Flux 3 Klein Distilled LoRA (SD Extract)

LoRAs extracted from FLUX 3 Image Klein: the step-distilled (SD) checkpoint minus the base checkpoint, at ranks 256, 64, 32 and 16. On the base model they give the distilled model's look in 16 steps, and unlike the distilled model, the base still takes CFG.

No trigger words. Use at strength ~1.0 (seems to work at 0.9 with extra steps)

You need the Flux 3 Flow Map Patch

The distilled model has a time-step layer that the base model does not (ref_time_in), and a LoRA cannot add layers. Without it, the base with this LoRA renders with a fine horizontal grain every 4 pixels and a blue cast. The Flux 3 Flow Map Patch node in WAS Node Suite v3 reads that layer (about 20 MB) from the distilled checkpoint and attaches it to the base model. The grain is gone with it.

Setup in ComfyUI

1. Load Diffusion Model: flux3_image_klein_bf16.safetensors (the base model)

2. LoraLoaderModelOnly: one of these LoRAs, strength 1.0

3. Flux 3 Flow Map Patch: checkpoint set to flux3_image_klein_sd_bf16 or flux3_image_klein_sd_int8_convrot

4. KSampler: 16 steps, euler, simple

   • CFG 1: the distilled picture from the same seed

   • CFG 2.5 to 3.5: more contrast and colour

Which rank

File Size At CFG 1 flux3_image_klein_fp16_distill_r256 1.3 GB Closest to the distilled model flux3_image_klein_fp16_distill_r64 330 MB Very close; the best balance flux3_image_klein_fp16_distill_r32 160 MB Same scene, a little more drift flux3_image_klein_fp16_distill_r16 80 MB Clean, but the composition wanders

With CFG above 1 the four ranks come out much closer to each other. See the comparison grid in the gallery.

Other files you need

From black-forest-labs/flux-3-image-klein-comfy:

• flux3_image_klein_bf16.safetensors in models/diffusion_models

• flux3_image_klein_sd_bf16.safetensors or flux3_image_klein_sd_int8_convrot.safetensors in models/diffusion_models (the patch reads only its two time-step weights)

• qwen3vl_4b_int8_convrot.safetensors in models/text_encoders

• flux2-vae.safetensors in models/vae

These LoRAs are also on Hugging Face: WAS/flux3_klein

License

This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License. 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.

These LoRAs are a Derivative of FLUX 3 Image Klein, modified by WAS, and distributed under the FLUX Kommunity License v1.0. Any rights to use them are granted directly by Black Forest Labs Inc. under that license.

• Allowed use: non-commercial use.

• Commercial images: Qualifying Users, with gross annual revenue under US$5,000,000, may use generated images commercially, on the license's conditions.

• No endorsement: these LoRAs are not an official Black Forest Labs product and are not endorsed by them.

• No warranty: they are provided as is.