Sign In

Eris

1.0k
3.6k
23
Verified:
Pruned SafeTensor
Type
Checkpoint Merge
Stats
3,648
Reviews
Published
Mar 20, 2023
Base Model
SD 1.5
Hash
AutoV2
210D0D5D5D

Eris, the Greek goddess of strife and discord.

This is my first pure 3D render/Cartoon model. When I started working on this model I had almost zero knowledge on writing prompts for 3d renders/CGI. As such, my main goal for this model was quickly realized - Make a model that accepts photorealistic prompts and returns beautiful 3D rendered outputs. Normal 3D render/CGI prompts also work with this model.

Important Settings

The following settings assume you're using webui by Automatic1111

  • USE CFG VALUES BELLOW 8*

    • This model does not respond well to CFG values above 8. With some caveats that will be explained bellow.

    • Do not go over 6 CFG with Euler a unless,

    • You are using Hires Fix. I found that you can go as high as 10 CFG with hires. fix enabled. This applies to all samplers.

    • DPM++ SDE Karras starts to degrade around 7 CFG if using clip skip = 1, no hires fix. However, if you're using DPM++ SDE Karras with clip skip = 2 the CFG values can go up to at least 10 CFG without significant (if any) degradation.

    • General Advice:

      • Do Not exceed 6 CFG with Euler a, unless using hires fix.

      • Do Not exceed 7 CFG with DPM++ SDE Karras with clip skip set to 1, unless using hires fix.

      • You Can go up to at least 10 CFG with DPM++ SDE Karras with clip skip set to 2, without using hires fix.

      • I could not test all samplers. If you notice broken compositions or distortions; lower your CFG value and/or enable hires fix. My own personal settings for hires fix can be found bellow.

    • Thanks to Duo V2 on the Unstable Diffusion discord for helping me get to the bottom of this issue.

  • VAE: vae-ft-mse-840000-ema-pruned.safetensors

  • Clip skip : 2

    • This varies. Most of the time Clip skip = 2 works best with this model. In the event you're not getting the results you want. First, try Clip skip = 1.

  • ETA Noise Seed Delta: 31337

    • Found in - Settings -> Sampler parameters -> Eta noise seed delta

  • Sigma Noise: 1

  • Sigma Churn: 0 or 1

    • Optional - When using Euler, Heun, DPM2 samplers try setting Sigma Churn = 1

    • Found in - Settings -> Sampler parameters -> sigma churn

  • Optional Negative Textual Inversion (TI): bad_prompt_version2

  • Optional Recommended Samplers:

    • DPM++ SDE Karras // 18-35 steps

    • Euler // 40 - 80 steps // Sigma Churn : 1

    • DPM++ 2M Karras // 40 - 70 steps

    • DPM++ 2S a Karras // 30 - 75 steps

    • Heun // 20 - 30 steps // Sigma Churn : 1

    • DPM2 // 30 - 75 steps // Sigma Churn : 1

    • These are just recommendations; feel free to experiment with different values and samplers.

  • Optional Hires. Fix settings (these are my go-to parameters):

    • Upscaler : R-ESRGAN 4x+

    • Denoising strength: 0.3 - 0.35

    • The rest of the settings are up to you. They don't really impact the quality of the output.

    • Note: Hires Steps, if you're using high steps for SD - You can choose a hires step value lower than your SD step value. This will increase the generation times.

      • For example; Let's say you're using Euler sampler @ 60 steps. You can then set the Hires Steps to ~40. Apposed to the default which is 0. Hires Steps = 0 means it will follow the same step value as sampler (Euler in this case). Lower Hires Step values will reduce generation time.

  • Optional Face Restoration:

    • I use CODEFORMER @ 0.8 strength. Whether or not you use face restoration is up to you.

  • Optional Quick Settings:

    • For easier access to some of the settings in webui, you can move some settings sliders/dropdowns to the main interface (beside model selection).

    • Quick settings location -

      • Settings -> User interface -> Quicksettings list

    • My Quick settings -

      • sd_model_checkpoint, CLIP_stop_at_last_layers, s_churn

      • Where:

        • CLIP_stop_at_last_layers is Clip Skip

        • s_churn is sigma churn

PDF documentation is in the works.


Check out my other models

SDXL

SD1.5

LoRA

Questions or Feedback?

Visit my thread on the Unstable Diffusion Discord Server