Wagner Bruna
Wagner Bruna
The first changeset fixes safety checker behavior for models which do not incorporate the safety model internally (safetensors and some diffusers models). The other two are somewhat-related minor memory optimizations....
On img2img, the number of steps correspond to the last precalculated sigma values, but the internal alphas_cumprod and compvis_sigmas were being computed over the entire step range. Also, tweaks the...
This is on master-10c6501: 1. `./sd --model SD/cyberrealisticSemi_v30.safetensors -W 384 --seed 1 -p 'red shirt' --sampling-method ddim_trailing -o base.png` to create the base image (just to be sure DDIM itself...
Some SDXL LoRAs (eg. PCM) can exceed 12k nodes. Also tested with DMD2 4 step: ```text stable-diffusion.cpp/ggml/src/ggml.c:5764: GGML_ASSERT(cgraph->n_nodes < cgraph->size) failed (...) #12 0x000055769307cfa3 in ggml_build_forward_expand () #13 0x0000557692f0ff56 in...
On master-10c6501, SDXL embeddings crash with an assertion failure, either on a CPU build or when passing --clip-on-cpu. The Vulkan backend works fine. The following test is with CyberRealisticPony_v7 and...
Implement the linear LCM scheduler, as defined e.g. in Diffusers' LCMScheduler, and select it by default with the LCM sampler. The main change is ensuring the timesteps match the LCM...
This helps avoid a name conflict (for instance with the original function) when linking against the library.
Attempts to block LoRA names in the prompt from accessing files outside the specified LoRA model directory, which could be a security concern. Related discussion in: #820