search

rupeshs / fastsdcpu

Fast stable diffusion on CPU
MIT License
1.01k stars 87 forks source link

RealVisXL v4.0 not generating images properly #184

Closed AlgorithmicKing closed 1 month ago

AlgorithmicKing commented 1 month ago

3de61218-7233-405d-abf7-afbb94eba64e-1 this is the image that it generated and it generates it every time I try to generate the image

AlgorithmicKing commented 1 month ago

well now it generates an image (after changing the interference steps to 25 or max value) but the images look extremely bad according to https://imgsys.org/ RealVisXL v4.0 is the best open source image model. here are the images that it generated prompt = nature landscape, size = 1024 1024 f6a6256d-7992-47ac-973a-82fbb1845554-1 prompt = nature landscape or fantasy landscape i dont remember, size = 256 256 4728a4a2-843f-46b8-95f7-0fadf4708328-1

AlgorithmicKing commented 1 month ago

prompt = the moon, size = 512 512 e2742b39-7cbc-470f-8fb7-6af0000ea1fd-1

AlgorithmicKing commented 1 month ago

same prompt same size 7594ae6f-499f-4d50-bdb4-909d1bef6ab1-1

rupeshs commented 1 month ago

@AlgorithmicKing Able to generate image without any issue. image

Config


{
    "lcm_model_id": "stabilityai/sd-turbo",
    "openvino_lcm_model_id": "rupeshs/sd-turbo-openvino",
    "use_offline_model": false,
    "use_lcm_lora": true,
    "lcm_lora": {
        "base_model_id": "SG161222/RealVisXL_V4.0",
        "lcm_lora_id": "latent-consistency/lcm-lora-sdxl"
    },
    "use_tiny_auto_encoder": false,
    "use_openvino": false,
    "prompt": "the moon",
    "negative_prompt": "",
    "strength": 0.6,
    "image_height": 768,
    "image_width": 768,
    "inference_steps": 3,
    "guidance_scale": 1.0,
    "number_of_images": 1,
    "seed": -1,
    "use_seed": false,
    "use_safety_checker": false,
    "diffusion_task": "text_to_image",
    "lora": {
        "weight": 0.5,
        "fuse": true,
        "enabled": false
    },
    "controlnet": null,
    "rebuild_pipeline": false
}
rupeshs commented 1 month ago

@AlgorithmicKing Follow the config, image size 768 x 768, steps 3 Mode - LCM-LoRa select LCM-Lora model as latent-consistency/lcm-lora-sdxl image

AlgorithmicKing commented 1 month ago

how do i set the lcm lora model to SG161222/RealVisXL_V4.0? i dont see the option

rupeshs commented 1 month ago

@AlgorithmicKing update stable-diffusion-models.txt Also refer : https://github.com/rupeshs/fastsdcpu?tab=readme-ov-file#models image

AlgorithmicKing commented 1 month ago

thanks a lot

AlgorithmicKing commented 1 month ago
  1. also how did you know to set those specific configuration where did you get where can i get it for other models?
rupeshs commented 1 month ago

@AlgorithmicKing please checkout readme it has lots of details for you. https://github.com/rupeshs/fastsdcpu/blob/main/Readme.md