| --- |
| license: other |
| license_name: flux-kommunity-license |
| license_link: LICENSE.md |
| base_model: |
| - black-forest-labs/flux-3-image-klein-base |
| - black-forest-labs/flux-3-image-klein |
| pipeline_tag: text-to-image |
| tags: |
| - lora |
| - comfyui |
| - flux-3 |
| - flux-3-image-klein |
| - distillation |
| --- |
| |
| # FLUX 3 Image Klein distillation LoRAs |
|
|
| The step-distilled FLUX 3 Image Klein checkpoint minus the base checkpoint, as LoRAs at four |
| ranks. Put one on the base model with the **Flux 3 Flow Map Patch** node from |
| [WAS Node Suite](https://github.com/WASasquatch/was-node-suite-comfyui), and the base samples |
| the way the distilled model does in 16 steps: the distilled picture from the same seed at CFG 1, |
| and more contrast and colour at CFG 2.5 to 3.5, which the distilled model does not take. |
|
|
| > 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 modified from FLUX 3 Image Klein by WAS. See [License](#license). |
|
|
|  |
|
|
| One prompt, seed 42, 1280x2048, euler / simple, LoRA strength 1.0. Top row: the distilled model |
| at CFG 1 and 16 steps, and the base model at CFG 3.5 and 50 steps. Each row below is the base |
| with one LoRA and Flux 3 Flow Map Patch, 16 steps, at CFG 1, 2.5 and 3.5. The figure under each |
| title is PSNR against the distilled render. |
|
|
| ## Files |
|
|
| Put them in `ComfyUI/models/loras`. |
|
|
| | File | Rank | Size | PSNR against the distilled render, CFG 1 | |
| |---|--:|--:|--:| |
| | `flux3_image_klein_distill_r256.safetensors` | 256 | 1.35 GB | 27.3 dB | |
| | `flux3_image_klein_distill_r64.safetensors` | 64 | 337 MB | 22.2 dB | |
| | `flux3_image_klein_distill_r32.safetensors` | 32 | 169 MB | 20.8 dB | |
| | `flux3_image_klein_distill_r16.safetensors` | 16 | 84 MB | 19.0 dB | |
|
|
| The PSNR figures come from the one prompt and seed in the grid. |
|
|
| ## What else you need |
|
|
| ComfyUI 0.41 or later, and a WAS Node Suite release that includes **Flux 3 Flow Map Patch**. From |
| [black-forest-labs/flux-3-image-klein-comfy](https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy): |
|
|
| | File | Goes in | |
| |---|---| |
| | `flux3_image_klein_bf16.safetensors`, the base model | `ComfyUI/models/diffusion_models` | |
| | `flux3_image_klein_sd_bf16.safetensors` or `flux3_image_klein_sd_int8_convrot.safetensors`, the distilled model | `ComfyUI/models/diffusion_models` | |
| | `qwen3vl_4b_int8_convrot.safetensors` | `ComfyUI/models/text_encoders` | |
| | `flux2-vae.safetensors` | `ComfyUI/models/vae` | |
|
|
| Flux 3 Flow Map Patch reads only two weights from the distilled file, about 20 MB, so it never |
| loads the distilled model itself. |
|
|
| ## Using them in ComfyUI |
|
|
| 1. **Load Diffusion Model** on `flux3_image_klein_bf16.safetensors`. |
| 2. **LoraLoaderModelOnly** on one of these LoRAs, strength 1.0. |
| 3. **Flux 3 Flow Map Patch**, with `checkpoint` set to a distilled file, one with `_sd_` in its name. |
| 4. **KSampler** at 16 steps, `euler`, `simple`. CFG 1 gives the distilled picture; CFG 2.5 to 3.5 |
| adds contrast and colour. |
|
|
| Example workflow: |
| [`flux3-klein-distill-lora.json`](https://github.com/WASasquatch/was-node-suite-comfyui/blob/main/docs/workflows/flux3-klein-distill-lora.json). |
| It renders one seed at CFG 1 and CFG 2.5 into an image compare. |
|
|
| ## Why the patch |
|
|
| The distilled checkpoint has a layer the base does not: `ref_time_in`, which embeds the sigma |
| each step lands on and adds it to every modulation. A LoRA can only change layers that already |
| exist, so it cannot carry that one. Without the patch, base + LoRA renders with a fine stripe |
| every 4 pixels down the rows and a blue cast. Flux 3 Flow Map Patch reads the two `ref_time_in` |
| weights from a distilled file and attaches them to the base model. |
|
|
| ## How they were made |
|
|
| For every layer, the distilled weight minus the base weight, computed in fp32 from the two bf16 |
| checkpoints. Each matrix is a truncated SVD at the file's rank, stored as fp16 |
| `lora_up`/`lora_down` with no value clipping. The 86 norm scales are exact fp32 differences. |
| `ref_time_in` is not included, since the base has no layer for it. |
|
|
| ## License |
|
|
| These LoRAs are a Derivative of FLUX 3 Image Klein and are distributed under the |
| [FLUX Kommunity License v1.0](LICENSE.md). Any rights to use FLUX 3 Image Klein and these LoRAs |
| are granted to you directly by Black Forest Labs Inc. under that License. It permits |
| non-commercial use, and lets Qualifying Users use Outputs commercially on the conditions it sets |
| out; read it before use. |
|
|
| They were modified from FLUX 3 Image Klein by WAS, and are not an official product of Black |
| Forest Labs Inc., nor endorsed, approved or validated by it. They are provided as is, without |
| warranty of any kind, and WAS accepts no liability arising from their use. The Attribution Notice |
| is in [NOTICE.md](NOTICE.md). |
|
|
