flux2-klein-4b-stamp-filter

LoRA adapter for unsloth/FLUX.2-klein-4B that turns a photo into a grainy, print-textured field-notes illustration on a white background.

This is an image-conditioned (image-to-image) adapter. It was trained on photo → illustration pairs with the photo fed to the model as reference tokens, so at inference you must pass a condition image — plain text-to-image with this adapter is not the intended use.

Trigger word: stamp

Example

Input photo (left) and the adapter's output (right), held-out photo, 4 sampling steps.

Usage

import torch
from diffusers import Flux2KleinPipeline
from PIL import Image

pipe = Flux2KleinPipeline.from_pretrained(
    "unsloth/FLUX.2-klein-4B", torch_dtype=torch.float16
)
pipe.load_lora_weights("vladimir94/flux2-klein-4b-stamp-filter")
pipe.to("mps")  # or "cuda"

cond = Image.open("photo.jpg").convert("RGB")
out = pipe(
    image=cond,
    prompt="stamp",
    num_inference_steps=4,  # Klein is step-distilled; 4 steps is enough
    guidance_scale=1.0,     # ignored by step-distilled models anyway
).images[0]
out.save("stamp.png")

Output size follows the condition image's aspect ratio (training used aspect-preserving resize, no cropping).

Training details

• Base model: unsloth/FLUX.2-klein-4B (Flux2Transformer2DModel)
• Task: paired image-to-image — condition photo appended as extra latent tokens, loss on target tokens only
• Dataset: 233 photo → isolated-drawing-on-white pairs, all captioned stamp
• Rank / alpha: 8 / 8, targeting to_qkv_mlp_proj, to_q, to_k, to_v, to_out.0
• Resolution: up to 768×768 total area, aspect preserved, downscale-only, no center crop
• Steps: 700 (~3 epochs), lr 1e-4 constant, fp16 (full LoRA, not quantized), batch 1
• Hardware: Apple Silicon (MPS), ~3 hours
• Final mean loss: 0.60

Limitations

• Requires a condition image; without one the output is unpredictable.
• Best on landscape/nature photos similar to the training set (field-notes subjects).
• 4-step sampling only — the base model is step-distilled; more steps can wash out detail.

Framework versions

• PEFT 0.18.1
• diffusers 0.40.0.dev0