MacOS black image generation.

[l0stl0rd]

I keep running into the same error with most samplers, for a while now and with the latest commit f1.0.2v1.10.1-previous-214-gdad1d17f.

I did try on 2 different Apple Silicon Machines on with MacOS 14.6.1 and one with beta 5 of MacOS 15.

stable-diffusion-webui-forge/modules/processing.py:990: RuntimeWarning: invalid value encountered in cast x_sample = x_sample.astype(np.uint8)

or

stable-diffusion-webui-forge/modules/sd_samplers_common.py:75: RuntimeWarning: invalid value encountered in cast x_sample = x_sample.astype(np.uint8)

Update:

sometime I get both, did 20 steps DPM++ 2M using 1024 x 1024 px and at step 11 I get stable-diffusion-webui-forge/modules/sd_samplers_common.py:75: RuntimeWarning: invalid value encountered in cast x_sample = x_sample.astype(np.uint8)

At the end I get /stable-diffusion-webui-forge/modules/processing.py:990: RuntimeWarning: invalid value encountered in cast x_sample = x_sample.astype(np.uint8)

Wondering if it is caused by the same math rounding issues some used to have. https://github.com/vladmandic/automatic/issues/2012

Samplers tested: DDIM = works DPM++ 2M = broken DPM++ SDE = broken DPM++ 2M SDE = broken DPM++ 2sa = broken DPM++ 3M DES = works Euler a = broken Euler = works UniPC = works

These work fine in A1111 however (on the same machines).

[l0stl0rd]

issue remains with version : f2.0.1v1.10.1-previous-223-g86ee2d94

[TheToLoveRuDarkness]

Hello, I updated to 14.6.1 yesterday and now I have the same problem. Is there already a solution?

[9wow]

+1 (SDXL)

Sometimes Euler a (Normal) works if you keep trying the generate after the black image. If you see it turn green in the preview double click interupt and try again.

[l0stl0rd]

The issue sure is a bit weird sometimes I think it is fixed change sampler or scheduler and the error is back.

I think it is not the samplers that are broken but some schedulers.

Euler a, on auto, normal l, ays11 did work, karras and simple did not.

[9wow]

I think it is not the samplers that are broken but some schedulers.

Yea, I guess I'll have to stick to SGM Uniform for now until this (hopefully) gets addressed.

[IPv6]

Also hit similar problem - some scheduler/sampler pairs always produce black images on M3 - and thankfully same problem was with Comfy and same solution helped me. It is "--force-upcast-attention" command line option, with this option Karras and DPM++ 2M SDE work without black images. Full command line for reference: --api --skip-torch-cuda-test --upcast-sampling --force-upcast-attention --no-half-vae --all-in-fp16 --always-high-vram --use-cpu interrogate

[jeffclayton]

Also hit similar problem - some scheduler/sampler pairs always produce black images on M3 - and thankfully same problem was with Comfy and same solution helped me. It is "--force-upcast-attention" command line option, with this option Karras and DPM++ 2M SDE work without black images. Full command line for reference: --api --skip-torch-cuda-test --upcast-sampling --force-upcast-attention --no-half-vae --all-in-fp16 --always-high-vram --use-cpu interrogate

this helped a lot. I am using a macbook pro with 16gb ram (low vram for this) using pytorch and I was able to load Juggernaut XL without black images by cutting off just the last part in the webui-user.sh:

export COMMANDLINE_ARGS="--api --skip-torch-cuda-test --upcast-sampling --force-upcast-attention --no-half-vae"

[Chandy8626]

I'm having the same problem whenever I increase the batch count above one. The first image comes out, but any subsequent images are black boxes. I can see the images while they are generating, but when it's finished, they are black.

I'm using a M1 MacBook with 16GB RAM. I've been able to use A1111, but not Forge. I would prefer Forge for various reasons.

I added COMMANDLINE_ARGS="--api --skip-torch-cuda-test --upcast-sampling --force-upcast-attention --no-half-vae" to webui.sh but it didn't fix the issue.

Does anyone have a better idea?

[aidanxoxo]

Im using a M1 Pro with 16GB RAM and have the same problem of subsequent images being black. the command line did not fix the issue, too

[l0stl0rd]

there seems to be no way around it unless I add --force-upcast-attention I get black pictures even with pytorch 2.7.

My webui-macos-env.sh not looks like this.

export install_dir="$HOME" export COMMANDLINE_ARGS="--skip-torch-cuda-test --upcast-sampling --force-upcast-attention --no-half-vae --use-cpu interrogate" export PYTORCH_ENABLE_MPS_FALLBACK=1

if [[ "$(sysctl -n machdep.cpu.brand_string)" =~ ^."Intel".$ ]]; then export TORCH_COMMAND="pip install torch==2.1.2 torchvision==0.16.2" else export TORCH_COMMAND="pip install torch==2.7.0 torchvision==0.22.0" fi

My guess is and I am not a programmer, that the problem comes down to MPS not supporting FP8

[akilanarasu1994]

Also hit similar problem - some scheduler/sampler pairs always produce black images on M3 - and thankfully same problem was with Comfy and same solution helped me. It is "--force-upcast-attention" command line option, with this option Karras and DPM++ 2M SDE work without black images. Full command line for reference: --api --skip-torch-cuda-test --upcast-sampling --force-upcast-attention --no-half-vae --all-in-fp16 --always-high-vram --use-cpu interrogate

I can confirm that this works on my M2 Max with 96gb RAM.