jimmycarter/krea2-turbo-bbox

Merged turbo (few-step distilled) transformer(s) for a Krea-2 finetune. Built as turbo_epoch = epoch_checkpoint + (krea/Krea-2-Turbo - krea/Krea-2-Raw) — the finetune's epoch weights carry the base→turbo distillation delta so they can be sampled in 8 steps, CFG-free.

Each epoch/EMA variant lives under its own <epoch_name>/transformer (or <epoch_name>/ema_<decay>/transformer) subfolder in this repo. Latest variant pushed: epoch-4-step-20910 (raw), at epoch-4-step-20910/transformer.

This is a transformer-only repo (everything else — VAE, text encoder, tokenizer, scheduler — is identical to krea/Krea-2-Raw). Load a variant as the transformer of a stock Krea-2 pipeline via subfolder:

import torch
from diffusers import Krea2Pipeline, Krea2Transformer2DModel

tf = Krea2Transformer2DModel.from_pretrained(
    "jimmycarter/krea2-turbo-bbox", subfolder="epoch-4-step-20910/transformer", torch_dtype=torch.bfloat16)
pipe = Krea2Pipeline.from_pretrained("krea/Krea-2-Raw", transformer=tf, torch_dtype=torch.bfloat16)
pipe.to("cuda")

# These are DISTILLED weights — sample at turbo settings:
image = pipe("a photo of a cat", num_inference_steps=8, guidance_scale=0.0).images[0]
image.save("out.png")

Note: is_distilled is a pipeline-level flag (model_index.json), not a transformer config field, so it can't be baked into this transformer-only upload. The pipeline above uses the fixed distilled timestep shift (mu=1.15) when it treats the checkpoint as distilled; if your pipeline build doesn't infer that, pass mu=1.15 / set is_distilled=True explicitly.