What for?

SDXL-based models (NoobAI-XL) converted to Rectified flow / RF and more:

• Mugen | RF + Flux 2 VAE conversion – CivitAI / 🤗-Link
  • no support on Machina's sd-scripts because no Flux2 VAE support (yet?)
• ChenkinRF | based on ChenkinNoob Eps v0.2; RF conversion only, default SDXL VAE – CivitAI / 🤗-Link
  • compatible with Machina's sd-scripts and other trainers supporting RF objective for SDXL
• Shitty experiment | RF + Anzhc's EQ-VAE – CivitAI
  • compatible with Machina's sd-scripts and other trainers supporting RF objective for SDXL as well as VAE reflection (optional but will improve image border edges)

Hard requirements

• torch_linear_assigment or SciPy
• A working venv, you can take the one from LoRa_Easy_Training_Scripts/backend/sd-scripts/venv and install whatever is missing.

Additional Commands

• --flow_model Required to train into Rectified Flow target

• --flow_use_ot ot = Optimal Transport; Not needed but use anyway, want info? ask lodestone

• --flow_timestep_distribution=uniform|logit_normal

  • In any case either should work
  • Mugen: logit_normal suggested
  • ChenkinRF: used a mix of logit_normal (early training) and uniform, logit_normal suggested
  • Shitty experiment: was trained with uniform.

• --flow_logit_mean needed by logit_normal timestep_distribution

  • Mugen (and others probably): -0.2

• --flow_logit_std needed by logit_normal timestep_distribution

  • Mugen (and others probably): 1.5

• Shift, choose one

  • --flow_uniform_static_ratio allows static values for shift, default 2.5
    • rough suggested values:
      • ~6.0 – Mugen
      • 2.5 – ChenkinRF
      • ~2.5 – Shitty Experiment
  • --flow_uniform_base_pixels default 1048576 (1024x1024), allows dynamic shifting for >1024 training
  • --flow_uniform_shift allows dynamic shifting

---

• --vae_type=flux2 Required for Mugen; Self explanatory, can work with Flux 1 too

• --latent_channels=32 Required for Mugen; Self explanatory, use 16 for Flux 1 if you into that

• --vae_custom_scale Mugen / Flux 2 VAE: 0.6043 | Anzhc eq-vae: 0.1406

• --vae_custom_shift Mugen / Flux 2 VAE: 0.0760 | Anzhc eq-vae: -0.4743

• --vae_reflection_padding suggested for Anzhc's eq-vae, my shitty experiment wasn't trained with this

Misc options

• --contrastive_flow_matching Magic thingie that makes training results a bit sharper | Mugen didn't use this

  • --cfm_lambda needed by the one above, default 0.05, I prefer 0.02

• --skip_existing skips latents check, avoids wasting hours checking ++10M latents

• --use_sga attaches Lodestone's stochastic accumulator for gradient accumulation; currently borked

You can use launch_all_train.sh to start training, or copy the command to an activated venv, works on windows or linux w/e the file itself has a working template

Info on the shitty test

(For some info on Mugen, see the readme in the HuggingFace repo)

Trained with --flow_use_ot --flow_timestep_distribution uniform --flow_uniform_static_ratio 2.5 AdamW8bit Kahan Summation at 2e-5, caption and tag dropout of 0.1 with keep token separator based on NoobAI's own dataset, for one combined epoch of 486k unique images, scheduler was constant with warmup with 10% warmup.

Dataset

Uses Deepghs' danbooru webp from 0000 to 0065 with the exact same tags as NoobAI's own run extra files not present in NoobAI's dataset were purged.

Why tho?

Just sharing the codebase used to train the Rectified Flow model, so both people that know can scrutinize it, and people that think they do can shit on it.

How to train?

Activate venv, install missing shit, edit config and dataset toml as per your needs run launch_all_training.sh, or run accelerate manually, simple as that.