RDBT [NetaYume]

Recalibrated distribution

Do NOT download TCFP8 version. I messed up with the metadata. Huge stability issues. I will delete this version.

Use the new uploaded bf16 version.

This model is part of the test theories to improve diffusion models.

Trained from NTYM4 with ~70k images.

Aiming for

• Better textures and art details.

• Better and stable prompt coherence.

• Balanced contrast and lighting. Never overflow/oversaturated.

Guide

Prompt: Basically the same as NetaYume. Except:

• Style prompt is required. This model does not have default style. The default tv anime style in NetaYume has been nuked.

• Use "Digital anime art style by @xxxx." at the end of the prompt to prevent Gemma 2 paying too much and incorrect attention to the artist name.

• Quality tags are not needed. Dataset has higher quality than avg "masterpiece".

• You don't need tons of tags to describe a character. Just use the most unique ones. e.g. "elf girl frieren, fox girl tamamo \(fate\)". See: img.

• Prefer simple natural language at the start, and tags at the end.

Settings:

• Timesteps shift 3~4.5 for better details. (from node ModelSamplingAuraFlow).

• CFG scale: 1. Although CFG 1~1.5 is doable, if you want.

• Sampler: Prefer euler_a + normal.

About CFG distilled model:

• You can't control CFG scale and negative prompt. Those are trained inside the model.

• CFG scale = 1 is a special value. It means disabling CFG and neg prompt.

• Because you don't need to run a forward pass for the negative prompt, you can generate 2x faster.

Some training details

Total dataset contains ~70k images. Not equally weighted.

Only layers.[2:25] were trained.

Captions are mainly from Gemini. Natural language only, no tags.

Not a LoRA this time?

Multi stage training. No LoRA.

Versions

v0.1 cfg distilled: bf16 full model.

v0.1 cd tcfp8: (has issues, do not download, will be deleted soon) cfg distilled, also a tensorcorefp8 version for ComfyUI.