toxicdog/FLUX.3-Image-Klein-INT4-ConvRot-ComfyUI

FLUX 3 Image Klein INT4 ConvRot for ComfyUI

Repackaged model files for ComfyUI and others, with INT4 ConvRot converted for testing in Radiant Canvas for fast macOS inference: Radiant Canvas on the App Store.

This is an unofficial, modified version of black-forest-labs/flux-3-image-klein (the distilled checkpoint). It is not an official Black Forest Labs product and has not been endorsed, approved or validated by Black Forest Labs. The ComfyUI BF16 transformer used for this conversion came from black-forest-labs/flux-3-image-klein-comfy.

File

• diffusion_models/flux3_image_klein_sd_int4_convrot.safetensors: the distilled FLUX 3 Image Klein transformer in ComfyUI's native ConvRot W4A4 format.

For ComfyUI, place the file in:

ComfyUI/models/diffusion_models/

Use a recent ComfyUI build with FLUX 3 Image Klein and native INT4 ConvRot support, and load it with the standard diffusion model loader in place of flux3_image_klein_sd_bf16.safetensors. The text encoder, prompt upsampler and VAE are not included here. Get them, and the workflows, from black-forest-labs/flux-3-image-klein-comfy. That repository recommends always using a prompt upsampler.

Conversion details

• Original model revision: 337337621b8526e120f2275d250242f4a6bedd75
• ComfyUI repack revision: 5c9b714851dc072b04a23ef6d2d36421f5ccb3a0
• Source: flux3_image_klein_sd_bf16.safetensors
• Layout: same tensor names as the official flux3_image_klein_sd_int8_convrot.safetensors. The q, k and v projections are fused into one qkv weight per block, and the per-block comfy_attention.config (INT8 attention where supported) is kept.
• Quantized layers: 172 (qkv, attn_out, mlp_in and mlp_out in all 5 text, 5 image, 5 image-condition and 28 single-stream blocks), the same layer set as the official int8 ConvRot file
• Quantization: comfy-kitchen 0.2.37 TensorCoreConvRotW4A4Layout
• ConvRot group size: 256
• Quantization group size: 64
• Output size: 3,190,202,531 bytes
• SHA-256: 1a8270e15d9e55c538e203f6869d72ed805215818395325e9b810aeb2a8f545b

Input embeddings, time embeddings, modulation, final layer and normalization weights are kept in BF16. The file loads in ComfyUI as a FLUX 3 model with fused QKV and ConvRot W4A4 layers.

This is a test conversion. Check output quality for your own prompts and workflows before relying on it.

License

This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License. Any rights to use the FLUX Model and these Derivatives are granted directly by Black Forest Labs Inc. to you under that license. See NOTICE.md for the attribution notice and a description of the modifications.

Copyright Black Forest Labs Inc. IN NO EVENT SHALL BLACK FOREST LABS INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.