README.md

---
title: hy3dlab Image-to-Image
emoji: 🖼️
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 6.22.0
python_version: 3.10.13
app_file: app.py
pinned: true
license: apache-2.0
short_description: Revision-pinned general and LoRA AI image editor
---

# hy3dlab Image-to-Image

Public backend duplicated from
[`prithivMLmods/Qwen-Image-Edit-2511-LoRAs-Fast`](https://huggingface.co/spaces/prithivMLmods/Qwen-Image-Edit-2511-LoRAs-Fast)
at commit `e1c9d2e8421d7934841d67f809d87ac7a6cdf720`.

General editing through the pinned replacement transformer is the default mode
(`lora_adapter="__base__"`). The same seven-field endpoint also accepts the 19
revision-pinned LoRA IDs recorded in `provenance.py`. Adapter changes use a
single-resident state machine: the prior adapter is unloaded before another is
loaded, and activation plus inference are serialized by one pipeline lock.

The Space App includes a first-party one-or-two-image editor with Prompt,
General/Style-LoRA selection, Quick Prompts, 19 fixed examples, queue status,
before/after comparison, session history, and PNG download. Browser code is
self-hosted and uses the same strict seven-field API contract.

## API boundary

- Named endpoint: `/edit_image`
- Non-GPU example endpoint: `/load_example`
- Input shape: the upstream seven fields
- Output shape: `{ "image": "data:image/png;base64,...", "seed": number }`
- Technical validation: one or two PNG/JPEG/WebP data URLs, at most 10 MiB and
  16 MP each (20 MiB and 32 MP total), Prompt length, numeric ranges, an exact
  allowlisted edit mode, and one concurrent pipeline operation
- Error output: stable low-cardinality codes; raw provider exceptions and user
  content are not returned or logged
- Input and output are processed in memory and are not deliberately written to
  application storage
- Generated output remains a single PNG with a maximum 1024px long side. The
  `Upscaler` adapter does not make this API a 4K-output service.

## Content-moderation decision

Application-layer content moderation is intentionally **not implemented** for
input images, prompts, or output images. The upstream transformer is retained.
This is a user-accepted product decision, not a claim that the model or service
passed an NSFW safety review, and it does not override Hugging Face policy or
applicable law.

## Reproducibility

Remote model, transformer, kernel, and all enabled LoRA revisions are pinned in
`provenance.py`. See `MODEL_BOM.md` for the recorded supply-chain facts and open
release gates; enablement is not a claim that every license/brand question has
been cleared. Python direct dependencies are pinned in `requirements.txt` and
`pre-requirements.txt`. The browser client is vendored at the installed
`@gradio/client` 2.3.1 build and hash-checked by the static test suite; a Linux
transitive lock remains a post-build task.

No Hugging Face owner token is required by the browser-facing public API design.
Deployment credentials must never be stored in this repository or sent to a
browser.