CthomasdesignandClaude Fable 5
5fcf343a4d
_shared gains default_model()/default_lora() (env overrides KREA2_TEST_MODEL/KREA2_TEST_LORA, else <repo>/../Models discovery with clear errors) and fox_source()/ref_uint8() for the reference source image boilerplate repeated across scripts. test_img2img/test_edit/ test_edit_grounded no longer hardcode machine paths; all tools compile and resolve on a fresh checkout. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
40 lines
1.6 KiB
Python
40 lines
1.6 KiB
Python
"""Validate the edit path: txt2img parity untouched, then a real identity-LoRA edit."""
|
|
|
|
import numpy as np
|
|
from PIL import Image
|
|
|
|
from _shared import PROMPT_SHORT, default_lora, default_model, fox_source, load_pipeline, ref_uint8
|
|
|
|
pipe, _ = load_pipeline(default_model())
|
|
LORA = default_lora()
|
|
|
|
# 1) parity guard: adding the edit code must not disturb txt2img
|
|
t2i = pipe.generate(PROMPT_SHORT, width=512, height=512, steps=8, seed=0)[0]
|
|
print(f"[guard] txt2img still matches reference (pixel-identical): "
|
|
f"{np.array_equal(np.array(t2i), ref_uint8())}")
|
|
|
|
src = fox_source()
|
|
|
|
# 2) edit WITHOUT the LoRA (baseline) vs WITH the identity LoRA — both should run
|
|
def edit(prompt, seed, use_lora):
|
|
pipe.set_loras([(LORA, 1.0)] if use_lora else [])
|
|
im = pipe.generate_edit(prompt, src, width=512, height=512, steps=8, seed=seed)[0]
|
|
pipe.set_loras([])
|
|
return im
|
|
|
|
sref = np.asarray(src, np.float32) / 255.0
|
|
for use_lora in (False, True):
|
|
im = edit("the fox wearing a red scarf", seed=0, use_lora=use_lora)
|
|
a = np.asarray(im, np.float32) / 255.0
|
|
mae = float(np.abs(a - sref).mean())
|
|
tag = "LoRA" if use_lora else "base"
|
|
im.save(f"reference/edit_{tag}.png")
|
|
print(f"[edit:{tag}] ran; mean|Δ| to source = {mae:.4f} -> reference/edit_{tag}.png")
|
|
|
|
# contact sheet: source | edit(base) | edit(LoRA)
|
|
sheet = [np.asarray(src.resize((384, 384)))]
|
|
for tag in ("base", "LoRA"):
|
|
sheet.append(np.asarray(Image.open(f"reference/edit_{tag}.png").resize((384, 384))))
|
|
Image.fromarray(np.concatenate(sheet, axis=1)).save("reference/edit_sheet.png")
|
|
print("wrote reference/edit_sheet.png (source | base | LoRA)")
|