Build and publish Gradio demos for LoRA models

A skill for building and publishing a Gradio demo on Hugging Face Spaces that runs inference with a user-provided LoRA - for base models like Qwen-Image, Qwen-Image-Edit, LTX-Video, Wan, FLUX, or SDXL. The default target is ZeroGPU hardware with diffusers as the default inference library where supported, and the output is a real, published (private-by-default) Space the user can try in the browser, not a local script. The central design principle: two LoRAs sharing a task can still need completely different demos (a pose-control video LoRA and an outpainting video LoRA both take video in and out, but their inputs, preprocessing, and controls differ entirely), so the demo must be handcrafted for the specific LoRA rather than a generic template - loading and running fast, exposing exactly the controls that LoRA needs and nothing more, surfacing progress/seed/clear errors, honoring the LoRA's own model-card recommendations (trigger words, step count, guidance scale, LoRA scale, example inputs), and using creative UI patterns (interactive canvases, before/after sliders) where they help. The six-phase workflow, questions batched rather than drip-fed across turns: Phase 1 gathers LoRA info - checking for an already-authenticated HF token before ever asking the user for one (only prompting once, with write scope, if the repo is private/gated and no valid cached token can read it), listing repo files to find the right .safetensors weights file (README-recommended wins, then pytorch_lora_weights.safetensors, then the latest checkpoint, else ask), and reading the model card to determine base model, task (one of text-to-image, image-to-image, text-to-video, image-to-video, video-to-video), trigger words, the recommended inference recipe (trusting a card's example snippet for steps/guidance/CFG/LoRA-scale/dtype parameters while preferring pipe.load_lora_weights() over whatever loading code the snippet shows), example prompts/media for Gradio examples, and the LoRA's specific sub-task since a relighting, face-swap, and style LoRA can all technically be "image-to-image" but need different UIs. Phase 2 picks the base pipeline and inference recipe; Phase 3 designs the UI specific to that LoRA's task and inputs; Phase 4 writes app.py, requirements.txt, and README.md together, presenting all three for one batched user approval; Phase 5 publishes the Space as private; Phase 6 smoke-tests the published Space.

Use it whenever someone asks to create, generate, ship, or publish a Space, demo, Gradio app, or playground for a LoRA they trained or host on the Hub - for diffusion base models the skill supports (Qwen-Image, Qwen-Image-Edit, LTX-Video, Wan, FLUX, SDXL, and others).

Input is a LoRA repo ID on the Hugging Face Hub (and an access token if the repo is private/gated). Output is a published, private-by-default Hugging Face Space with app.py, requirements.txt, and README.md tailored to that specific LoRA's task, inputs, and model-card recommendations.

from huggingface_hub import HfApi, get_token

Uses huggingface_hub for repo/token/model-card access, diffusers as the default inference library, Gradio for the UI, and ZeroGPU as the default Space hardware target.

Users who trained or host a LoRA on the Hugging Face Hub and want a real, shareable, browser-based demo built around that specific LoRA's task and recommended settings, without hand-writing the pipeline, UI, and Space publishing steps themselves.