Flux 1 doesn’t exist as a current recommendation anymore. Black Forest Labs, the company founded by the former Stable Diffusion researchers behind the original model, released Flux 2 in November 2025, and every Flux 1 variant, Pro, Dev, and Schnell, has since been superseded. If you’re setting up Flux for the first time in 2026, the model you actually want is Flux 2, and the setup process and prompting approach both differ meaningfully from guides still built around the older generation.
What Actually Changed From Flux 1 to Flux 2
Flux 2 isn’t a minor version bump. It adds multi-reference conditioning, the ability to reference up to ten images at once while keeping a character, product, or style consistent across every generation, a capability Flux 1 never had at all. Output resolution jumped to 4 megapixels for both generation and editing, text rendering became substantially more reliable (Flux 1 was notorious for garbled or “text soup” output on anything with legible words in it), and Black Forest Labs added hex color steering for precise, exact color matching rather than approximate prompted colors. Generation speed also improved dramatically, with the fastest current variant producing images in under a second on capable hardware.
The Current Flux 2 Lineup
Flux 2 ships as a family of models rather than a single release, and picking the right variant matters more than picking Flux 2 in the abstract.
Flux 2 [pro]
The flagship, proprietary API-only model, built for the highest editing consistency and strongest prompt-following of the family. This is the version most commercial and production creative work should default to when quality matters more than cost or local control.
Flux 2 [dev]
The open-weight, non-commercial-license variant aimed at photorealism and multi-reference composition work. It’s the closest thing to a direct research and hobbyist successor to Flux 1 Dev, available for self-hosting rather than API-only access, with the licensing restriction that comes with that trade.
Flux 2 [flex]
Specialized specifically for typography and fine detail preservation, the variant to reach for when a project’s success depends on legible text rendering, UI mockups, or precise small-detail work that the other variants aren’t optimized for.
Flux 2 [klein]
The local-first, consumer-hardware variant, replacing Flux 1 Schnell’s role in the lineup. It comes in two sizes: a 4-billion-parameter distilled version released under the fully permissive Apache 2.0 license, generating images in under a second and running on roughly 13GB of VRAM, putting it within reach of consumer GPUs like an RTX 3090 or 4070, and a larger 9-billion-parameter version under a non-commercial open-weight license for higher quality at a real hardware cost increase.
How to Actually Set Up Flux 2
Your setup path depends entirely on which variant fits your hardware and licensing needs, so decide that first rather than picking a setup method and working backward. For [pro], setup is minimal: it’s API-only, accessed through Black Forest Labs’ own endpoints or a cloud platform hosting it, meaning there’s no local installation at all, just API credentials and a request format to learn.
For self-hosting [dev], [flex], or [klein], the most common path runs through ComfyUI, the node-based interface that’s become the de facto standard for running Flux locally. Install ComfyUI, download the specific Flux 2 checkpoint matching your chosen variant from Hugging Face, and place it in ComfyUI’s model directory according to the standard folder structure. The [klein] 4B variant is the realistic starting point for most people setting this up for the first time, since its modest 13GB VRAM requirement and Apache 2.0 license mean no commercial-use restrictions and no need for high-end hardware to get a functional first result.
If ComfyUI’s node-based interface feels like more complexity than you want to manage, several cloud platforms now host Flux 2 directly inside a simpler interface, trading some configuration control for a faster path to your first generation, worth considering if your priority is getting usable output quickly rather than building a fully customized local pipeline.
Best Prompts and Techniques for Flux 2
Flux 2’s biggest prompting shift from Flux 1 is that vague, mood-based prompting undersells what the model can actually do. Because multi-reference conditioning and hex color steering are new, specific capabilities, prompts that take advantage of them directly produce meaningfully better results than prompts written the way you’d have written them for Flux 1 or a generic diffusion model.
For product or brand consistency work, structure a prompt around your reference images explicitly: describe what should carry over from each reference (the product itself, the background style, the lighting) rather than assuming the model will infer which elements matter. For text-heavy generations, posters, UI mockups, infographics, specify the exact text content in quotes within your prompt rather than describing it vaguely, since Flux 2’s improved text rendering specifically rewards prompts that give it precise strings to render rather than a general description of what the text should convey.
For color-critical work, branded content especially, use hex color steering directly rather than describing a color in natural language. “A crimson red background” leaves room for interpretation Flux 2 will fill in its own way. Specifying the actual hex value removes that ambiguity entirely, which matters considerably for anything that needs to match an existing brand palette exactly.
A Structured Prompt Template Worth Starting From
A reliable starting structure for Flux 2, regardless of variant, follows subject, action or state, environment, style, and technical specification in that order. Something like: “a ceramic coffee mug [subject], sitting on a reclaimed wood table [environment and state], morning light through a nearby window [lighting], photorealistic product photography style [style], shot at f/2.8 with shallow depth of field [technical specification].” That structure gives the model a clear hierarchy to work through rather than a single dense sentence where the important details compete for attention with less important ones.
For editing rather than pure generation, be explicit about what should stay unchanged, not just what should change. Flux 2’s editing strength depends on the model correctly identifying the boundary between what you want modified and what you want preserved, and stating that boundary directly in the prompt produces more reliable results than assuming the model will infer it from context.
Troubleshooting Common Setup Problems
The most frequent ComfyUI setup failure is a checkpoint placed in the wrong model subdirectory, which produces a “model not found” error even when the file downloaded correctly. Confirm the specific subfolder Flux 2 checkpoints expect, distinct from where Stable Diffusion checkpoints go, since ComfyUI organizes models by architecture and a misplaced file simply won’t appear in the model selection dropdown rather than throwing an obviously diagnostic error.
Out-of-memory errors when running [dev] or the [klein] 9B variant almost always mean the specific variant exceeds your available VRAM, not a configuration mistake. Dropping to the [klein] 4B variant, or enabling any available memory-efficient loading options in ComfyUI, usually resolves this faster than trying to force a larger model onto insufficient hardware. If generation completes but output looks visibly degraded or blurry compared to example images you’ve seen online, check that you’re running the checkpoint at its intended resolution and step count rather than settings copied from a different variant’s recommended configuration, since [klein]’s speed-optimized settings differ meaningfully from [dev]’s quality-focused defaults.
How Flux 2 Actually Compares to Midjourney and Other Competitors
Flux 2 isn’t positioning itself as a direct Midjourney competitor on aesthetic terms, and it’s worth being clear about that distinction before choosing between them. Midjourney remains the easier default for anyone who wants striking, stylized images fast with no engineering effort and no interest in self-hosting. Flux 2’s actual pitch is control and production-pipeline fit: reference-image consistency, reliable text rendering, and self-hosted deployment options for teams that need image generation to slot into an existing product or workflow rather than function as a standalone creative tool.
That distinction should drive your choice more than a general “which is better” comparison. If image generation is a feature inside something you’re building, a product, an app, an automated content pipeline, Flux 2’s consistency and self-hosting options make it the stronger foundation. If you’re generating images for their own sake and want the best-looking result with the least setup, Midjourney’s more polished, no-engineering-required experience is still the faster path to a satisfying result.
Common Mistakes When Moving From Flux 1 Prompts
The most common mistake for anyone with existing Flux 1 prompt habits is treating Flux 2 as a drop-in replacement rather than a model with genuinely new capabilities worth prompting for directly. A Flux 1 prompt built around a single reference image and a vague style description will still generate something in Flux 2, but it leaves the multi-reference and hex-color capabilities entirely unused, capabilities that are often the actual reason to prefer Flux 2 in the first place. Rewrite prompts to take advantage of what’s actually new rather than porting old habits forward unchanged.
A second common mistake is picking [pro] by default without checking whether [klein] or [dev] would meet the actual need at a fraction of the cost or with no licensing restriction at all. Not every project needs the flagship model’s editing consistency, and testing a smaller variant first before committing to API costs on [pro] is worth the extra ten minutes for anything that isn’t already a confirmed production requirement.
Common Questions About Flux 2
Is Flux 1 still worth using in 2026?
Not as a first choice. Every Flux 1 variant has been superseded by a corresponding Flux 2 model, and Black Forest Labs’ own guidance treats Flux 1 as a legacy generation rather than an actively recommended option for new projects.
Which Flux 2 variant should I start with if I have a normal consumer GPU?
Flux 2 [klein] in its 4-billion-parameter form, released under Apache 2.0 with no commercial restrictions and requiring only around 13GB of VRAM, runs on hardware like an RTX 3090 or 4070 and is the most accessible entry point into the current lineup.
Can I use Flux 2 for commercial work without paying for the API?
Yes, specifically with the [klein] 4B variant under its Apache 2.0 license, which permits unrestricted commercial use. The [dev] and [klein] 9B variants carry non-commercial open-weight licenses instead, so check the specific variant’s license before using its output commercially.
Why does my text rendering still look wrong in Flux 2?
Confirm you’re specifying exact text content in quotes within your prompt rather than describing it vaguely, and consider the [flex] variant specifically, since it’s built for typography and fine detail preservation where the other variants prioritize different strengths.
The Bottom Line
Any Flux setup or prompting guide still built around Flux 1 in 2026 is describing a superseded model family. Flux 2’s real advantages, multi-reference conditioning, reliable text rendering, and hex color steering, only pay off if your prompts and setup are built around them directly rather than carried over unchanged from an older generation’s habits.
References and Sources
MindStudio, “What Is FLUX 2 Pro? Black Forest Labs’ Next-Gen Image Model”: https://www.mindstudio.ai/blog/what-is-flux-2-pro
InnFactory, “Black Forest Labs FLUX: Versions, Benchmarks and EU Availability”: https://innfactory.ai/en/ai-models/flux/
Thunder Compute, “Flux ComfyUI: The Complete Guide (September 2026)”: https://www.thundercompute.com/blog/flux-comfyui-ai-image-generation
ModelsLab, “FLUX.1 vs FLUX.2: Migrate Your API Calls (2026)”: https://modelslab.com/blog/image-generation/flux1-vs-flux2-migration-guide-2026
FluxNote, “FLUX.2: Black Forest Labs’ New AI Image Model Explained (2026)”: https://fluxnote.io/guides/flux-2
About The Author
