Flux2 Distilled 8step ComfyUI BF16 Fp8mixed

FLUX.2 8-Step Distilled Transformer for ComfyUI (BF16 + FP8 Mixed)

This repository contains two ComfyUI-ready transformer checkpoints converted from baseten/distilled_8step_FLUX.2-dev.

It is a transformer-only conversion package. It does not include the text encoder, tokenizer, scheduler, or VAE from the upstream Diffusers repository.

Files

File	Precision / format	Size	SHA256
flux2_distilled_8step_bf16.safetensors	BF16 ComfyUI transformer	61 GB	b4508322f3d522f829d716897fe676d5a6b119a35198263778efddfae92b9a7b
flux2_distilled_8step_fp8mixed.safetensors	ComfyUI FP8 mixed transformer	34 GB	3e3650964c921cd66760c2387f141678877d686f33aa3afb0a13ae4967b4e05d

Provenance

The upstream weight source used for this conversion was:

• Source model: baseten/distilled_8step_FLUX.2-dev
• Local snapshot commit: 81118b71f648071cdf3c1a9f1387b8aa50cc4f07
• Upstream format: Hugging Face Diffusers Flux2Pipeline
• Upstream transformer path: transformer/diffusion_pytorch_model*.safetensors

The upstream Baseten model is an 8-step distilled FLUX.2-dev derivative. Its repository metadata describes the transformer as a distilled 8-step FLUX.2-dev checkpoint. This repository does not claim a new training run, a new base model, or a LoRA; it only repackages the upstream transformer into ComfyUI-compatible tensor names and precision formats.

The broader model lineage should be understood as:

1. black-forest-labs/FLUX.2-dev as the original FLUX.2-dev family model.
2. baseten/distilled_8step_FLUX.2-dev as the upstream distilled Diffusers checkpoint.
3. This repository as a ComfyUI transformer conversion of that Baseten checkpoint.

What Was Changed

The conversion maps the Diffusers transformer keys into the ComfyUI FLUX.2 naming layout used by recent ComfyUI builds.

Main conversion steps:

• Converted Diffusers transformer tensor names to ComfyUI FLUX.2 tensor names.
• Added FLUX.2 global modulation mappings:
  • double_stream_modulation_img.linear.weight -> double_stream_modulation_img.lin.weight
  • double_stream_modulation_txt.linear.weight -> double_stream_modulation_txt.lin.weight
  • single_stream_modulation.linear.weight -> single_stream_modulation.lin.weight
• Mapped timestep and guidance embedder layers to ComfyUI:
  • time_guidance_embed.timestep_embedder.linear_1.weight -> time_in.in_layer.weight
  • time_guidance_embed.timestep_embedder.linear_2.weight -> time_in.out_layer.weight
  • time_guidance_embed.guidance_embedder.linear_1.weight -> guidance_in.in_layer.weight
  • time_guidance_embed.guidance_embedder.linear_2.weight -> guidance_in.out_layer.weight
• Renamed Q/K norm weights from .weight to .scale where ComfyUI expects scale tensors.
• Saved one direct BF16 checkpoint.
• Saved one FP8 mixed checkpoint using torch.float8_e4m3fn layer quantization and ComfyUI mixed-precision metadata.

The FP8 mixed file was produced by quantizing the layers present in a ComfyUI FLUX.2 FP8 mixed reference checkpoint and preserving ComfyUI-style _quantization_metadata plus input-scale tensor layout. It is intended for lower VRAM usage and faster local iteration, with the usual numerical tradeoffs of FP8 quantization.

Validation

The converted files were checked locally with ComfyUI model detection:

• Detected model class: Flux2
• Image model: flux2
• Double blocks: 8
• Single blocks: 48
• Hidden size: 6144
• Context dimension: 15360
• FP8 mixed: detected as ComfyUI mixed_ops
• Quantized FP8 mixed layers: 128

Usage in ComfyUI

Place either file in:

ComfyUI/models/diffusion_models/

Then select it from the appropriate ComfyUI diffusion model / UNet loader for FLUX.2.

You still need compatible FLUX.2 text encoder and VAE files separately. This conversion was prepared for a ComfyUI setup containing:

• mistral_3_small_flux2_fp8.safetensors in models/text_encoders/
• flux2-vae.safetensors in models/vae/

For the distilled model, start with 8 inference steps and adjust scheduler, guidance, resolution, and VRAM settings according to your workflow.

License and Use

This repository is a derivative conversion of the upstream FLUX.2-dev distilled model and should be used under the same upstream license restrictions and acceptable-use terms. In particular, treat this as non-commercial unless you have the necessary rights from the upstream licensors.

Users are responsible for complying with the licenses and terms of:

• black-forest-labs/FLUX.2-dev
• baseten/distilled_8step_FLUX.2-dev
• Any separate text encoder, tokenizer, scheduler, or VAE files used with these weights

Acknowledgements

All model quality comes from the upstream FLUX.2-dev lineage and the Baseten distilled checkpoint. This repository only provides a practical ComfyUI packaging and precision conversion.