rockerBOO/flux.1-krea-dev-convrot-int8

FLUX.1 Krea [dev] (INT8 ConvRot)

INT8 (row-wise, ConvRot) quantized version of black-forest-labs/FLUX.1-Krea-dev, a text-to-image model co-developed by Black Forest Labs and Krea, built on the FLUX.1 [dev] architecture.

Files

File	Size	Source
flux1-krea-dev-int8convrot.safetensors	13.2 GiB (14.1 GB)	flux1-krea-dev.safetensors (22.2 GiB / 23.8 GB)

Only the transformer/DiT checkpoint is quantized here — text encoders (CLIP-L, T5-XXL) and VAE are not included in this repo and should be loaded from the upstream repo as usual.

Requirements

Unlike NVFP4/MXFP8, INT8 row-wise quantization runs on any CUDA GPU — no Blackwell required. ConvRot (Hadamard rotation) improves quantization quality at row-wise granularity and needs Triton for its inference kernels.

Load with ComfyUI (or any loader supporting .comfy_quant metadata for INT8 row-wise + ConvRot weights) as a diffusion model checkpoint.

Quantization method

Quantized with the convert-to-quant (ctq) CLI tool using INT8 row-wise quantization with ConvRot (group-wise Hadamard rotation, group size 256) and Comfy-quant format. The --flux2 exclusion filter was reused here since its layer-name patterns (guidance_in, time_in, img_in, txt_in, final_layer) happen to match FLUX.1's naming too, keeping those layers high-precision. --exclude-layers was added on top to also keep the per-block modulation linears (img_mod/txt_mod in the double blocks) high-precision, since FLUX.2's stream_modulation key pattern doesn't match FLUX.1's img_mod.lin/txt_mod.lin naming.

Steps

1. Download the checkpoint (single file, no sharding):

   hf download black-forest-labs/FLUX.1-Krea-dev flux1-krea-dev.safetensors
   

2. Quantize directly:

   uv run --with convert-to-quant --with triton --with safetensors --with tqdm --with numpy ctq \
     -i flux1-krea-dev.safetensors \
     -o flux1-krea-dev-int8convrot.safetensors \
     --int8 --scaling_mode row --convrot --flux2 --exclude-layers 'img_mod|txt_mod' \
     --comfy_quant --save-quant-metadata --simple --low-memory
   

3. Delete the raw bf16 file, keeping only the quantized output.

flux1-krea-dev-int8convrot.safetensors: 1316 tensors, 268 layers quantized to INT8 row-wise + ConvRot (verified via the _quantization_metadata header) across all 19 double blocks and 38 single blocks. High-precision (unquantized) layers: img_mod/txt_mod (double blocks), guidance_in, time_in, img_in, txt_in, final_layer.

Flags used:

• --int8 — INT8 quantization instead of FP8
• --scaling_mode row — per-row scale factors (required for ConvRot)
• --convrot — group-wise Hadamard rotation (group size 256, default) to reduce INT8 quantization error
• --flux2 — exclusion filter; its key patterns (guidance_in, time_in, img_in, txt_in, final_layer) happen to match this FLUX.1-based checkpoint's layer names
• --exclude-layers 'img_mod|txt_mod' — additionally keep per-block modulation linears high-precision
• --comfy_quant — Comfy quantization tensor/metadata format
• --simple — skip SVD optimization, use simple quantization
• --save-quant-metadata — embed _quantization_metadata in the safetensors header
• --low-memory — stream tensors during quantization to reduce RAM usage

License

FLUX.1 [dev] Non-Commercial License (included in this repo), inherited from black-forest-labs/FLUX.1-Krea-dev / black-forest-labs/FLUX.1-dev.

Non-commercial use only — see the license for full terms, including restrictions on commercial use, Distribution requirements, and the required Attribution Notice.