Hybrid precision rebuild of krea/Krea-2-Turbo for ComfyUI.
This checkpoint keeps the main Krea-2-Turbo FP8-scaled generation backbone intact, but replaces the internal txtfusion 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 Changed
Changed to BF16:
- txtfusion.layerwise_blocks.0.* attention/MLP linears
- txtfusion.layerwise_blocks.1.* attention/MLP linears
- txtfusion.refiner_blocks.0.* attention/MLP linears
- txtfusion.refiner_blocks.1.* attention/MLP linears
Already BF16 in the FP8-scaled model:
- txtfusion.projector.weight
- txtmlp.*
- textfusion norm / QK-norm scale tensors
Kept FP8:
- shared DiT blocks.*
- main image-generation backbone
## Why
The FP8-scaled checkpoint stores many txtfusion 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 Details
For each replaced txtfusion linear layer:
- FP8 .weight tensor was replaced with the corresponding BF16 tensor from the official BF16 checkpoint.
- .weight_scale sidecar was removed.
- corresponding _quantization_metadata["layers"] entry was removed.
This makes the selected txtfusion layers load as normal BF16 weights instead of quantized FP8 weights.
## Intended Use
For ComfyUI users experimenting with Krea-2 text-conditioning behavior, especially workflows or custom nodes that inspect, patch, or scale the txtfusion path eg. ComfyUI-Krea2T-Enhancer.
## 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 Model
Derived from:
https://huggingface.co/krea/Krea-2-Turbo
## License
Subject to the Krea 2 Community License and upstream acceptable-use requirements.
Description
v 1.0.0
FAQ
Comments (10)
That's some nice technical detail, but but did all that do? What's improved here?
Nothing is supposed to be improved here under normal conditions. This is an edge-case fix for my node: when it scales those targeted layers in an FP8 E4M3 model, the quantized math can become unstable and stop producing the intended effect. The change simply keeps the node functioning correctly in that specific case; it is not a general image-quality improvement.
@Capitan01R sorry what even "krea 2 turbo" is ??
@amazingbeauty it is a newly released text to image model
And then released as FP8, not even a BF16 version in sight.
@mrmrswiggly612 this model is NOT BF16, maybe read the description as it says which layers were changed to BF16........... only the text-fusion layers and the txt mlp related to them
@Capitan01R Yes, I know. I was hoping to prompt you into a BF16 version. Guess I was too subtle.
@mrmrswiggly612 No problem, the BF16 version is already available in the official HF repo
@Capitan01R is it worth it ? nearly every 2 weeks new model and new finetune and all lacks for few and they replace each others like chess. it's follow prompt or can general realistic photos ? or nothing from both ?
@amazingbeauty the model is actually good compare to the newly released text to image models plus it's very fast in term of it/s, overall it is very good indeed.