A character-transfer LoRA for FLUX.2-klein that, unusually, drives both the un-distilled foundation (FLUX.2-klein-base-9B) and the step-distilled inference variant (FLUX.2-klein-9B). The same file gives a quality-first path (28 inference steps on the base) and a speed-first path (4 inference steps on the distilled), so you can iterate fast and ship slow with one set of weights.

What it does

Given a stitched conditioning image (source 960×544 on the left, reference 544×544 on the right), the model regenerates the 960×544 source half with the reference character's identity transferred onto the source's pose, composition, and background. See the input-stitching section below for the exact layout.

Quick start (diffusers, foundation base, 28 steps — quality path)

python

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-base-9B",
    torch_dtype=torch.bfloat16,
).to("cuda")

pipe.load_lora_weights(
    "24yearsold/flux2-klein-character-transfer-portable-r128",
    weight_name="pytorch_lora_weights.safetensors",
)

cond_image = build_stitched_conditioning(source_path, reference_path)  # 1504x544 RGB
prompt = open("character_transfer_prompt.txt", encoding="utf-8").read()  # ~300 CJK chars
out = pipe(
    image=cond_image,
    prompt=prompt,
    num_inference_steps=28,
    guidance_scale=3.5,      # un-distilled foundation honours real CFG
).images[0]
out.save("transferred_quality.png")

Quick start (diffusers, distilled base, 4 steps — speed path)

python

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-9B",   # <-- distilled variant
    torch_dtype=torch.bfloat16,
).to("cuda")
pipe.load_lora_weights(
    "24yearsold/flux2-klein-character-transfer-portable-r128",
    weight_name="pytorch_lora_weights.safetensors",
)
out = pipe(
    image=cond_image,
    prompt=prompt,
    num_inference_steps=4,    # <-- 4 steps matches the distilled recipe
    guidance_scale=1.0,
).images[0]
out.save("transferred_fast.png")

You can adjust LoRA strength on either path with pipe.set_adapters("default", adapter_weights=0.85) (anything 0.0..1.0+). The LoRA was trained at α=1.0; 0.8–1.0 is the recommended range.

Input stitching

+--------------------+----------+
|     source 960x544 | ref 544² |    -> 1504 × 544 RGB
|  (center-cropped)  |  (white- |
|                    |  padded) |
+--------------------+----------+

Reference is PAD-to-square with white fill (composites RGBA over white if the reference has alpha), then resized to 544×544. Source is center-cropped (or letterboxed) to 960×544. Concatenate left-to-right. Output is the regenerated 960×544 left half.

Why two bases work with one LoRA

FLUX.2-klein-9B (distilled) and FLUX.2-klein-base-9B (foundation) share the same transformer architecture (joint_attention_dim=12288, 8 dual-stream

24 single-stream blocks, attention_head_dim=128, same VAE, same Qwen3 text encoder). They differ only in the trained weight values: the distilled variant has step-and-guidance distillation baked in for 4-step inference; the foundation variant is the un-distilled weights.

This LoRA was fine-tuned on the foundation, which exposes the full gradient signal — and the resulting attention-delta turns out to attach cleanly to either weight base. We verified this with a full all-checkpoint side-by-side eval (see the W&B runs below).

Usage in ComfyUI

ComfyUI ≥ v0.10.0 (released 2026-01-20) loads this LoRA directly via the standard Load LoRA node thanks to Comfy-Org/ComfyUI #11981, which added FLUX.2's to_qkv_mlp_proj and to_out key mapping to comfy/utils.py:flux_to_diffusers. Drop the safetensors into models/loras/ and use either the foundation or the distilled klein-9B as your Load Checkpoint source. Strength 0.8–1.0.

Training recipe (P13)


Base	`black-forest-labs/FLUX.2-klein-base-9B` (un-distilled foundation)
Rank	128 (alpha=128)
Init	gaussian (NOT PiSSA)
Target modules	`to_k/q/v/out.0 + to_qkv_mlp_proj + single_transformer_blocks.{0..23}.attn.to_out`
Trainable params	~222 M
Schedule	cosine to 0.25× peak, 4 056 steps over 6 epochs
LR	2 × 10⁻⁴ (warmup 50, num_cycles=0.3333)
Effective batch	3 (3 GPU × bsz 1 × no grad-accum)
Precision	bf16
Weighting scheme	`none` (uniform σ-sampling; NOT logit_normal)
Hardware	3× RTX 4090 24 GB
Wall clock	~5.5 h training + ~1.5 h eval/diagnose

Held-out val_loss at ck-4056 (per-σ MSE, 8 held-out triplets)

u	loss
0.10 (texture)	0.3252
0.25	0.2294
0.50	0.2484
0.75	0.3591
0.90 (struct)	0.5480
mean	0.3420

Trajectory was still descending at endpoint, so an 8-epoch follow-up is the natural next move if you want a slightly better LoRA.

Reference renders

samples/p13_lora_on_distilled_ablation.png — 8 triplets × 4 columns: foundation-no-LoRA, foundation+LoRA (28 steps), distilled-no-LoRA (4 steps), distilled+LoRA (4 steps). The two no-LoRA columns confirm neither base does character transfer on its own; the LoRA does all the work and the quality/speed gap between the two LoRA columns is modest.

samples/triplet_0..7_foundation_28steps.png — finals on the foundation base samples/triplet_0..7_distilled_4steps.png — finals on the distilled base

Wandb (side-by-side, same LoRA, two bases)

Foundation @ 28 steps, all 24 ckpts × 8 triplets: https://wandb.ai/team-rocket/flux-klein-character-transfer/runs/jj5qaxrn
Distilled @ 4 steps, all 24 ckpts × 8 triplets: https://wandb.ai/team-rocket/flux-klein-character-transfer/runs/78ynegnw
Direct 4-col ablation (the headline image): https://wandb.ai/team-rocket/flux-klein-character-transfer/runs/tql9zwhd

Related models from the same project

24yearsold/flux2-klein-9b-character-transfer-lora-r128 (P12) — earlier LoRA trained against the distilled klein-9B. Lower MSE on the held-out triplets but tied to the distilled base; ships its own inference recipe.
24yearsold/flux2-klein-9b-character-transfer-pissa-r64 (P7) — original PiSSA r=64 merged release.

This portable r=128 release supersedes both for new work.

License

Apache 2.0. The base models (black-forest-labs/FLUX.2-klein-base-9B, black-forest-labs/FLUX.2-klein-9B) have their own licenses; please follow theirs when using.