External/SnapOtter
1
0
Fork 0
mirror of https://github.com/snapotter-hq/SnapOtter.git synced 2026-08-03 07:46:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix: enable tiling in Real-ESRGAN to prevent CUDA OOM on 8GB GPUs (#200)

Browse Source 
Process images in 512px tiles instead of all at once, drastically
reducing peak VRAM usage. If OOM still occurs, retry with 256px tiles
after clearing the CUDA cache. Covers both upscale and face enhance.

Closes #191
This commit is contained in:
SnapOtter
2026-06-05 17:29:56 +08:00
committed by GitHub
parent c365fde599
commit 074c96e8c3
1 changed files with 44 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/ai/python/upscale.py
+44 -7
View File
@@ -129,10 +129,13 @@ def main():
num_grow_ch=32,
scale=4,
)
tile_size = 512
upsampler = RealESRGANer(
scale=4,
model_path=REALESRGAN_MODEL_PATH,
model=ai_model,
tile=tile_size,
tile_pad=10,
half=use_gpu,
device=device,
)
@@ -140,7 +143,22 @@ def main():
img_array = np.array(img.convert("RGB"))
emit_progress(30, "Enhancing image with AI")
output_array, _ = upsampler.enhance(img_array, outscale=scale)
try:
output_array, _ = upsampler.enhance(img_array, outscale=scale)
except RuntimeError as oom_err:
if "out of memory" not in str(oom_err).lower():
raise
torch.cuda.empty_cache()
print(
f"[upscale] OOM with tile={tile_size}, retrying with tile=256",
file=sys.stderr,
flush=True,
)
emit_progress(35, "Retrying with smaller tiles")
upsampler.tile = 256
output_array, _ = upsampler.enhance(img_array, outscale=scale)
emit_progress(80, "AI enhancement complete")
result = Image.fromarray(output_array)
method = "realesrgan"
@@ -161,12 +179,31 @@ def main():
channel_multiplier=2,
bg_upsampler=upsampler,
)
_, _, face_output = face_enhancer.enhance(
img_array,
has_aligned=False,
only_center_face=False,
paste_back=True,
)
try:
_, _, face_output = face_enhancer.enhance(
img_array,
has_aligned=False,
only_center_face=False,
paste_back=True,
)
except RuntimeError as oom_err:
if "out of memory" not in str(oom_err).lower():
raise
torch.cuda.empty_cache()
print(
"[upscale] OOM during face enhance, retrying with tile=256",
file=sys.stderr,
flush=True,
)
emit_progress(84, "Retrying face enhancement")
upsampler.tile = 256
face_enhancer.bg_upsampler = upsampler
_, _, face_output = face_enhancer.enhance(
img_array,
has_aligned=False,
only_center_face=False,
paste_back=True,
)
result = Image.fromarray(face_output)
emit_progress(88, "Face enhancement complete")