Overview

Flux.2-Klein-9B-MatchingPose is a LoRA adapter for FLUX.2 Klein 9B that transfers pose from a mannequin reference image onto a character while preserving the character's identity, face, and style. Created by nhathoangfoto, this model solves a critical problem in pose transfer: standard ControlNet approaches either lose character identity during pose enforcement or fail to match poses accurately. The model works by using a faceless mannequin as a clean pose anchor, allowing the character's identity to fill in without drift. It requires the diffusers library and loads as a .safetensors file through the Flux2Pipeline. The recommended inference setup uses 4 steps for distill models or 20 steps for the base model, with LoRA strength between 0.9 and 1.1 and guidance scale from 1 to 4.0. This model is the second stage of a two-stage pipeline, paired with the companion Mannequin LoRA that generates the clean pose reference in stage one.

Best use cases

Character pose transfer for game and animation assets. When you need to place a specific character (identified by face and clothing) into exact poses for game sprites, animation frames, or storyboarding, this LoRA excels because it decouples pose from identity. The mannequin reference encodes only skeletal structure and proportions, so your character renders into that structure without facial distortion or identity bleed. This is far more reliable than text-based pose description or attempting ControlNet-based pose transfer on realistic photos, which typically corrupts facial features.

Fashion and product photography with consistent models. If you have a signature character or model and need them in multiple poses for catalog or social media work, this pipeline guarantees consistent identity across all pose variations. Generate your desired pose as a mannequin reference once, then apply it to your character multiple times. Each output preserves the exact same face and styling while only the body position changes, making batch generation of pose variations trivial and maintaining brand consistency.

Narrative storyboarding with pose-locked characters. For comic panels, graphic novels, or storyboards where specific characters must appear in predetermined poses across multiple frames, this model ensures your character's appearance stays fixed while their pose matches your compositional reference. Unlike trying to describe poses in natural language (which conflicts with the LoRA's pose-matching behavior), you control poses through actual image references, eliminating ambiguity and reducing iteration cycles.

Avoiding identity collapse in pose-conditional generation. Standard approaches to pose control in diffusion models either use multi-ControlNet stacks (slow, unreliable) or force poses through text prompts (results in character inconsistency and identity loss). This LoRA solves this by leveraging the mannequin as an identity-free proxy. The two-stage workflow—Mannequin LoRA first, then MatchingPose second—systematically removes identity information from the pose encoding, allowing cleaner character insertion.

Limitations

Requires clean mannequin references. This model produces best results only when paired with mannequin images generated from the companion Mannequin LoRA. Using real photographs as pose references introduces identity bias into the pose encoding, degrading output quality. Dirty, cluttered, or multi-subject mannequin images reduce pose transfer sharpness. You cannot simply extract poses from arbitrary photos; you must generate explicit mannequin references first, adding an extra pipeline step.

Fails on extreme poses. Acrobatic, heavily contorted, or extreme poses reduce transfer accuracy noticeably. The model was trained on standard poses, so handstands, backbends, or unusual flexibility-dependent positions may produce anatomically incorrect limb placement or misaligned joints. Standard upright, sitting, and moderate-angle poses work consistently; beyond that, accuracy degrades.

Face and hand detail quality. While identity is preserved, face and hand details sometimes require secondary fixing or upscaling at high resolution. The LoRA operates on 9B base model with 28 inference steps at most, limiting fine detail capture in facial features and finger articulation. You may need additional face restoration or hand-specific refinement passes for commercial-quality output.

Single-subject training limitation. The model trained primarily on single-character images. Multi-person scenes, crowd shots, or complex multi-subject compositions produce inconsistent pose matching or may fail to apply the reference pose correctly to all subjects. This is not a limitation of the LoRA itself but reflects its training distribution.

Body proportion normalization. If your character description implies a drastically different body type than the mannequin reference—for example, describing a child when the reference mannequin is adult-proportioned—the model normalizes proportions toward the mannequin template rather than creating extreme size variance. You must match character description gender, age, and build to the mannequin's proportions to avoid this.

No commercial license clarity issues, but base model licensing applies. This LoRA releases under Apache 2.0, which permits commercial use. However, the base model FLUX.2 Klein 9B has its own separate license terms that you must review and comply with. Check those base model license conditions before deploying this adapter commercially.

Inference speed modest for real-time use. With 20 steps on the base model, generation takes longer than ultra-lightweight alternatives. Exact timing depends on hardware, but this is not suitable for real-time interactive applications or batch processing thousands of images under tight latency budgets.

How it compares

Flux.2-Klein-9B-SmartCharacterSwap handles character and face swapping with occlusion awareness, excelling at preserving lighting and color grading from target scenes. Choose MatchingPose if your primary goal is exact pose control across characters; choose SmartCharacterSwap if you need seamless face or character replacement while matching scene lighting and handling hands/occlusions. SmartCharacterSwap is more suited to photo-realistic compositing; MatchingPose is built for pose consistency across character variations.

Flux2-Klein-9B-Consistency improves overall Klein consistency without trigger words or cues, focusing on character stability across multiple generations. If your need is consistent character appearance in multiple unrelated poses, Consistency LoRA handles that with no extra reference images required. MatchingPose requires explicit pose references but guarantees exact pose matching; Consistency improves general character retention but does not enforce specific body positions.

Flux-2-Multi-Angles-LoRA-v2 generates multiple viewing angles of the same subject. Choose MatchingPose if you need precise pose control with identity preservation; choose Multi-Angles if you want automatic viewpoint variation of a subject without manual pose reference creation. Multi-Angles is faster for batch angle generation; MatchingPose gives you absolute control over exact limb positioning.

Flux2-Klein-9B-True-V1 is a fine-tuned undistilled version of FLUX.2-Klein 9B offering clearer, more realistic outputs with precise editing and reduced detail collapse compared to distilled versions. If you need superior base image quality and don't need pose control, True-V1 is a better foundation model. MatchingPose is specialized for pose transfer specifically; True-V1 improves general realism but does not address pose consistency.

refcontrol-FLUX.2-klein-9B-reference-pose-lora fuses reference image identity with pose control maps (skeleton/keypoints), using the trigger phrase "apply pose from image 1 with reference from image 2." This is conceptually similar to MatchingPose but accepts keypoint maps directly rather than mannequin images. Choose MatchingPose if you prefer working with full mannequin renders; choose RefControl if you have pose skeletons or keypoint data and want to avoid the intermediate mannequin generation step. RefControl offers more flexibility for technical pose input; MatchingPose is simpler for visual reference workflows.

Technical specifications

Architecture and base model: This is a LoRA (Low-Rank Adapter) for FLUX.2 Klein 9B Base, developed using the ai-toolkit training framework by Ostris. It does not modify the base 9B parameter model itself but adds trainable low-rank matrices to specific layers, keeping the adapter file size small while achieving specialized behavior.

Input requirements: Requires two inputs—a character reference image (photograph, AI-generated, or illustration containing the target face and identity) and a mannequin pose reference image (best generated from the companion Flux.2-Klein-9B-Mannequin LoRA). Text prompt with the trigger word matchingpose9b placed at the beginning.

Output format: Single image output at the base model's native resolution (typically 1024×1024 or 768×768 depending on FLUX.2 Klein base configuration).

Trigger word: matchingpose9b—must be placed at the start of the positive prompt to activate pose-matching behavior.

Recommended inference settings:

  • LoRA strength: 0.9 to 1.1 (slightly stronger activation than typical LoRAs)

  • Inference steps: 4 steps for distill models, 20 steps for base model

  • Guidance scale: 1.0 to 4.0

  • Framework: Diffusers library with Flux2Pipeline

File format: Distributed as .safetensors file for safety and compatibility.

License: Apache 2.0 for the LoRA itself. Base model FLUX.2 Klein 9B has separate license terms requiring separate review.

Training framework: ai-toolkit by Ostris.

Companion model: Designed to pair with Flux.2-Klein-9B-Mannequin for the complete two-stage pipeline.

Model inputs and outputs

Inputs

  • Character reference image:Any photograph, AI-generated character, or illustration containing the face and identity you want to preserve. No strict resolution requirement, but cleaner images produce better results.

  • Mannequin pose reference image:Full-body faceless mannequin render, ideally generated from the companion Mannequin LoRA, on clean white or minimal background.

  • Text prompt:Description of the character (appearance, clothing, style) beginning with the trigger word- matchingpose9b. Avoid redundant pose descriptions; let the mannequin reference handle body position.

Outputs

  • Single image:Generated character in the exact pose of the mannequin reference, with identity and face preserved from the character reference, rendered at base FLUX.2 Klein resolution. Natural skin, clothing, and lighting (not stiff mannequin appearance).

Getting started

```
from diffusers import Flux2Pipeline
import torch

Load base FLUX.2 Klein 9B model

pipe = Flux2Pipeline.from_pretrained(
"black-forest-labs/FLUX.2-klein-base-9B",
torch_dtype=torch.bfloat16
).to("cuda")

Load the MatchingPose LoRA

pipe.load_lora_weights(
"nhathoangfoto/Flux.2-Klein-9B-MatchingPose",
adapter_name="matchingpose"
)
pipe.set_adapters(["matchingpose"], adapter_weights=[1.0])

Generate image with pose matching

image = pipe(
prompt="matchingpose9b, a young woman with long black hair, wearing casual clothes, studio lighting",
num_inference_steps=20,
guidance_scale=3.5,
).images[0]
image.save("output.png")
`` For ComfyUI workflows, download the.safetensorsfile, place it inComfyUI/models/loras/`, and reference it in your node setup with the trigger word in your positive prompt.

Frequently asked questions

Q: Can I use this model commercially?

A: Yes. The LoRA itself is licensed under Apache 2.0, which permits commercial use. However, you must also comply with the separate license terms of the base FLUX.2 Klein 9B model, which you should review on its official repository.

Q: Do I need the Mannequin LoRA to use MatchingPose?

A: Yes, for best results. The model was designed as stage two of a two-stage pipeline with the companion Mannequin LoRA generating clean pose references in stage one. Using real photographs as pose references instead introduces identity bias and degrades pose transfer quality. The two-stage workflow is the recommended approach.

Q: How much VRAM do I need to run this?

A: Since this loads into an existing FLUX.2 Klein 9B pipeline, you need enough memory for that base model plus the LoRA weights. Exact requirements depend on your hardware and inference settings, but FLUX.2 Klein 9B typically requires 8–16 GB of VRAM for inference on consumer GPUs. The LoRA adds minimal overhead.

Q: Why does my character's proportions not match the pose reference?

A: If your character description implies a significantly different body type than the mannequin (e.g., child character with adult mannequin), the model normalizes toward the mannequin template. Always match your character description gender, age, and build to the mannequin's proportions to avoid this.

Q: Can I use real photos as pose references instead of mannequins?

A: You can, but results degrade because real photos encode identity features that conflict with the character you want to render. The mannequin approach works because it is identity-free—only pose and proportions. For reliable pose transfer, generate mannequin references from the companion Mannequin LoRA first.

Q: How do I prompt this model correctly?

A: Begin your prompt with the trigger word matchingpose9b, then describe the character's appearance, clothing, and style. Let the mannequin reference handle the pose itself—avoid describing body position in text, as this can conflict with the reference image and degrade results. Focus on "who" the character is, not "what pose" they are in.

Q: Is this model still maintained?

A: The model was released by nhathoangfoto, who actively develops AI models and accepts collaboration requests. Check the model repository for the latest updates and community discussions.

Q: What happens if my mannequin pose is extreme or contorted?

A: Extreme poses (acrobatic stances, heavy backbends, unusual flexibility) reduce transfer accuracy. The model trains on standard upright, sitting, and moderate-angle poses. For unusual positions, pose matching becomes less reliable and may produce anatomically incorrect limb placement. Stick to standard poses for best results.