Spaces:

---

Duplicated from  linoyts/Qwen-Image-Edit-2509-Fast

LPX55

/

Qwen-Image-Edit-2511-Turbo-Lightning

Like 121

Running on Zero

What this fixes

After the dimensions UI shipped in PR #4 (auto-fill from the first uploaded image, per-dimension fallback in infer()), the sliders never auto-fill on the first upload — and infer() always falls back to 1024. The root cause is in Gradio's Slider implementation:

self.minimum if value is None else value   # Slider line 147

A gr.Slider(value=None, minimum=256) reports 256 (its minimum), not None, to every event handler. PR #4's "unset" check (current_height is None / height is None) therefore never fires — the slider looks already-set to 256, and infer() always bails to the 1024 fallback.

Fix

Replaces value-based unset detection with user-interaction flags:

• Two per-session gr.State(False) components: height_user_set, width_user_set.
• Each slider gets a .input() listener (fires only on user interaction, never on programmatic gr.update() — confirmed via Gradio's docstrings and source) that flips its flag to True.
• update_dimensions_on_upload(images, height_set, width_set) now treats not height_set/not width_set as "unset" and auto-fills those dimensions from the uploaded image (preserving dimensions the user already set).
• infer(..., height_user_set=False, width_user_set=False, ...) mirrors this in the runtime fallback — only dimensions the user hasn't touched get derived from the first input image.
• These flags are appended to the input_images.upload inputs and the gr.on inference-trigger inputs list (positional order matches the updated infer() signature).

Verification

• python3 -m py_compile passes.
• Handler logic unit-tested with real gradio 6.17.3 + PIL:
  • untouched (2048×2048) → auto-fills 1024×1024 ✓
  • height user-set, width unset → height preserved, width auto-filled ✓
  • both user-set → both preserved (no-op gr.update()) ✓
  • empty gallery → both untouched ✓
• Diff: +22/-13 across 6 hunks, all in the dimension-handling regions (update_dimensions_on_upload, infer() signature/fallback, input_images.upload, gr.on inputs). No overlap with PR #5's LoRA regions — applied cleanly onto the post-#5 main.

Why not value-based

gr.Slider eagerly coerces None → minimum. Tracking via .input() is the documented, stable way to distinguish user-set values from untouched defaults — and it survives Gradio's coercion without fighting it.