Hybrid precision rebuild of <code>krea/Krea-2-Turbo</code> for ComfyUI.This checkpoint keeps the main Krea-2-Turbo FP8-scaled generation backbone intact, but replaces the internal <code>txtfusion</code> attention/MLP linears with BF16 weights from the official BF16 release.This is not a finetune, LoRA merge, retrain, or new architecture. No new data was used.## What ChangedChanged to BF16:- <code>txtfusion.layerwise_blocks.0.*</code> attention/MLP linears- <code>txtfusion.layerwise_blocks.1.*</code> attention/MLP linears- <code>txtfusion.refiner_blocks.0.*</code> attention/MLP linears- <code>txtfusion.refiner_blocks.1.*</code> attention/MLP linearsAlready BF16 in the FP8-scaled model:- <code>txtfusion.projector.weight</code>- <code>txtmlp.*</code>- textfusion norm / QK-norm scale tensorsKept FP8:- shared DiT <code>blocks.*</code>- main image-generation backbone## WhyThe FP8-scaled checkpoint stores many <code>txtfusion</code> attention/MLP matrices as FP8 tensors with quantization scale sidecars. This rebuild removes FP8 quantization from the textfusion path so ComfyUI text-conditioning experiments can operate on BF16 weights while keeping the memory profile close to the FP8 model.## Technical DetailsFor each replaced <code>txtfusion</code> linear layer:- FP8 <code>.weight</code> tensor was replaced with the corresponding BF16 tensor from the official BF16 checkpoint.- <code>.weight_scale</code> sidecar was removed.- corresponding <code>_quantization_metadata["layers"]</code> entry was removed.This makes the selected <code>txtfusion</code> layers load as normal BF16 weights instead of quantized FP8 weights.## Intended UseFor ComfyUI users experimenting with Krea-2 text-conditioning behavior, especially workflows or custom nodes that inspect, patch, or scale the <code>txtfusion</code> path eg. <a rel="ugc" href="https://github.com/capitan01R/ComfyUI-Krea2T-Enhancer">ComfyUI-Krea2T-Enhancer.</a>## Not Intended As- not an official Krea release- not a finetune- not a LoRA- not a safety bypass- not a claim of improved general quality## Base ModelDerived from:<a target="_blank" rel="ugc" href="https://huggingface.co/krea/Krea-2-Turbo">https://huggingface.co/krea/Krea-2-Turbo</a>## LicenseSubject to the Krea 2 Community License and upstream acceptable-use requirements.

Files with SHA256: 48a31b4ff78c3994a64949f301689fa9451a1922901087a3f8769d213bc343b8