Fal API Flux Dev with LoRA Image-to-Image

Take the LoRA-equipped Flux Dev node, add a starting image and a strength dial, and you get this one: image-to-image generation on FLUX.1 [dev], with up to five LoRAs stacked on top of it. It's the node you reach for when you want to restyle or reinterpret an existing image rather than generate one from nothing - a rough sketch turned into a finished piece, a photo pushed through a trained aesthetic, that kind of job.

The two inputs that make this an img2img node are image (required) and strength (0–1, default 0.8). strength is the standard denoise-amount tradeoff: near 1 gives the model more room to diverge from your input and lean on the prompt, near 0 keeps the result close to what you fed it and barely touches it. 0.8 is already fairly aggressive - if you're finding the output has drifted too far from your source image, that's the first number to pull down before you touch anything else.

Everything else mirrors the text-to-image LoRA node. prompt should read like a sentence rather than a tag list, since Dev's T5 encoder handles natural language properly. width/height default to 1024×1024 in steps of 8 (this sets the output canvas fal renders to, separate from your input image's own dimensions). num_inference_steps defaults to 28 (1–100), guidance_scale defaults to 3.5 (0.1–40, and the usual community advice is to drop it toward 1.8–2.0 if faces are coming out generic), num_images batches 1–4, and enable_safety_checker toggles fal's content filter. seed is optional. For LoRAs, lora_1 through lora_5 each accept a LORA_CONFIG from a "Fal API Flux LoRA Config" node - build one per LoRA you want stacked, set its URL and scale, and wire it in; none of the five slots are required.

Output is IMAGE.

Installing it. Same pack, same setup: ComfyUI Manager search for "ComfyUI-Fal-API-Flux" is the simplest route, or a manual clone of https://github.com/yhayano-ponotech/ComfyUI-Fal-API-Flux.git into custom_nodes followed by pip install -r requirements.txt. Either way, the node does nothing without a config.ini in the pack's folder carrying a [falai] section and your fal.ai api_key - that file isn't optional, and it isn't exposed anywhere in ComfyUI's own UI, so it's easy to forget on a fresh install.

Troubleshooting notes specific to this node. If the output looks basically identical to your input, strength is almost certainly too low, not a prompt problem - try nudging it up in steps of 0.1 rather than jumping straight to 1.0. If it looks nothing like your input, the opposite: pull strength down. For the LoRA side, the README's own advice holds here too - a lora_url has to be a direct link to a raw .safetensors file that fal's servers can fetch without a login, so a CivitAI page link typically won't work where a Hugging Face resolve/main/... link will.