Midjourney alternative open source: self-hosting and costs. Which open image generators you can run on your own computer, what their licenses let you sell, and when a hosted plan costs less trouble. Stable Diffusion 3.5 Medium is my pick for a Midjourney alternative you run yourself, because it needs a graphics card with only 12 GB of VRAM. If you would rather subscribe, Leonardo AI allows commercial use for £90.60 a year on its 12-month plan. Image Genius's own comparison data, checked in October 2026, puts Leonardo's 12-month plan at £7.55 a month and Open Art's at £9.81. Stable Diffusion 3.5 runs on home hardware. Stability AI shipped Stable Diffusion 3.5 in three versions: Large, Large Turbo and Medium, which followed on October 29. Stability says all of them run on consumer hardware and adapt easily to your own needs. The weights download from Hugging Face and GitHub. Large carries 8 billion parameters and aims at professional work at 1 megapixel. Medium is far smaller at 2.5 billion parameters. Civitai reports a maximum prompt context of 256 tokens for 3.5, half what Flux takes. At launch only ComfyUI supported 3.5 among the common local interfaces. Stability itself concedes that 3.5 gives up some consistency and polish in its images in return for being easier to adapt. By Stability's own ELO scoring, the 3.5 family still follows prompts better than Flux.1 Dev. Stability's Community License suits small sellers. Stability's Community License costs nothing for research, personal or commercial work until your revenue tops $1M a year. Above that line a paid Enterprise license applies, with implementation support, customization and professional services. Stability puts no cap on how many images or videos you create. You keep ownership of everything you generate. Stability also promises it will never ask you to delete images, fine-tunes or products built on them, even if you never pay. Businesses earning under the line self-report on a short form. They also register with Stability before any commercial use starts. If you pass the weights or a derivative on to others, a "Powered by Stability AI" notice has to appear on a related website, interface, blog post, about page or product documentation. Stability's license also forbids using the weights or their output to build or improve a foundational generative model. Since 31 July 2025, Stability's Acceptable Use Policy has also barred sexually explicit content from its Core Models. FLUX.1 schnell gives Apache 2.0 freedom. Black Forest Labs has released two of its text-to-image models with open weights. FLUX.1 [schnell], one of that pair, turns written descriptions into images with 12 billion parameters. Black Forest Labs put it out under Apache 2.0, which allows personal, scientific and commercial use. Schnell comes from the Dev variant by timestep distillation. Inferless's docs call it the fastest model in the FLUX.1 suite and say it suits local development and personal projects. It handles a wide spread of aspect ratios and resolutions. Its image-variations companion, FLUX.1 Redux [schnell], is also released under Apache 2.0, with weights and code on GitHub. FLUX.1 dev keeps local images non-commercial. FLUX.1 [dev] works the same way, as a rectified flow transformer that turns text into images. Its own Non-Commercial License covers it instead. Black Forest Labs' terms rule out commercial and production work with dev, for the model and for anything it makes. A business has to request a license, and Black Forest Labs may charge a fee, a royalty or a revenue share. Images made with FLUX.1 models on Replicate may be used commercially, Replicate states, though images made locally from downloaded weights may not. Replicate bills FLUX.1 [dev] once for each image at $0.025, so $1 covers 40. Black Forest Labs claims no ownership of what you make, but you cannot use those images to train, fine-tune or distill a competing model. Anyone redistributing the model has to include an attribution notice naming Black Forest Labs and the license. Use must also comply with the EU AI Act and GDPR. Hosted plans cost a fixed yearly sum. A year of Open Art on the same term comes to £117.72. That makes Open Art £27.12 a year dearer, as the Leonardo AI vs Open Art comparison sets out. Leonardo AI puts Flux Dev among its models, alongside Lucid Origin, Lucid Realism, Phoenix 1.0 and Phoenix 0.9. Open Art's first paid tier offers no API access. Stable Diffusion for sellers, schnell for no strings. Run Stable Diffusion 3.5 Medium if you own a mid-range graphics card and plan to sell your images, since its license is free below the revenue line. Civitai found that LoRAs trained on Large did not work with its weights. Choose FLUX.1 [schnell] if you want Apache 2.0 and its plain permission for commercial use. Schnell carries far more parameters than Medium does. Pick Leonardo AI if you want Flux Dev images you can sell, since it includes Flux Dev among its models and allows commercial use. Its cheapest plan offers no API access. Open Art serves you only for personal work, since it grants no commercial license.

Comfy-Org's FLUX.1-dev single-file package hits 1 million downloads. Comfy-Org bundles FLUX.1-dev and its control variants into single-file safetensors with built-in text encoders, cutting setup friction for ComfyUI users under 24GB VRAM. · 4 hrs ago Read 2 min * Comfy-Org/Flux-1-Dev repackages FLUX.1-dev into single-file safetensors for ComfyUI, with text encoders bundled in. * FP8 checkpoint is 17.2GB and runs on 8GB+ VRAM; full version is 23.8GB. * Repo also includes repackaged Canny, Depth (NVFP4), Fill, and Redux control variants. * Load with the standard Load Checkpoint node, no DualCLIPLoader wiring needed. * Ready-made workflow templates cover inpaint, outpaint, Canny, depth LoRA, and reference generation. * Over 1M downloads; non-commercial Flux.1-dev license still applies to all files. Comfy-Org's single-file packaging of Black Forest Labs' FLUX.1-dev has passed one million downloads on Hugging Face. The repository turns a four-component ComfyUI setup into two drop-in checkpoint options, allowing developers to load the model, VAE, and text encoders through one standard node. Fewer artifacts make workflows easier to share and reduce failures caused by missing files, incorrect paths, or mismatched precision settings. Black Forest Labs released FLUX.1-dev as a 12-billion-parameter, guidance-distilled image model. ComfyUI's split-file workflow treats its diffusion transformer, VAE, CLIP-L encoder, and T5-XXL encoder as separate assets. The ComfyUI team maintains the repository, repackaging the source weights for ComfyUI's loader conventions without retraining the model or adding network layers. Four loaders become one. Each file in checkpoints/ bundles the diffusion model, VAE, CLIP-L, and T5-XXL components. ComfyUI's Load Checkpoint node exposes the model, text-encoder, and VAE outputs required by the rest of the graph. This replaces the separate Load Diffusion Model, DualCLIPLoader, and Load VAE nodes used by the modular workflow. | Checkpoint | Size | Best fit | Main caveat | | flux1-dev.safetensors | 23.8 GB | Systems with more memory and storage headroom | Larger downloads and higher resource use | | flux1-dev-fp8.safetensors | 17.2 GB | Consumer GPUs and lower-memory systems | Quantization can introduce small output differences | File size does not translate directly into a fixed VRAM requirement. ComfyUI can offload model components to system memory, so some 8 GB GPUs can run the FP8 checkpoint with low-VRAM settings. Resolution, batch size, attention implementation, active nodes, and available system RAM all affect whether a workflow completes and how quickly it runs. Heavy offloading can make generation substantially slower. Pro article This story is for Pro members. You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.