Convert safetensors to bin

[paulocoutinhox]

Hi,

Im trying convert dreamshaper_8Inpainting.safetensors to bin, and it work when i run convert.py script.

python3 convert.py /Users/paulo/Developer/workspaces/python/stable-diffusion-webui/models/Stable-diffusion/dreamshaper_8Inpainting.safetensors
loading model from /Users/paulo/Developer/workspaces/python/stable-diffusion-webui/models/Stable-diffusion/dreamshaper_8Inpainting.safetensors
loading model from /Users/paulo/Developer/workspaces/python/stable-diffusion-webui/models/Stable-diffusion/dreamshaper_8Inpainting.safetensors completed
Stable diffuison 1.x
Saving GGML compatible file to /Users/paulo/Developer/workspaces/python/stable-diffusion.cpp/models/dreamshaper_8Inpainting-ggml-model-f16.bin
Processing tensor: alphas_cumprod with shape (1000,), float32 -> f32
Processing tensor: cond_stage_model.transformer.text_model.embeddings.position_embedding.weight with shape (77, 768), float16 -> f16
Processing tensor: cond_stage_model.transformer.text_model.embeddings.token_embedding.weight with shape (49408, 768), float16 -> f16
[...]
Processing tensor: model.diffusion_model.time_embed.2.weight with shape (1280, 1280), float16 -> f16
Convert done
Saved GGML compatible file to /Users/paulo/Developer/workspaces/python/stable-diffusion.cpp/models/dreamshaper_8Inpainting-ggml-model-f16.bin

The problem is when i try load it.

const std::string &model = modelOpt.value();
const std::string &prompt = promptOpt.value();

bool vaeDecodeOnly = true;

const char *rngTypeToStr[] = {
    "std_default",
};

Schedule schedule = KARRAS;
std::string negative_prompt;
float cfg_scale = 7.0f;
int w = 512;
int h = 512;
SampleMethod sample_method = DPMPP2M;
int sample_steps = 20;
float strength = 0.75f;
RNGType rng_type = STD_DEFAULT_RNG;
int64_t seed = -1;

StableDiffusion sd(-1, vaeDecodeOnly, true);

if (!sd.load_from_file(model, schedule))
{
    spdlog::error("[MappingStableDiffusion :: callbackGenerate] Error while load model: {}", model);
    r(Image{"ERROR-LOAD-MODEL"});
    return;
}

std::vector<uint8_t> img = sd.txt2img(
    prompt,
    negative_prompt,
    cfg_scale,
    w,
    h,
    sample_method,
    sample_steps,
    seed
);

Error:

[INFO]  stable-diffusion.cpp:2832 - loading model from '/Users/paulo/Developer/workspaces/python/stable-diffusion.cpp/models/dreamshaper_8Inpainting-ggml-model-f16.bin'
[INFO]  stable-diffusion.cpp:2860 - model type: SD1.x
[INFO]  stable-diffusion.cpp:2868 - ftype: f16
[ERROR] stable-diffusion.cpp:3047 - tensor 'model.diffusion_model.input_blocks.0.0.weight' has wrong shape in model file: got [3, 3, 9, 320], expected [3, 3, 4, 320]

[leejet]

The inpainting model has some differences compared to the original model and is not supported at the moment. You can use dreamshaper_8.safetensors for now.