Flux fill is prone to crashing

[yincangshiwei]

Expected Behavior

Normal output.

Actual Behavior

After a few drawings it occasionally looks like this：

Steps to Reproduce

flux fill test.json

Debug Logs

got prompt
using extra model: /root/ComfyUI/models/sams/sam_hq_vit_l.pth
<All keys matched successfully>
grounding-dino is using models/bert-base-uncased
final text_encoder_type: /root/ComfyUI/models/bert-base-uncased
# 😺dzNodes: LayerStyle -> SegmentAnythingUltra Processed 1 image(s).
loaded completely 18778.313576507568 18638.461921691895 True
torch.Size([1, 1, 1024, 768])
torch.Size([1, 1, 1024, 768])
100%|██████████| 20/20 [00:19<00:00,  1.03it/s]
Prompt executed in 48.17 seconds

Other

No response

[maizhouzi]

I encountered the same issue; it seems that during the iteration process, one or more steps turn into complete noise.

[drphero]

I'm also running into this same problem. Sometimes it happens on certain seeds, but on other images/prompts it happens every single time.

EDIT: I just tested this on an L40S via vast.ai and this problem didn't happen a single time during dozens of inpaints/outpaints.
I wonder if this has something to do with the GPU architecture? My normal card is an RTX Quadro 6000 which has the Turing architecture, while the L40S is Ada Lovelace. I've run into problems before with my card such as it not supporting bf16 as well as other things that require Ampere or newer.

[yincangshiwei]

It should be related to this:
black-forest-labs/flux#195