Spaces:

---

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

LPX55

/

Qwen-Image-Edit-2511-Turbo-Lightning

like 100

Paused

LoRA Auto Key-Fix (PR #7)

Adds automatic repair for "troublesome" LoRAs whose safetensors omit the <module>.alpha keys diffusers' Qwen LoRA converter expects (raising KeyError: 'img_in.alpha'), or include ComfyUI delta/patch keys (.diff, .diff_b) that diffusers' strict converter rejects (ValueError: state_dict should be empty...).

What changed

Three new helper functions (lines ~400–500):

1. _inject_missing_alpha_keys(state_dict) — injects <module>.alpha = rank for every LoRA module, in BOTH diffusion_model.<module>.alpha AND stripped <module>.alpha forms (diffusers may strip the prefix before lookup). Fixes KeyError: 'img_in.alpha'.

2. _filter_to_diffusers_lora_keys(state_dict) — keeps only keys diffusers can consume (*.lora_up.weight, *.lora_down.weight, *.lora_mid.weight, *.alpha + normalizes *.lora_alpha → *.alpha), drops ComfyUI delta keys (*.diff, *.diff_b) and any other noise. Returns cleaned dict + stats.

3. _duplicate_stripped_prefix_keys(state_dict) — duplicates any diffusion_model. prefixed LoRA keys in unprefixed form, ensuring compatibility regardless of whether diffusers strips the prefix.

4. _load_lora_weights_autofix(path, adapter_name) — wraps pipe.load_lora_weights. Tries normal load first; on KeyError or ValueError, downloads the file, runs the three helpers, loads the cleaned dict. Runs automatically for every custom LoRA (no opt-in flag needed).

Modified load_custom_lora

• Replaces direct pipe.load_lora_weights(path, ...) call with _load_lora_weights_autofix(path, CUSTOM_LORA_ADAPTER).
• Adds safetensors.torch.load_file import (was missing) and traceback import.

Why automatic is better than opt-in

The reference space (cruisewagner2220) uses a needs_alpha_fix: True per-LoRA config flag. Ours detects the failure mode at load time and self-repairs — covers alpha-missing AND leftover-keys errors without per-LoRA configuration.