b0n3-merge — Anima DiT Checkpoint

A merged Anima DiT checkpoint combining six carefully weighted LoRAs, with a focus on realistic image generation. No restrictions apply, great NSFW capability granted.

The stack includes two separate turbo LoRAs — Anima Turbo LoRA v0-2 and TURBO for ANIMA v1 — which were merged with extra care: standard conflict-resolution (TIES/attenuation) was disabled for this merge after testing showed it was cancelling out the turbo LoRAs' step-distillation effect. The result is a model that retains fast, low-step inference behavior alongside the realism and detailing contributions from the rest of the stack.

On top of the standard merge, TURBO for ANIMA's norm-layer patches (which live outside the usual LoRA key format) were applied separately as direct weight deltas.

Recommended: lower step counts to get the most out of the baked turbo behavior:

I get very good results with a combination of ER-SDE sampler and beta57 scheduler at 12 steps with a config of 1.1. The example images were generated with these settings.

LoRA's baked in with strength:

- anima_realism_enhancer v2 — 0.5

https://civarchive.com/models/2643318/anima-realism-enhancer

- Anima RL v0-1 — 0.8

https://civarchive.com/models/2583128/anima-rl

- Anima Turbo LoRA v0-2 — 0.8

https://civarchive.com/models/2560840/anima-turbo-lora

- l3n4i Anima v2 — 0.6

experimental unpublished lora

- TURBO for ANIMA v1 — 0.2

https://civarchive.com/models/2619830/turbo-for-anima-less-steps?modelVersionId=2941393

- ZodaPlus Detailer v1 — 0.8

https://civarchive.com/models/2632704/zodaplus-detailer

Thanks to the authors of the above lora's. Your work is much appreciated and enables other dabbling users like me to start out with something beautiful and creative!

Merged using the [Anima LoRA Merge GUI](https://civarchive.com/models/2484986) by hybskgks28275.