Flux

Start with Flux.1-dev via ComfyUI on RTX 4090 24 GB — Flux.1-dev is the best open-weight text-to-image model for photorealism and prompt adherence. The dev variant uses a guidance-distilled 50-step DiT (diffusion transformer) generating 1024×1024 images in ~8 seconds on RTX 4090. For lower VRAM (<12 GB), Flux.1-schnell (4-step distilled) generates in ~2 seconds on RTX 3060 12GB with acceptable quality — good for rapid iteration. Skip Flux.1-pro (Black Forest Labs API, closed-weight). The Flux architecture is a DiT with dual CLIP + T5-XXL text encoders — the T5-XXL alone requires ~11 GB VRAM at FP16. For 24 GB cards, load T5 in FP16; for <16 GB cards, use T5 Q8 via bitsandbytes. Flux uses the Flux-1-dev non-commercial license — commercial use requires a Black Forest Labs agreement.

For single-user generation: ComfyUI with Flux.1-dev FP16 on RTX 4090 24 GB. The ComfyUI Flux workflow: load T5-XXL FP16 text encoder (11 GB), load Flux DiT model (17 GB FP16), load VAE (~330 MB). Total: ~30 GB at FP32 attention — use --fp8_e4m3fn attention flag to fit on 24 GB. For maximum speed: Flux.1-schnell with TensorRT engine on RTX 5090 32 GB — ~0.8 seconds per 1024×1024 image (4-step distilled). For server/batching: ComfyUI in API mode behind a job queue — queue-based batching processes one image per forward pass; use multi-GPU for concurrent generation. For LoRA fine-tuning: Flux.1-dev supports LoRA via Kohya SS or OneTrainer — ~12 GB VRAM for rank-16 LoRA training on RTX 4090. See GPU buyer guide.