FluxBooru is a series of models that are trained on a collection of SFW booru images, aesthetic photos, and artistic uh.. anatomy. It doesn't know the kind of positioning information that 🐴 does, if you get my drift. This model was not trained on tags. The tags were used as ground truth for captioning via multiple VLMs.

What FluxBooru tries to offer is a better base model for finetuning and inferencing illustrative artworks. This has been a problem area for tuning the base Dev and Schnell models, which resist illustrative training like it's got a counter-intelligence operation going on inside it.

These models received serious amount of compute. 32x H100s were used with 16 of them dedicated to multi-node training, and two other nodes split between various training jobs.

Approx. 2500 H100 GPU hours have been burned on these weights.

• FluxBooru v0.1 is the first release (Dev-based)

  • Requires classifier-free guidance (CFG) at inference time

    • This slows down (ComfyUI impl) or slows it down less but requires more VRAM (batched CFG in Diffusers)

    • Use a value around 5.0 but lower values look less cooked, w/ anatomy issues

    • The Flux Guidance value can't be set to 1.0 but still has some kind of impact.

    • Has obvious issues and mostly serves as an artifact for continued training in the same direction, since it is less biased than v0.2 and was trained in full.

  • Requires ~20-25 steps

  • https://huggingface.co/terminusresearch/flux-booru-CFG3.5

• An attempt was made at de-distilling Schnell

  • but even it seems, 32x H100 isn't enough for that job, and it was abandoned.

  • Not worth testing, doesn't really function at all.

  • Offered incase anybody looking to do the same would like somewhat of a head start.

  • https://huggingface.co/terminusresearch/flux-schnellbooru/tree/bfcf16e7a0d28b0ad394ffa4ab5f62f75eef8ab5/transformer

• FluxBooru v0.2 is a continued training run of v0.1

  • Requires classifier-free guidance (CFG) at inference time

    • Same caveats about speed as with v0.1

    • CFG remains restricted to ~5.0-6.0 with slight improvements in acceptable range

    • Much better results while still not much character name knowledge learnt in such a relatively short training session (1.6 million seen samples, not even 1 full epoch)

  • Requires 20-25 steps

  • https://huggingface.co/terminusresearch/flux-booru-v0.2/tree/main

• FluxBooru v0.2-LoKr 1.6B is a LyCORIS LoKr adapter trained to target the full FluxBooru-v0.2 model

  • Requires classifier-free guidance (CFG) at inference time

    • But can be merged into Flux.1 Dev, and then use without CFG

  • Works in ComfyUI, Diffusers

  • Can be merged into the base FluxBooru-v0.2 model

  • Has much improved reliability over the base v0.2 model as the LoKr adapter trained much faster and more robustly than the full finetuning had

  • Works with an expanded CFG range

  • https://huggingface.co/terminusresearch/FluxBooru-v0.2-LoKr-1.6B

    • The validation samples here are with too-low of a CFG value and are not representative of the final model's quality

• FluxBooru v0.1-LoKr-1.3B is a LyCORIS LoKr adapter trained against the Flux.1 Dev base model

  • Does not require classifier-free guidance at inference time

    • The fastest option out of all of the above.

  • Works in ComfyUI, Diffusers

  • Can be merged into the base Flux.1-Dev or Schnell models

  • Weaker impact and not as fun to use. Might have had an issue during training.