AI · Fundamentals
Model types overview
SDXL, Flux.1 dev, Flux.1 Krea dev, Flux.2 dev, Flux.2 Klein — which model when.
The last article clarified which files a model is made of: diffusion model, text encoder, and VAE — bundled into one checkpoint with SDXL, shipped as separate files with Flux — plus LoRA and ControlNet as additions. This article addresses the question that comes up first in practice: which model do you actually load?
There isn't "the one" image generation model — there are several fundamentally different architectures that vary considerably in size, speed, quality, license, and hardware requirements. The choice isn't purely a matter of taste; it depends on what you have available and what you're trying to achieve.
This overview is deliberately wide-ranging: from models that run on a MacBook Air to models that push even capable hardware to its limits. That reflects this site's audience — from apprentices on a limited budget to professional production setups with the gear to match.
The models compared
SDXL
The most established and accessible entry point. SDXL runs stably with comparatively modest memory requirements — even a MacBook Air with 16 GB of RAM can handle it. Image quality is solid, and the ControlNet ecosystem around SDXL is mature and well documented, since the model has been in broad use for a while now.
The trade-off: SDXL doesn't reach the photorealistic level or prompt fidelity of newer architectures. For fast iteration, learning the fundamentals, and workflows with limited hardware, it's still the most solid choice.
Flux.1 dev
The next step up in quality and prompt understanding, while hardware demands remain reasonable. Flux.1 dev is guidance-distilled — a training technique that makes the model more efficient and changes how it's steered (more on that in the Flux.1 dev setup article). In practice, that means noticeably better results than SDXL without generation time exploding.
ControlNet Canny is available for Flux.1 dev and works reliably. That makes dev RAY-L-capable and the neutral base from which Krea can be positioned.
Flux.1 Krea dev
A sibling of Flux.1 dev: it sits on the same guidance-distilled base (flux-dev-raw), onto which Black Forest Labs and Krea ran a pure aesthetic post-training — with the stated goal of producing more realistic, less oversaturated images. Pipeline, encoders, and node structure are identical to Flux.1 dev; ControlNet Canny works just as reliably and is equally suited to RAY-L.
The difference is purely in the look: Krea dev has a distinctive, markedly photographic, "filmic" character instead of the smooth AI middle — currently my first choice for a photographic look, and therefore the recommendation for RAY-L. More on this in the Flux.1 Krea dev setup article.
Flux.2 dev
A considerably larger and more capable model than Flux.1 dev — the parameter count sits well above it, and the model combines image synthesis with its own language-understanding system for noticeably more accurate prompt translation, better consistency across multiple reference images, and higher resolution.
The price for that is compute time: on a MacBook Pro with an adapted setup, Flux.2 dev runs stably, but with significantly longer generation times than Flux.1 dev — on the order of 40 minutes per image, instead of seconds to a few minutes. That's unsuited to fast iteration, but entirely workable for deliberate, final results.
Currently relevant for RAY-L: ControlNet Canny does exist for Flux.2 dev — so the RAY-L path would be technically open. In practice it fails on render time: on a MacBook Pro with 64 GB the model runs together with its encoder and VAE, but at around 40 minutes per image that's no iterative workflow.
Flux.2 Klein — 4B and 9B
The small, fast members of the Flux.2 family — considerably quicker than SDXL and Flux.1 dev (own measurement: Klein 4B ~36 s versus SDXL ~143 s at 1024², warm). Both are step-distilled 4-step models; the difference is size: 4B for fast iteration and weaker hardware, 9B for more fine detail — in the image the two sit close together, 9B only visibly cleaner at strong zoom. They diverge on licensing: 4B under Apache 2.0 (free, including commercial use), 9B under the FLUX Non-Commercial License.
Currently relevant for RAY-L: no ComfyUI-ready ControlNet Canny exists for Klein — neither for 4B nor for 9B, and as of today likely none is coming. Unlike Flux.2 dev, where Canny exists and only render time stands in the way, Klein lacks the connection entirely.
Questions to decide by, not just spec sheets
Instead of a pure spec table, these questions are more helpful when choosing a model:
What can your hardware handle? SDXL and the small Klein variants run on modest setups, Flux.1 dev and Krea dev need a bit more, Flux.2 dev demands potent hardware and patience.
Do you need fast iteration or a single, high-quality result? For many test runs in a short time, smaller, faster models are the better fit. For a final, single image, a slower, more capable model is worth it too.
Is ControlNet Canny necessary for your workflow? When Blender geometry is meant to steer the composition — as in the RAY-L workflow — the availability of ControlNet Canny for a given model is currently the decisive selection criterion. As of today that's the case for SDXL, Flux.1 dev, and Krea dev; Flux.2 dev has Canny but is ruled out by render time, and Flux.2 Klein lacks it entirely.
Which license fits your plans? More on that in the next article — licensing models differ considerably between models, and "open weights" doesn't automatically mean unrestricted commercial use.
Summary
There's no single right answer to "which model should I use" — only the right answer for your hardware, your time budget, and your workflow. SDXL as an accessible, fast entry point with a mature ControlNet ecosystem. Flux.1 Krea dev as the current RAY-L recommendation, with a photographic look on the identical pipeline of Flux.1 dev, which as the more neutral base remains equally RAY-L-capable. Flux.2 dev as an outlook on considerably higher quality — with ControlNet, but only with a lot of patience (around 40 minutes per image). Flux.2 Klein as a fast, small alternative — 4B permissive under Apache 2.0, 9B more detail-precise, but both without ControlNet Canny.
For local use with ComfyUI, all the models mentioned are already available. On Windows with Nvidia hardware — the platform typical for these workflows — SDXL, Flux.1 dev, Krea dev, Flux.2 dev, and Flux.2 Klein run without special restrictions. One restriction concerns Apple Silicon (Mac): Flux.2 dev runs there only via a specially adapted setup path, and fp8 isn't computed on Apple Silicon at all — there it's fp16 or GGUF (details on the Krea page).
The next article clarifies what "open weights" actually means in licensing terms — a point that's often underestimated when choosing a model.