Revival

A Retro Anime Art Model.

This is a small anime model I created in-between working on larger projects.

SD 1.5 checkpoint

SDXL models are in-the-works. I'm taking my time to create better dataset tools and a new XL trainer.

Note about sample prompts: I used this model as a testbed for another project I'm working on. An LLM SD prompt generator extension for Automatic1111 (standalone version is also planned). Many, of the prompts used in the sample images above were generated using this system. More info on the prompt generator when it's closer to release.

Questions/Feedback/Updates?

Visit my thread on the Unstable Diffusion Discord

Settings

Quick Settings :

Settings -> User interface -> Quicksettings list
sd_model_checkpoints sd_vae CLIP_stop_at_last_layers s_chrun always_discard_next_to_last_sigma

ETA Noise Seed Delta :

Settings -> Sampler parameters -> Eta noise seed delta
31337

Sampler Method :

DPM++ SDE Karras
- Sampling steps = 25 - 40
- CFG = 7 - 9
Euler a
- Sampling steps = 20 - 30
- CFG = 6 - 7.5

These are just recommendations. Feel free to experiment with other samplers and different parameters.

Hires Fix

Models :

4x_foolhardy_Remacri - Download here
4x-UltraSharp - Download here
lollypop - Download here

Install location : stable-diffusion-webui -> models -> ESRGAN

Hires Fix Settings :

For all ESRGAN models (which the above are) I use the following settings :

Upscale by
- 1.5 if resolution is > 512x768
- 1.6 if resolution is <= 512x768
Denoising strength
- 0.25 - 0.35
Hires steps
- If sampling steps > 60 then,
- Hires steps = half of sampling steps.
- for example, sampling steps = 80, then Hires steps = 40.