huggingface-image-generation

name	huggingface-image-generation
description	Use Hugging Face for image generation and image editing workflows. Use when the user asks to build, integrate, debug, or deploy text-to-image or image-to-image features with Hugging Face, including Inference Providers, `@huggingface/inference`, `huggingface_hub` `InferenceClient`, model selection, prompt and negative prompt tuning, seeds, dimensions, and handling returned image bytes or blobs.

Hugging Face Image Generation

Use this skill for product-facing image generation and editing features, not for training or fine-tuning.

Check current docs first

Hugging Face image APIs and provider support change quickly. Verify the current official docs before coding:

• text-to-image: https://huggingface.co/docs/inference-providers/tasks/text-to-image
• image-to-image: https://huggingface.co/docs/inference-providers/en/tasks/image-to-image
• JS client docs: https://huggingface.co/docs/huggingface.js/main/en/inference/modules

Default approach

Prefer hosted inference first:

• Web or Node apps: @huggingface/inference
• Python apps or scripts: huggingface_hub.InferenceClient

Only switch to local Diffusers when the user explicitly wants self-hosting, GPU-local inference, or lower-level pipeline control.

Auth

• Read the token from HF_TOKEN.
• Never hardcode tokens in source, tests, docs, or examples.
• If a provider-specific key is required, keep it in a separate env var and document it explicitly.

App integration checklist

• Decide whether the workflow is text-to-image or image-to-image.
• Let callers pass model, prompt, negative_prompt, seed, width, height, guidance_scale, and num_inference_steps when supported.
• Treat provider and model support as dynamic. Check the current task docs and model card instead of assuming every parameter is available everywhere.
• Return binary image data safely: Blob, ArrayBuffer, or a file response in JS; PIL.Image or raw bytes in Python.
• Save generated files outside source trees unless the user explicitly wants checked-in assets.

Good defaults

• Start with moderate image sizes and step counts.
• Expose seed whenever reproducibility matters.
• If the user asks for an editor, prefer image-to-image with a source image plus prompt instead of regenerating from scratch.
• Frame model names as examples, not permanent defaults, because the recommended set changes frequently.

JavaScript note

Prefer InferenceClient from @huggingface/inference. Keep file handling explicit when converting between Blob, buffers, form uploads, and saved files.

Python note

Prefer InferenceClient from huggingface_hub for hosted inference. Use local Diffusers only when the task actually needs local generation or advanced control.