Readme
Qwen Image Edit 2509 Plus – Skin LoRA
Camera-aware edits, cinematic scene swaps, product fusion, relighting, restoration, upscale, skin retouch, and photo→anime conversions—all powered by Qwen Image Edit 2509 (Plus) with Lightning optimizations. Each preset loads a different Hugging Face LoRA by default but still exposes the full generic interface (custom LoRA weights, Lightning toggle, multi-image edits, etc.).
Presets & Links
Inputs (shared across every preset)
| Name | Description | Default |
|---|---|---|
image |
One or more JPEG/PNG/GIF/WebP references. | required |
prompt |
Edit instruction appended after the preset’s bilingual hint. | preset-specific hint |
aspect_ratio |
match_input_image, 1:1, 16:9, 9:16, 4:3, 3:4. |
match_input_image |
go_fast |
Lightning path (8 steps) when true. Set false for the 40-step detailed pass. | true (multi-angle overrides to false) |
num_inference_steps |
Force 1–40 steps. Leave blank to rely on the go_fast preset. | blank |
lora_weights |
Hugging Face slug (owner/repo or owner/repo/file.safetensors) or direct HTTPS link (.safetensors, .zip, .tar.*). Blank uses the preset’s bundled LoRA; multi-angle blank = dx8152/Qwen-Edit-2509-Multiple-angles, etc. |
preset default |
lora_scale |
Strength of the active LoRA (0–4). Each preset shows its tuned value in the UI. | preset default |
true_guidance_scale |
CFG value passed through to the engine. | 1.0 |
seed |
Optional integer for deterministic runs. | blank |
output_format |
webp, jpg, or png. |
webp |
output_quality |
0–100 quality for webp/jpg. | 95 |
disable_safety_checker |
Skip Stable Diffusion + Falcon NSFW filters. | false |
How LoRA selection works
- Slugs, blob URLs, resolve URLs, and direct HTTPS links all normalize to the same cached file. Chinese characters and other UTF-8 filenames are preserved.
- Archives (.zip/.tar/.tgz/.tar.gz/.tar.xz/.tar.bz2) are unpacked automatically; the first
.safetensorsfound is used unless you specify the filename in the slug. - Each download is cached under
/tmp/lora_cache/<hash>in both raw and transformed form, so repeated runs hit disk instead of the network. - Lightning (go_fast=true) stays active alongside your LoRA unless you disable it; set go_fast=false for a base-model pass or to run a custom LoRA solo.
- Setting
lora_scale=0effectively disables the adapter while keeping it cached for later.
Tips
- Start with the preset hints—they’re bilingual triggers copied straight from the LoRA authors; add your own instruction after the hint.
- Need a different LoRA? Paste any Hugging Face slug or resolve URL into
lora_weights; the UI prints the available filenames when you point at a repo. - Multi-angle’s camera sliders can drive an edit without any extra prompt text; use small increments (±15° tilt/rotate) for clean results.
- Keep
go_fast=truefor speed and LoRA blending; flip it off when you want the 40-step high-contrast look or when debugging a custom LoRA. - Leave
disable_safety_checkerat false unless you’re testing locally—both Stable Diffusion and Falcon NSFW filters run otherwise.
Credits
- Base pipeline: Qwen/Qwen-Image-Edit-2509 (Plus variant).
- Lightning adapter: lightx2v/Qwen-Image-Lightning.
- LoRAs: linked above; all credit to their Hugging Face authors.
- Proxy runtime: Pruna optimizations, Cog packaging, and Replicate-hosted GPU backends.