Credits:
This model was trained and Finetuned by silveroxides
You can find all their models and work on Hugging Face here: https://huggingface.co/silveroxides/models
30GB Version of black-forest-labs/FLUX.2-dev 64.4GB
You can find workflow with samples in the workflow_assets folder.
https://huggingface.co/silveroxides/FLUX.2-dev-fp8_scaled/tree/main/workflow_assets
The workflow contains the information and links needed to get started with using this model.
This fp8_scaled model is faster than the official one released by ComfyOrg when used with the Loader in the workflow.
The custom node that loads the model in the workflow is necessary to obtain fastest inference on lower VRAM GPU.
Description
FAQ
Comments (6)
So this is NOT the fp8-mixed that they released on hf?
No, it's a different one. The checkpoint "fp8_scaled" means it uses FP8 precision quantization with dynamic scaling to improve efficiency. This enhances VRAM usage and performance on GPUs. The Comfy-Org flux2_dev_fp8mixed version—I'm not sure what the "mixed" means, mixed of what?—weighs 35.5 GB, while this one weighs 30GB.👌
Is there any difference in output quality compared to the official Comfy FP8 mixed model?
Hello, the text_encoders you shared are the abrided version marked "small" made by ComfyUI, not the complete version of text_encoders from the studio!
Comfyui crashing due to low system ram. Have 64GB DDR5, doesn't seem to be enough for a 30GB model. Using 17GB mistral 3 small text encoder.
Is there a NSFW lora for these bigger models?