leejet
Strange result
I got a correct result now. Is this result random?
This should not happen. It looks like something between the diffusion model and the vae happend...
Can you give more context, like os, exact parameters and device?
You can also run with -v to extract more info from sd.cpp .
Is this result random?
The only source of randomness should be the seed. If you get different results with the same seed and same other settings, then something is going very wrong, either with your computer or with the software. Last time I saw nondeterministic outputs it was because something was reading data from the buffer of a destroyed vector.
Broken image metadata:
change 'flux.cpp' to 'kontext.cpp' Steps: 20, CFG scale: 1.000000, Guidance: 3.500000, Eta: 0.000000, Seed: 42, Size: 512x512, Model: , RNG: cuda, Sampler: euler_a, TE: clip_l.safetensors, TE: t5xxl_fp16.safetensors, Unet: flux1-kontext-dev-Q8_0.gguf, VAE: ae.safetensors,
Good image:
change 'flux.cpp' to 'kontext.cpp' Steps: 20, CFG scale: 1.000000, Guidance: 3.500000, Eta: 0.000000, Seed: 42, Size: 512x512, Model: , RNG: cuda, Sampler: euler, TE: clip_l.safetensors, TE: t5xxl_fp16.safetensors, Unet: flux1-kontext-dev-Q8_0.gguf, VAE: ae.safetensors,
Flow models like Flux and SD3 only work well with a few samplers, and euler_a is not one of them. Using euler like you did in the second image is indeed recommended.
I got a correct result now. Is this result random?
This should not happen. It looks like something between the diffusion model and the vae happend...
Can you give more context, like os, exact parameters and device? You can also run with
-vto extract more info from sd.cpp .
I'm unable to reproduce the erroneous result anymore.
Is this result random?
The only source of randomness should be the seed. If you get different results with the same seed and same other settings, then something is going very wrong, either with your computer or with the software. Last time I saw nondeterministic outputs it was because something was reading data from the buffer of a destroyed vector.
I used the same seed and parameters.