AstolfoRF (3EP) / AstolfoVL (2.5EP) / AstolfoXL (2EP)

Probably the first (and the only) individual Full Finetuning with multi-GPU and... as open-source as it can (may not be copyleft).

LoKR works, but no thanks. 我不做人了，早苗！

Discord: "Good luck".

Specification

• Base model (3EP): AstolfoVL, version 2.5EP (VPRED) RF

• Base model (2.5EP): AstolfoKarMix-XL, version Evo-2EP VPRED

• Base model (2EP): AstolfoKarMix-XL, version NIL1.5 v1.2

• Base model (1EP): AstolfoMix-XL, version 255c

• Tech report: ch06

• Training metrics (tensorboard): HF

• Dataset (images > latents): danbooru2024, e621_2024

• Dataset (tags + captions): meta_lat.json

• 1 step = 16 images, 4x RTX 3090 24G.

• 778k steps for 1EP, 8.0 + 4.6 = 12.6M images

• Tag + NLP caption with A1111 token trick

• Trainer codes: The PR won't be merged

• Train parameter (1EP, 2EP): adamW8bit, UNET 1.5e-6, TE 1.2e-5, BS4 (4 GPU) grad accu 4, 71% UNET (Speed + must underfit)

• Train parameter (2.5EP, 3EP): adamW8bit, UNET 1.5e-5, TE OFF, BS4 (4 GPU) grad accu 4, 100% UNET (finetune to different objective parameters)

• 75-100+ days for 1EP. Train 1 EP only. Save per 10k steps.

• Train result and loss curve: Tensorboard in HF

• Core concept: Unsupervised learning

• Expectation: MID (100% no filter no quality tag) or "reality" "golden mean", proven in Public Arena (ELO 1500). Currently in s3.

• Actual: Need TIPO non empty negative prompt

How to use

• (Reference only, unchanged since 2025) CFG4, Euler, Shift 3.0.

• Train LoRA / merge on top of this model. Compatability should still close to 215c base model. Realistic human content is still supported. "Trust me bro".

• Artist tags may not work, but I did trained. Just dump your "NAI" prompts here.

• Use TIPO to expand tag based prompts with NLP.

• Short tags will suffer from background latent noise. Tags can be observed from E621 or danbooru.

• All images are just seen once. There is no task or KPI to chase, or that omniscient state has been archived. The loss curve is as flat as it neither converge or diverge.

Full docuementation will be published, which is as long as the AstolfoMix series.