that help with your next step.
FLUX includes different generations and variants. Record the exact base shown by the publisher or Civitai version page, such as FLUX.1 Dev, rather than just writing “Flux.” A file advertised for another generation, variant, or parameter size needs an explicit compatibility check. BFL model families.
What is a checkpoint?
A checkpoint is saved model weights. Some downloads package several components together; others provide the generation model separately from its text encoder and VAE. In practical use, ask which components the download actually contains and which loader the workflow expects. A .safetensors filename alone does not answer that question.
The text encoder converts your prompt into information the generation model can use. The VAE connects the model's latent representation with visible image pixels. Use the companion files required by that model and workflow. ComfyUI documents the distinction between model files, their folders, and loader nodes in its model guide.
What is a LoRA?
A LoRA is a smaller learned adaptation applied to a base model. It can influence a style, subject, or concept. It does not replace the base model, and it must be compatible with it. Its model page may specify trigger words and recommended strengths. LoRA introduction.
Try one LoRA first. Keep the prompt and seed fixed, render without it, then render with it at the author's suggested starting strength. Compare the result before stacking more adapters. If it overwhelms the image, reduce its influence or remove it; extra adapters are not automatically an improvement.
Krea documents a specific pairing: LoRAs trained on Krea 2 RAW can be applied to Turbo. That is a documented relationship within Krea 2, not permission to reuse arbitrary FLUX or SDXL adapters. Krea's release notes.
What is ControlNet?
ControlNet adds structural guidance, such as pose, edges, or depth. A preprocessor may turn a reference image into that guidance map; it is a separate part of the workflow. Match the control model to both the base model and the kind of map it expects. ControlNet introduction.
Use a LoRA when you want a learned style or concept. Consider structural guidance when you need a pose, silhouette, or layout to follow a reference. They can be combined in a compatible workflow, but learn them separately first. A reference-image editing feature is not automatically ControlNet.
Distilled, quantized, and fine-tuned
Distilled: adapted to run with a different, often shorter sampling procedure. Use its documented steps and guidance settings.
Quantized: represented at reduced numerical precision to reduce storage or memory needs. The exact format must be supported by the loader; quality and speed can differ.
Fine-tuned: further trained toward a particular behavior or domain. Check its base model and the author's intended workflow.
These describe different properties. A model can be both fine-tuned and quantized. None of the labels alone tells you its license, hardware requirements, or compatibility with an adapter. Hugging Face quantization concepts, Krea's model report.
Read a Civitai page before downloading
Record the exact version, resource type, base model, file format, required companion files, recommended workflow, and usage terms. Check whether the download is a checkpoint or a LoRA; the cover image is not enough. Keep the source URL and file digest with your project.
Model weights, adapters, and hosted generation can have different terms. Buying a VibeField workflow or skill does not grant rights to third-party weights. Read the specific model's terms before commercial use; do not assume every FLUX variant has the same license. BFL licensing, Krea 2 model card.
Ask your agent to inspect the stack
Prompt
Read the model pages and my ComfyUI workflow. Identify the exact base family and variant, then classify each file as generation weights, text encoder, VAE, LoRA, or ControlNet. Record source URLs and versions. Flag missing components and unsupported compatibility assumptions. Recommend a baseline with no adapters before downloading or running anything.
Your first experiment
Open the official Krea 2 Turbo template and follow its model setup.
Run one image without adapters and keep its workflow, prompt, seed, and output.
Change only the prompt and compare.
Add one documented compatible LoRA and compare again.
Try your chosen Civitai FLUX model with its own matching workflow. Do not swap its filename into the Krea graph.
Download the model comparison worksheet. For a ceramic-cup product image, judge whether the cup silhouette remains clear, the lighting matches the reference, and the output is usable at the intended size. Record the exact model revision and workflow before running.
Baseline A uses no adapter. Run B changes only one documented compatible LoRA and its recorded strength; retain prompt, seed, dimensions, sampler, and workflow version. For structural guidance, make a separate experiment with a matching control model and required preprocessing. Save the preprocessed map so a weak result can be traced to the input or conditioning strength. Do not compare a changed prompt, changed model, changed seed, and added adapter as if one caused the difference.
The worksheet records output paths and observations; it contains no invented generation results. vibefield-model-check validates declared family/source fields, not the weights themselves. A successful real run is still required. Download access, local compute cost, hosted charges, and model usage terms are separate questions.
Checked 2026-11-13: the publisher card for FLUX.1-Krea-dev identifies the FLUX.1 Dev non-commercial weight licence; Krea-2-Turbo identifies the Krea 2 Community License. Inspect the full terms for your use and every added resource. This guide grants no third-party model rights.
Questions
What models does Krea use?
Krea is also a hosted product, which should not be confused with a single model. Here we identify Krea 2 Turbo/RAW and the separate FLUX.1 Krea Dev by exact publisher card. Check the hosted product's current selection for a hosted task.
What is Flux AI used for?
Here FLUX means Black Forest Labs' image-model family, used for image generation and variant-specific image tasks. Other products use the words "Flux AI"; their capabilities and terms do not establish anything about these weights.
Is Flux AI free to use?
There is no single answer for every FLUX variant and access route. Check the exact model's licence and any hosted provider fee. Downloadable weights do not imply unrestricted commercial model use or free compute.
What's the difference between a LoRA and a checkpoint?
A checkpoint stores model weights. A LoRA stores a learned adaptation that needs a compatible base model. The extension alone cannot identify which role a file plays.
What does ControlNet do in Stable Diffusion?
It supplies structural conditioning, such as an edge or pose map, through a compatible control model. Preparing the correct conditioning image is a separate step from loading the model.
What models work with ControlNet?
Use the exact control model's documented base architecture, input type, and supported workflow. Do not assume an SD1.x, SDXL, FLUX, or Krea control component transfers between families.
Is Krea AI better than Midjourney?
Define the task and compare controlled outputs, access, runtime, and terms. This guide has no measured cross-product benchmark and does not infer one from a Google question.
Is Stable Diffusion obsolete?
The question's premise is not a research finding. Evaluate whether the exact model and workflow meet your task, quality, hardware, and usage requirements.
Pick the right image model the first time
Stop downloading files that do not fit: know which checkpoint, LoRA, VAE and ControlNet belong together for Krea 2 or FLUX.
schedulescheduleComing soon
Skill
Choose by exact family, variant, runtime and usage requirements.