This fantastic model was NOT created by me. It was created by yang520333. You can find the original at HuggingFace. This is from the step 5000 or 5600 model recommended by the author. The other models can be downloaded from HuggingFace. It seems to work well in a super resolution workflows. It worked well for instance, in flux-3-image-klein-upscale. I have also used it on both SFW and NSFW generations.

FLUX.3 Image Klein · Natural Facial Skin Texture LoRA

Add natural facial skin texture to portraits.

Workflow update — October 29, 2026: The new Natural Facial Texture workflow is now recommended. In the author's tested images, it resolved the positional drift observed with the previous workflow, avoided the previously observed Chinese-text corruption, and substantially improved overall quality. The previous workflow remains available. Reported 4K generation time is approximately 3 minutes with the new workflow, compared with 1 minute 40 seconds previously. See the workflow comparison and performance sections below.

Model overview

This image-editing LoRA adds natural-looking facial skin texture, subtle pores, and fine microtexture to portraits with overly smooth, untextured, or weakly textured skin. Its goal is to introduce natural skin detail while retaining the subject's identity, facial features, expression, skin tone, lighting, and composition as closely as possible.

• Training base: black-forest-labs/flux-3-image-klein-base.

• Model type: An image-editing LoRA adapter. A compatible base model and its supporting components must be loaded separately.

• Weight format: Safetensors, BF16, rank 64, exported with AI Toolkit.

• Available checkpoints: 14 training checkpoints are currently available.

• Recommended checkpoints: 5,000 and 5,600 training steps are the author's recommendations and the versions tested most extensively. Other checkpoints have received less testing. The degree of fit differs across training stages, so their results may vary; they are provided as alternatives rather than equally validated recommendations. A later training step does not necessarily produce a better result; compare them for your input and preferred texture intensity.

• Trigger words: No additional custom trigger word is required. Describe the desired edit in natural language.

Using the LoRA in ComfyUI

1. Install or update ComfyUI and the required nodes to versions that support FLUX.3 Image Klein editing and this LoRA format.

2. Place a selected .safetensors file under ComfyUI/models/loras/, optionally in a subfolder, and refresh the model list.

3. Import a compatible image-editing workflow. Start with the recommended workflow below, select your local model components and LoRA, and load the portrait you want to edit.

4. Describe the subtle skin texture you want to add and the identity, facial features, skin tone, and lighting you want to retain. Adjust the LoRA strength for the image and compare results.

Both supplied workflows retain their original JSON contents. Importing a workflow does not install its model files or custom-node dependencies. The local input images and temporary comparison previews referenced by the workflows are not included.

The author's commonly used prompt:

The following is the author's exact prompt, preserved verbatim:

Add realistic facial skin details and natural micro-texture to the character in image 1 while preserving the character's identity, facial features, skin tone, expression, lighting, and all other image content.

Workflow update — October 29, 2026

Recommended: Natural Facial Texture workflow

Download the recommended workflow

According to the author, this workflow was configured from a built-in ComfyUI template, while the previous workflow was configured from an example in the official repository. Both files include the author's LoRA and other settings; they are not presented here as unmodified official originals.

The author's latest observations:

• The previous workflow showed positional drift, occasional Chinese-text corruption, and some loss of image quality.

• In the author's tested images, the new workflow resolved the previously observed positional drift, did not show the same Chinese-text corruption, and substantially improved overall quality.

• Standard 1K and 2K generation times remain approximately the same. At 4K, generation takes approximately 3 minutes, compared with 1 minute 40 seconds with the previous workflow.

These are author-reported observations from the tested images. This release update did not rerun inference or perform an independent pixel-alignment or Chinese OCR evaluation. Resolving positional drift does not mean that every pixel value remains unchanged after adding texture, nor does it guarantee zero drift or perfect text preservation for every input.

Saved configuration:

• LoRA: 5,000-step checkpoint, strength 1.0.

• KSampler: 16 steps, CFG 1, Euler / simple, denoise 1.

• Seed behavior: Randomized.

• Reference handling: ReferenceMultiLatent set to keep resolution.

Reselect your own portrait and locally installed model files after importing. The saved model components are listed below. Some download hints embedded in node properties refer to different quantized variants; check the files actually selected in the loaders.

Component selected in the workflowCommon model directoryflux3_image_klein_int8_convrot.safetensors — image modelComfyUI/models/diffusion_models/qwen3vl_4b_bf16.safetensors — text encoderComfyUI/models/text_encoders/qwen3.5_9b_flux3_prompt_upsampler_int8_convrot.safetensors — prompt upsamplerComfyUI/models/text_encoders/flux2-vae.safetensors — VAEComfyUI/models/vae/Flux3 image klein Natural facial skin texture_000005000.safetensors — this adapterComfyUI/models/loras/flux3 image klein/ matches the saved relative path

The base-model components must be obtained separately. Only the facial-texture adapters are distributed in this repository.

Custom nodes referenced by the recommended workflow:

• YxwCreateBoundingBoxesClassic: Yxw052/ComfyUI-ClassicBoundingBoxes.

• Image Comparer (rgthree): rgthree-comfy.

The new workflow does not reference the previous workflow's easy showAnything node. Install missing dependencies through ComfyUI Manager where available, or follow the corresponding projects' installation instructions. The JSON does not include their program code.

Confirmed differences between the two workflow files

The following is a static comparison of the supplied JSON files, not a conclusion about why an upstream developer made a change.

SettingPrevious box-edit workflowRecommended facial-texture workflowReference preprocessingReferenceMultiLatent: resize, approximately 1MPReferenceMultiLatent: keep resolution, encoding the supplied dimensionsOutput latent dimensionsA separate EmptyFlux2LatentImage, with dimensions supplied through an output-size subgraphThe reference node's empty_latent output, matching the first encoded reference latent's shapeExplicit sampling-shift patchIncludes ModelSamplingFlux; shift comes from log(alpha), with alpha=16 saved in the main interfaceDoes not contain that explicit patch nodeSaved sampling steps30 in the main interface, overriding the subgraph's internal default16Sampler / schedulerEuler / simpleEuler / simple

The ComfyUI reference-node implementation distinguishes these reference-processing modes: resize scales the reference before VAE encoding, while keep resolution encodes it at its supplied dimensions. The empty_latent output is the first reference latent multiplied by zero, retaining its shape. It does not pass the original image content directly into the sampler as a populated initial latent.

A plausible explanation, not yet isolated experimentally: Keeping a larger reference latent at 4K can increase computation and memory demands while retaining more fine reference detail, which may help preserve small text and structure. Resizing the reference to approximately 1MP may discard some of that detail before generation. However, the workflows also differ in output-size handling, sampling shift, and saved step count. The observed quality and timing changes cannot be attributed to one setting without a controlled comparison.

For an input near 1MP, the reference-processing difference is usually smaller. Actual 2K and 4K timings also depend on output dimensions, quantization, offloading, and other settings.

Unverified hypothesis: The author wonders whether the difference is related to an official fix for malformed limbs or anatomy. No matching upstream change, commit history, or controlled test has been established. This remains a hypothesis; the confirmed findings are the configuration differences above and the author's reported results.

Previous: Box-edit workflow, retained for reference

Download the previous box-edit workflow

This JSON remains unchanged. It includes subgraphs and requires a ComfyUI version that supports those subgraphs and FLUX.3 nodes. Its node metadata records core version 0.42.0. The release checks cover JSON structure, dependency references, and file hashes; the workflow was not rerun as part of this publication update.

Saved main-interface configuration: The 5,000-step facial-texture LoRA at strength 1.0, with 30 sampling steps, Euler / simple. These are the file's saved values, not universal recommendations. Main-interface selections override internal subgraph defaults; verify the actual loaded components after importing.

Its main interface selects the same model component filenames listed above. Although the previous workflow's filename contains bf16, its saved image-model selection is INT8 ConvRot. The workflow filename alone does not establish the precision used during inference.

Custom nodes referenced by the previous workflow:

• YxwCreateBoundingBoxesClassic: Yxw052/ComfyUI-ClassicBoundingBoxes.

• Image Comparer (rgthree): rgthree-comfy.

• easy showAnything: comfyui-easy-use.

After importing, select your own portrait in Image 1, then use the bounding-box canvas if needed. The saved local input filenames and temporary comparison-preview URLs are not included as assets. Bounding boxes guide the model's edit region; they do not guarantee unchanged pixels outside the boxes.

Historical guidance retained in the previous workflow's Note:

• Undistilled Base: Compare LoRA strengths from 0.5 to 1.0; the author commonly uses 0.6.

• Distilled model: Try starting at 0.7 or higher.

• Sampling: The older Note recommends 30–50 steps. This historical recommendation does not override the new workflow's saved 16-step configuration.

• Training data: The author reports 78 paired images at 1024 × 1024.

• Earlier high-resolution observations included positional drift and weaker texture enhancement at 4K. Refer to the current update above for the author's newer results.

These are author-provided observations and suggestions, not settings independently retested for this release. The original Note remains intact in the previous JSON.

VRAM and generation-time reference

The following figures are the author's reported experience and measurements. They were not independently benchmarked for this publication and are not universal minimum requirements.

VRAMPreviously reported experience8 GBBarely workable12 GBAdequate16 GBComfortable operation, including 4K output

RTX 5070 Ti, distilled model, warm start:

Reported output resolutionPrevious box-edit workflowRecommended facial-texture workflowStandard 1K, previously described as 1024 resolutionApproximately 11 secondsApproximately unchanged, around 11 secondsStandard 2KApproximately 35 secondsApproximately unchanged, around 35 seconds4KApproximately 1 minute 40 secondsApproximately 3 minutes

Warm start means subsequent generation after the model has loaded; first-load time is excluded. The labels 1K / 1024, 2K, and 4K follow the author's descriptions; exact width and height were not provided for each timing.

Both workflow JSON files are available, but the timings were not accompanied by complete per-run parameters and logs. The saved JSON settings may not match every timed run, so this is not a controlled benchmark at identical settings. Quantization, offloading, sampling settings, and input dimensions affect both speed and memory use. In particular, the previous VRAM experience should not be treated as a guarantee for the new workflow when retaining a full-resolution 4K reference.

Training and inference checkpoints are different: This LoRA was trained on Klein Base. The timings above come from the author's distilled-model tests. They are not timings for the undistilled Base checkpoint, and this publication did not validate every adapter checkpoint and inference configuration individually.

Limitations

• The added texture is generated. It does not recover or establish the subject's actual pore pattern.

• Retaining identity, skin tone, lighting, and composition is an editing goal rather than a guarantee of unchanged pixels.

• The author's improvements in drift, Chinese-text preservation, and quality describe the tested images, not a universal guarantee for all inputs.

• Texture intensity depends on the input, checkpoint, prompt, and inference settings. If you see rough or repetitive texture, color shifts, or excessive sheen, reduce the LoRA strength or compare another checkpoint.

License and attribution

This LoRA is based on FLUX.3 Image Klein Base. The base weights are released under the FLUX Kommunity License. Use and distribution are subject to LICENSE.md and its applicable terms. Public availability for download does not imply an Apache 2.0 license or unrestricted commercial permission; consult the full license.

This is a community LoRA published by yang520333, not an official Black Forest Labs product. No endorsement, approval, or validation by Black Forest Labs is claimed. The author modified the base model's behavior through a LoRA adapter to add facial skin texture. This repository contains adapter weights rather than the complete base-model weights. See NOTICE.md for attribution and the modification notice.