프로그램 사용/ai 프로그램

stable diffusion 모델 변경(SD -> SDXL)

구차니 2026. 5. 12. 10:15

antigravity 가라사대~

512x 512에서 1024x1024로 바꾸란다.

왜 되지?!?!

1. 해상도 (Resolution) 변경 🌟 (가장 중요)
SD 1.5는 기본적으로 512x512 해상도로 학습되었습니다.
SDXL은 1024x1024 기반으로 학습되었습니다. SDXL에서 512x512로 생성하면 이미지가 깨지거나 노이즈 덩어리처럼 나오게 됩니다.
변경: width와 height를 1024로 변경하거나, 원하는 비율에 맞춰 SDXL 권장 해상도(예: 832x1216, 1152x896 등)로 변경해야 합니다.
2. CFG Scale (분류기 자유도) 하향
현재 코드에 설정된 cfg_scale: 15 (img2img는 30)는 SD 1.5에서도 꽤 높은 편이지만, SDXL에서는 너무 높은 수치입니다. CFG가 너무 높으면 색감이 타고 이미지가 일그러집니다.
변경: SDXL은 프롬프트 이해력이 뛰어나므로 cfg_scale을 5.0 ~ 7.0 사이로 낮추는 것이 좋습니다.
3. 네거티브 프롬프트 (선택 사항)
SDXL (특히 파생 모델들)은 SD 1.5만큼 복잡하고 긴 네거티브 프롬프트나 Embedding(EasyNegative 등)에 크게 의존하지 않습니다. 오히려 너무 무거운 네거티브 프롬프트는 생성을 방해할 수 있습니다.
현재 코드에 있는 base_neg 정도면 나쁘지 않으나, 필요하다면 줄일 수 있습니다. (우선은 그대로 두셔도 무방합니다.)

 

512x512 생성 실패, VAE automatic 설정

 

1024x1024 생성 성공, VAE automatic 설정

 

 

----

API  •  GitHub  •  Gradio  •  Startup profile  •  Reload UI
version: v1.10.1-96-g1937682a  •  python: 3.10.12  •  torch: 2.1.2+cu121  •  xformers: N/A  •  gradio: 3.41.2  •  checkpoint: 6ce0161689

 

[링크 : https://civitai.com/models/795765/illustrious-xl]

[링크 : https://huggingface.co/OnomaAIResearch/Illustrious-xl-early-release-v0/tree/main]

 

illustriousXL 모델로 하면 이상하게 나오는데

 

처음에 해봤던 V1.5 emanoly는 정상적으로 잘 나온다.(물론 이상한 비행기 형태인건 여전하지만)

 

 

There's not much to it. The SDXL models work similar to other models. They are just more resource intensive.
Update your A1111 installation. I find it easiest to just run "git pull" from a command line.
Then download an SDXL model and VAE and put them with your other SD models. (The official SDXL release consists of a base model and a refiner but most people don't seem to bother with refiners.)
If your graphics card has less than 12 GB VRAM, then add the "--medvram-sdxl" argument.
Don't expect your 1.5 prompts to perform well with SDXL. You'll need to relearn how to write good prompts. Go to Civitai, grab a model that you like, look at the example images and study their prompts and settings. SDXL works best at around 1024×1024 pixels.

[링크 : https://www.reddit.com/r/StableDiffusion/comments/1bwidjg/updated_guide_on_how_to_run_sdxl_on_a1111/]

[링크 : https://huggingface.co/stabilityai/sdxl-vae]

 

/mnt/Downloads/stable-diffusion-webui/models$ ls -alR VAE*
VAE:
total 8
drwxrwxr-x  2 falinux falinux 4096  5월  4 21:38  .
drwxrwxr-x 11 falinux falinux 4096  5월  4 21:51  ..
-rw-rw-r--  1 falinux falinux    0  5월  4 21:38 'Put VAE here.txt'

VAE-approx:
total 432
drwxrwxr-x  2 falinux falinux   4096  5월  6 21:57 .
drwxrwxr-x 11 falinux falinux   4096  5월  4 21:51 ..
-rw-rw-r--  1 falinux falinux 213777  5월  4 21:38 model.pt
-rw-rw-r--  1 falinux falinux 213777  5월  6 21:57 vaeapprox-sdxl.pt

 

$ cat /mnt/Downloads/stable-diffusion-webui/repositories/generative-models/configs/inference/sd_xl_base.yamlml
model:
  target: sgm.models.diffusion.DiffusionEngine
  params:
    scale_factor: 0.13025
    disable_first_stage_autocast: True

    denoiser_config:
      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
      params:
        num_idx: 1000

        weighting_config:
          target: sgm.modules.diffusionmodules.denoiser_weighting.EpsWeighting
        scaling_config:
          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
        discretization_config:
          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization

    network_config:
      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
      params:
        adm_in_channels: 2816
        num_classes: sequential
        use_checkpoint: True
        in_channels: 4
        out_channels: 4
        model_channels: 320
        attention_resolutions: [4, 2]
        num_res_blocks: 2
        channel_mult: [1, 2, 4]
        num_head_channels: 64
        use_spatial_transformer: True
        use_linear_in_transformer: True
        transformer_depth: [1, 2, 10]  # note: the first is unused (due to attn_res starting at 2) 32, 16, 8 --> 64, 32, 16
        context_dim: 2048
        spatial_transformer_attn_type: softmax-xformers
        legacy: False

    conditioner_config:
      target: sgm.modules.GeneralConditioner
      params:
        emb_models:
          # crossattn cond
          - is_trainable: False
            input_key: txt
            target: sgm.modules.encoders.modules.FrozenCLIPEmbedder
            params:
              layer: hidden
              layer_idx: 11
          # crossattn and vector cond
          - is_trainable: False
            input_key: txt
            target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2
            params:
              arch: ViT-bigG-14
              version: laion2b_s39b_b160k
              freeze: True
              layer: penultimate
              always_return_pooled: True
              legacy: False
          # vector cond
          - is_trainable: False
            input_key: original_size_as_tuple
            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
            params:
              outdim: 256  # multiplied by two
          # vector cond
          - is_trainable: False
            input_key: crop_coords_top_left
            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
            params:
              outdim: 256  # multiplied by two
          # vector cond
          - is_trainable: False
            input_key: target_size_as_tuple
            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
            params:
              outdim: 256  # multiplied by two

    first_stage_config:
      target: sgm.models.autoencoder.AutoencoderKLInferenceWrapper
      params:
        embed_dim: 4
        monitor: val/rec_loss
        ddconfig:
          attn_type: vanilla-xformers
          double_z: true
          z_channels: 4
          resolution: 256
          in_channels: 3
          out_ch: 3
          ch: 128
          ch_mult: [1, 2, 4, 4]
          num_res_blocks: 2
          attn_resolutions: []
          dropout: 0.0
        lossconfig:
          target: torch.nn.Identity

 

SD VAE 에서 refresh 하고

sdxl_vae.safetensors 를 선택하고 apply settings 해준 후

 

그리기 시도하니 얼추 비슷... 한건가?

 

원래는 이런 이미지

Hatsune Miku,limited palette,black background,colorful,vibrant,glowing outline,neon,blacklight,looking at viewer, masterpiece, very aesthetic Negative prompt: worst quality,bad quality,bad hands,very displeasing,extra digit,fewer digits,jpeg artifacts,signature,username,reference,mutated,lineup,manga,comic,disembodied,futanari,yaoi,dickgirl,turnaround,2koma,4koma,monster,cropped,amputee,text,bad foreshortening,what,guro,logo,bad anatomy,bad perspective,bad proportions,artistic error,anatomical nonsense,amateur,out of frame,multiple views, Steps: 28, CFG scale: 7.5, Sampler: Euler a, Seed: 3625896228, Size: 832x1216, Model: Illustrious-XL-v0.1, Version: f1.0.2-v1.10.1RC-latest-691-g37223711, Model hash: 3e15ba0038, Schedule type: Automatic, Discard penultimate sigma: True

 

--lowvram

 

[링크 : https://huggingface.co/stabilityai/models?search=sdxl]

[링크 : https://huggingface.co/stabilityai/sdxl-turbo/tree/main]

[링크 : https://huggingface.co/stabilityai/sdxl-vae/tree/main]

 

 

 

저작자표시 (새창열림)