How much VRAM do you need? SD 1.5, SDXL and Flux, honestly

These are working ranges, not guarantees; driver, front end and settings move them. Post your card and your numbers to correct them. SD 1.5 - Model file: about 2 GB in fp16 (pruned). - Runs on 4 GB cards with A1111's --medvram or --lowvram flags, or ComfyUI's automatic low-VRAM mode. 6-8 GB is comfortable for 512x768 with a ControlNet. - Runs on CPU, slowly; LCM makes CPU practical (see the LCM thread). SDXL - Base model file: about 6.9 GB in fp16. - 8 GB works with offloading flags; 12 GB is comfortable; refiner plus base together want more. - CPU is possible but painfully slow. Flux.1 (Black Forest Labs) - A 12-billion-parameter model; the full-precision weights are much larger than SDXL. - Quantised versions (fp8, and GGUF via the ComfyUI-GGUF custom nodes) bring it within reach of 8-12 GB cards at some quality and speed cost. - Licences differ: Flux.1 [schnell] is Apache 2.0; Flux.1 [dev] has a non-commercial licence. Read the model card before selling output. Training LoRA training on SD 1.5 is commonly done on 8 GB with memory-saving options (gradient checkpointing, 8-bit optimisers); SDXL LoRA training wants roughly 12-24 GB depending on settings. Free and cheap GPU options - Google Colab's free tier has offered NVIDIA T4 GPUs (16 GB) subject to availability and time limits. - Your own Modal account runs a GPU per request. Hathor Studio's engines page has the setup for both: https://hathor.soapbox.community/engines