Synthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Synthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets".
Paper summary: Jane: Now, wrapping up this discussion on "Synthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets", what's the simple version of why we should care about this work?
Tom: It boils down to the fact that they showed a model trained on chronic wound photos can generate synthetic leprosy images that are just as diverse as real ones, which is a very strong result because it’s such low-data stuff <ref:2609.13226#pg2> >
Lu: The authors flag a limitation here, which they mention in section three point three, noting that many of the training images on both sides of the transfer carry large synthetic flat regions from preprocessing artifacts > <ref:2609.13226#pg2>
Meng: That's a big caveat for me. If those flat regions are just artifacts from preprocessing, how reliable are we when we try to deploy these generated images for actual clinical use?
Jane: It means that while the diversity numbers look good, they need to ensure those synthetic artifacts don't introduce misleading visual information when the model is used in a real-world setting >
Lalam: Lalam is confident that based on what they measured, the absence of mode collapse in the seven hundred eight-image fine-tuning regime is the central positive result of this study > <ref:2609.13226#pg1>
Tom: So, to summarize, this paper shows that chronic wound photography is a viable donor domain for leprosy lesion synthesis because low-level appearance transfers well across it >
Jane: Exactly. The remaining barrier they identified isn't image quality itself but achieving good semantic control over the final output images <ref:2609.13226#pg2> >
Conclusion: Tom: So we're wrapping up this look at that paper, "Synthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets." Basically, they took a massive set of chronic wound photos and used them to train a diffusion model to create synthetic leprosy images.
Jane: And the big takeaway is that this model can generate new leprosy images that are just as diverse as the real ones they started with. It’s not just making blurry copies; it’s actually capturing the variety in skin tone and lesion shapes from those wound photos.
Lu: The central finding, which I think is really interesting, is that this transfer learning works surprisingly well for low-level appearance details like texture and boundary shapes because chronic wounds share a lot of visual statistics with leprosy lesions.
Meng: But they did admit there’s a catch when you look at the training data pipeline. They mentioned that most of their training images from both sides had these big synthetic flat regions from the preprocessing steps, which is something we have to be careful about when deploying this stuff.
Tom: That’s fair, Meng. The authors are being transparent about where their training signal comes from and what those artifacts look like. It tells us that while the model learns the surface texture well, we need to watch out for how it handles those synthetic areas in a real scenario.
Jane: So, if you're driving and you see an image generated by this system, what does that mean for you? It means it’s visually convincing on the surface, but we still have to consider if those synthetic flat areas might look weird when used in a clinical setting.
Lalam: From my side as the model, I can confirm that the internal diversity of the generated set is statistically indistinguishable from the real leprosy images they tested against. The absence of mode collapse in this fine-tuning setup is what they are really highlighting here.
Tom: Yeah, so they're not just showing a cool visualization; they’re proving that chronic wound photography is a viable starting point for synthesizing medical images, provided we manage those synthetic artifacts carefully.
Jane: It shifts the focus from just "can we make realistic images?" to "how do we control the semantic details of these generated images?" because they found that was still a bit of a barrier.
Lu: Exactly. The next step isn't just making it look good; it’s figuring out how to better guide that diffusion process so we can control the actual medical features, not just the surface appearance.
Meng: So, the paper suggests that while this transfer works for low-level stuff, we need to figure out a better way to inject semantic control into the model itself if we want it to be truly useful beyond just visual realism.
Tom: That’s the next frontier then. The paper sets up the foundation by proving the transfer works, but now they have to tackle that deeper level of control.
Jane: And that leads us perfectly into what they suggest for future work—how to engage this multi-scale conditioning mechanism they mentioned earlier in a way that gives better control over the final output.
cs.CV
Submitted: 2026-09-01
Updated: 2026-11-18
Comments: 15 pages, 12 figures, 6 tables. Code: https://github.com/Yusufa09/SyntheticImageGeneration. v2: retrained with a held-out test split and an expanded evaluation (KID, paired comparisons, confidence intervals). Fine-tuning now measurably shifts outputs toward real leprosy images; this supersedes the base-vs-fine-tuned comparison and conclusion in v1, which had an evaluation error
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 90/100
The gist: Synthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets asks whether a model trained on abundant chronic wound photography…
Key concepts
- Mask-Conditioned Latent Diffusion Model
- This is an AI model that generates images by starting with random noise and iteratively refining it based on specific instructions. In this study, the generation process is guided by a 'mask'—a map detailing where lesions are located. The model learns to create realistic textures and shapes specifically within the boundaries defined by this mask.
- Transfer Learning
- This technique involves taking a model that has already been trained on one large task (like analyzing thousands of wound photos) and repurposing its knowledge for a new, related task (like generating leprosy images). The model uses the general visual understanding it gained from the wound data to help generate realistic features for the target leprosy images.
- Mode Collapse
- Mode collapse occurs when a generative model fails to produce a variety of different outputs, instead only generating a small subset of possibilities. The study found no mode collapse; its internal perceptual diversity matched the real leprosy set, meaning the synthetic images are varied enough and not stuck producing just one type of image.
- LPIPS (Learned Perceptual Image Patch Similarity)
- LPIPS is a metric used to measure how perceptually similar two images are by comparing their features extracted from a pre-trained neural network. It assesses similarity based on what a human might perceive, rather than just pixel differences. The study used this to confirm the synthetic images look visually similar to real leprosy photos.
Terminology
Summary
Synthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets asks whether a model trained on abundant chronic wound photography transfers to this low-data regime. p. 0 The generated set shows no mode collapse: its internal perceptual diversity (0. p. 0662) is statistically indistinguishable from that of the real leprosy set (0. p. 0672, 95% CI [0. p. 0 p. 0664, 0.680]). Chronic wound photography is therefore a viable donor domain for leprosy lesion synthesis: low-level appearance transfers well, and the remaining barrier is semantic control rather than image quality. p. 0
Pruning redundant reasoning tokens
The paper presents a three-stage pipeline designed to synthesize synthetic leprosy images using transfer learning from large chronic wound datasets. This pipeline involves first automatic mask generation using a DeepLabV3-ResNet50 segmentation network on two wound datasets that ship without them. p. 3 Second, it assembles a mask-conditioned latent diffusion model from Stable Diffusion 1. p. 05 components and trains it on 3,280 region-of-interest wound crops. p. 0 Third, it fine-tunes this model on 708 leprosy image–mask pairs drawn from 764 images of approximately 150 patients. p. 0
How the training signal is built
The training process for the diffusion model involves several specific conditioning signals and architectural choices. The UNet input convolution was widened from 4 to 11 channels to admit three mask feature maps and a blurred low-frequency context latent. p. 0 Mask features consist of the binary mask together with two Euclidean distance transforms, one measuring distance inward from the boundary and one outward, each normalized to [0, 1] per crop. p. 3 The context hint carries low-frequency information like skin tone and illumination direction after being Gaussian-blurred with radius 6. p. 3 The model is trained using a standard noise-prediction objective under a DDPM schedule [12] reweighted by Min-SNR loss weighting. p. 4
Where the gains come from
The success of the transfer stems from how low-level appearance transfers well, as chronic wound photography shares statistics like skin tone and irregular lesion boundaries. The paper notes that fine-tuning only marginally shifted the output distribution, which is traced to lesion geometry reaching the network through input concatenation alone. p. 0 The evaluation uses LPIPS perceptual distance anchored by a real-versus-real baseline computed on the same 242 anchor images to interpret cross-set comparisons. p. 0
Evaluating the results
The study evaluates performance using LPIPS [27] which measures perceptual distance between images as a weighted L2 distance in the feature space of a pretrained network. The generated set matches the real set’s diversity, with fine-tuned images on average 0. p. 7662 apart from one another against 0. p. 7672 for real leprosy images. p. 7 The domain gap is quantified as a real-to-fine-tuned distance of 0. p. 8716, which sits outside the baseline confidence interval but under half of one standard deviation. p. 8
Limitations and future directions
The study notes several limitations, including the domain scope being limited to chronic ulceration rather than the full Ridley–Jopling spectrum. Furthermore, the conditioning mechanism is loose, as geometry reaches the network through input concatenation alone rather than a multi-scale path that was implemented but not engaged. p. 5 Future work suggests engaging the multi-scale conditioning and ablating it to test the mechanism for loose mask adherence. p. 5 The paper also highlights that a majority of training images on both sides of the transfer carry large synthetic flat regions from preprocessing artifacts. p. 9
Conclusion
A mask-conditioned latent diffusion model trained on 3,280 wound crops and fine-tuned on 708 leprosy crops generates images whose internal perceptual diversity (0.662 LPIPS) is statistically indistinguishable from that of the real leprosy corpus (0. p. 0672, 95% CI [0.664, 0.680]). The absence of mode collapse in a 708-image fine-tuning regime is the central positive result. p. 7 Chronic wound photography is therefore a viable donor domain for leprosy lesion synthesis, low-level appearance transfers well across it, and the remaining barrier is semantic control rather than image quality. p. 0 The broader premise holds up. p. 1
Data and Code Availability
Code for preprocessing, segmentation, diffusion training, and evaluation is available at https://github.com/Yusufa09/SyntheticImageGeneration. p. 0 All image data are from public sources cited above: the Kaggle Wound Segmentation Images collection [14], the AZH Wound Database [25, 2], WoundsDB [5], and CO2Wounds-V2 [23]. p. 1 No new patient data were collected. p. 0
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Transfer low-level appearance effectively to data-scarce domains by using chronic wound photography as a donor domain for synthesis of leprosy lesions, as
low-level appearance transfers well.
-
Implement a three-stage pipeline involving automatic mask generation, mask-conditioned latent diffusion pretraining on 3,280 wound crops, and fine-tuning on limited leprosy data to generate synthetic images whose internal perceptual diversity matches the real set's diversity (0.662 LPIPS vs 0.672).
-
Use a real versus real LPIPS baseline computed on the same anchor set as cross-set comparisons to interpret generated image distances, stating that
without this reference, cross-set LPIPS values cannot be interpreted at all.
-
Improve semantic control by considering multi-scale conditioning mechanisms and explicitly testing them to see if they provide better geometric adherence than simple input concatenation alone, which was found to be the effective mechanism in the reported run.
Sources
- Synthesizing Diabetic Foot Ulcer Images with Diffusion Model
- Rethinking Atrous Convolution for Semantic Image Segmentation
- DermGAN: Synthetic Generation of Clinical Skin Images with Pathology
- Beyond Labels: Zero-Shot Diabetic Foot Ulcer Wound Segmentation with Self-attention Diffusion Models and the Potential for Text-Guided Customization
- Classifier-Free Diffusion Guidance
- CO2Wounds-V2: Extended Chronic Wounds Dataset From Leprosy Patients
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs