InnoVision

InnoVision is a general-purpose base model for anime and semi-realistic image generation. It uses various merging and fine-tuning techniques to allow you to create anything from illustration-style to ¾-realistic art. It responds to danbooru-tag style prompting (e.g. 1girl, from side) as well as some natural-style prompting. Combining both produces best results.

90% of the sample images shown were generated on the first try with some common test prompts, without further attempts, to give an accurate impression of the model's performance. Your results should be better if you re-try and use more advanced prompt strategies.

The benefit of this model is that it supports SD1.5-style "shotgun" prompting. The model also isn't as heavily-reliant on ADetailer/FaceFix as most of the other models I've released, but it is still recommended. Quite frequently the model produces decent hands.

Controlling output style

Use the following prompts in the negative and positive prompt fields as needed:

• anime/anime style/2d/thick lines

• realistic/realism/hyperrealism/3d/photograph/volumetric lighting

Some themes only work better in "anime space" than in "semi-realism" space and vice-versa.

Recommended base negative prompt:

worst quality, low quality, deformed, bad anatomy

More advanced negative prompts give better results.

Web-UI settings used for samples

Basic settings:
Steps: 30
Sampler: Euler a
Scheduler: Automatic
CFG scale: 9 (adjust as needed)
Clip skip: 1 or 2
WxH: 896 x 1280
No Hires fix was applied (or is needed), but you may want to experiment with it

Advanced settings:
Token merging ratio: 0.5
Downcast alphas_cumprod: True

For landscapes I recommend using DPM2 Karras and in your setting “Extra noise multiplier for img2img and hires fix” at 0.07 - use this with img2img at denoise 0.4 and CFG 12 and resize by whatever factor your hardware supports. Works better than upscale models.​

Notes

This model is likely to be a good base for fine-tuning to anime, photorealism, etc since it has aspects of both. Because of the many concepts the model is capable of, the results may be quite good, so please let me know if you use the model in a fine-tune. Personally, I will be using it for my fine-tuning projects so please look forward to those models as well!

This model has not been tested with explicit-NSFW. Your results may vary.

Recipe

A straight 0.5 merge of the excellent AlbedoBase XL - v2.1 and the experimental perturbator model I've released GloryToAllMankind - v1.0 (GTAM)

Perturbation merges are a method I developed where you create a target (usually unusable, exaggerated model) that expands the capabilities of the model you merge it with. Merging the two models created better prompt-adherence than with either one.

Try merging GTAM with your favourite models!