Image Generation AI Tools

Text prompts and reference images go in, raster output comes out, with no hosted API metering each render. The people running these locally want something an API makes awkward: high volume batches, fine-tuning on private data, conditioning on pose or depth maps. The field splits roughly three ways in 2026. The older UNet lineage around Stable Diffusion XL is outclassed on raw quality but carries an enormous fine-tune and LoRA ecosystem, including Juggernaut XL, RealVisXL, and Illustrious-XL, and it fits in 8 to 12 GB of VRAM. The newer diffusion transformers, FLUX.1 and FLUX.2, Qwen-Image, Sana, and Z-Image, follow prompts far more closely and render legible text, at the cost of much heavier weights. Distilled variants such as SDXL-Turbo and Latent Consistency Models sit across both camps, trading fidelity for one to eight step sampling. ComfyUI is the sensible first runner, since new models ship a node graph for it first; Fooocus is the lower friction option, and Diffusers is the right answer when the work belongs in a script. Pair either with Stable Diffusion XL to learn the mechanics cheaply, move to FLUX.1 once the pipeline is stable, and add ControlNet only when prompt text stops being enough. The trap here is licensing rather than capability. FLUX.1 dev weights are non-commercial, several Stability releases carry revenue-triggered community terms, and popular fine-tunes quietly inherit whatever their base model allowed. Chroma exists partly in response, an Apache-2.0 rebuild from FLUX.1-schnell. Check the base license and the VRAM floor before committing.