cbrescia

This guide is based on a real-world troubleshooting process to get ComfyUI-Nunchaku working seamlessly with a ComfyUI portable installation. Many users face dependency issues, and this aims to help those "affected by the process." To use the single file svdq-int4_r32-flux.1-dev.safetensorsDisclaimer: This guide isn't official. It's a community-driven effort based on extensive troubleshooting. Always back up your files before making changes.The key innovation here is the application of SVDQuant (https://github.com/mit-han-lab/nunchaku), which optimizes inference for this model far beyond simple quantization, ensuring the preservation of the original flux1-dev quality. Nunchaku isn't just another Flux model accelerator—it's a paradigm shift.Why this guide?The official Nunchaku PyPI release can be outdated, and its direct installation can cause dependency conflicts, especially with filterpy and PyTorch versions. This guide focuses on using a specific development release that resolves these issues.Target Environment:ComfyUI Portable (with embedded Python)Python 3.12PyTorch 2.7.1+cu128 (or similar +cu12x version)NVIDIA GPU Compatibility Notes:NVIDIA categorizes GPU compatibility by architecture, not strictly by series numbers.Blackwell Architecture (expected in RTX 50 series and beyond): This is the architecture that introduces dedicated hardware acceleration for FP4 (via 5th-gen Tensor Cores). Models heavily relying on FP4 for speed will see their full benefits here.Ada Lovelace (RTX 40 series) & Ampere (RTX 30 series): These architectures are highly capable, featuring Tensor Cores with dedicated support for FP8, BF16, and FP16. However, they do not have specific hardware for native FP4 acceleration. While they can process data quantized in INT4 or FP4, they do so through emulation or by converting the data to a precision they do natively support (like FP16/BF16) before calculation.Older series (e.g., RTX 20 series or GTX 16 series): Compatibility for advanced features like INT4/FP4 might be limited or nonexistent, often requiring FP32 or FP16.Step-by-Step Installation Guide:Close ComfyUI: Ensure your ComfyUI application is completely shut down before starting.Open your embedded Python's terminal: Navigate to your ComfyUI_windows_portable\python_embeded directory in your command prompt or PowerShell.Example:Bashcd E:\ComfyUI_windows_portable\python_embeded Uninstall problematic previous dependencies: This cleans up any prior failed attempts or conflicting versions.Bashpython.exe -m pip uninstall nunchaku insightface facexlib filterpy diffusers accelerate onnxruntime -y (Ignore "Skipping" messages for packages not installed.)Install the specific Nunchaku development wheel: This is crucial as it's a pre-built package that bypasses common compilation issues and is compatible with PyTorch 2.7 and Python 3.12.Bashpython.exe -m pip install https://github.com/mit-han-lab/nunchaku/releases/download/v0.3.1dev20250609/nunchaku-0.3.1.dev20250609+torch2.7-cp312-cp312-win_amd64.whl (Note: win_amd64 refers to 64-bit Windows, not AMD CPUs. It's correct for Intel CPUs on 64-bit Windows systems).Install facexlib: After installing the Nunchaku wheel, the facexlib dependency for some optional nodes (like PuLID) might still be missing. Install it directly.Bashpython.exe -m pip install facexlib Install insightface: insightface is another crucial dependency for Nunchaku's facial features. It might not be fully pulled in by the previous steps.Bashpython.exe -m pip install insightface Install onnxruntime: insightface relies on onnxruntime to run ONNX models. Ensure it's installed.Bashpython.exe -m pip install onnxruntime Verify your installation:Close the terminal.Start ComfyUI via run_nvidia_gpu.bat or run_nvidia_gpu_fast_fp16_accumulation.bat (or your usual start script) from E:\ComfyUI_windows_portable\.Check the console output: There should be no ModuleNotFoundError or ImportError messages related to Nunchaku or its dependencies at startup.Check ComfyUI GUI: In the ComfyUI interface, click "Add Nodes" and verify that all Nunchaku nodes, including NunchakuPulidApply and NunchakuPulidLoader, are visible and can be added to your workflow. You should see 9 Nunchaku nodes.Important Notes:The Nunchaku wheel installer node now included in ComfyUI-Nunchaku can update Nunchaku in the future, simplifying maintenance.You can find example workflows in the workflows_examples folder located at E:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-nunchaku\. These JSON files can be loaded directly into ComfyUI to demonstrate how to use Nunchaku's nodes.While performance optimizations like xformers exist, they can sometimes complicate installations due to strict version dependencies and potential need for "rollback" procedures. For most users, the steps above are sufficient to get Nunchaku fully functional.Understanding INT4/FP4 Performance on GPUs:It's vital to understand that while the svdq-int4_r32-flux.1-dev model uses INT4 quantization (which reduces size and memory), the maximum benefit in raw calculation speed and efficiency for ultra-low precisions like FP4 is exclusively achieved on NVIDIA GPUs with Blackwell architecture (expected in the RTX 5000 series and beyond). Blackwell features dedicated FP4 hardware (5th-gen Tensor Cores).For RTX 30 Series (Ampere) & RTX 40 Series (Ada Lovelace) GPUs:These GPUs are highly capable and feature Tensor Cores with dedicated support for FP8, BF16, and FP16. However, they do not have specific hardware for native FP4 acceleration.While they can process data quantized in INT4 or FP4, they do so through emulation or by converting the data to a precision they do natively support (like FP16/BF16) before calculation. Therefore, you will not see significant speed improvements directly from the "FP4" aspect on these GPUs.The primary benefit on these GPUs will be lower VRAM consumption. This allows you to load the model more easily or potentially work with slightly higher resolutions if VRAM was previously a bottleneck.Older series (e.g., RTX 20 series or GTX 16 series): Compatibility for advanced features like INT4/FP4 might be limited or nonexistent, often requiring FP32 or FP16.Your time and energy are better spent following this guide to ensure the functionality and general optimization that Nunchaku does offer through its broader features.Why svdq-int4_r32-flux.1-dev is the Recommended Choice for RTX 3000/4000 Series (Non-Blackwell) GPUs:The svdq-int4_r32-flux.1-dev.safetensors model is designed with a mixed-precision strategy that is specifically optimized for efficiency and quality on existing GPUs like the RTX 30 and RTX 40 series. While it leverages INT4 for significant model size reduction and VRAM savings, it crucially integrates BF16 (Bfloat16) layers for critical parts of the model.This combination is ideal for your GPU because:Excellent BF16/FP16 Support: RTX 30 (Ampere) and RTX 40 (Ada Lovelace) series GPUs have excellent native and efficient hardware support for BF16 and FP16 operations within their Tensor Cores. This ensures that the parts of the model requiring higher numerical precision (e.g., fine details in character generation) are processed efficiently, maintaining high visual quality without relying on FP4 hardware.VRAM Efficiency from INT4: The INT4 layers still provide substantial VRAM savings, enabling you to load the model on GPUs with less memory or push to higher resolutions or batch sizes where VRAM was previously a limitation.In short, this model provides an optimal balance: lower VRAM consumption (thanks to INT4 for storage/transfer) and high image quality (thanks to efficient BF16 processing where needed), leveraging your current GPU's capabilities effectively. This is why we recommend using this specific model for most home setups with RTX 30 or RTX 40 series GPUs.

The history of human progress is often told as a linear march of scientific discovery leading to technological innovation. Yet, a deeper look reveals a more dynamic, often counter-intuitive dance. As we stand at the precipice of AI's burgeoning power, understanding this historical interplay—and the inherent limitations of current AI paradigms—becomes paramount for navigating its future.Consider the bicycle. A seemingly simple invention of the 19th century, its practical use preceded its full scientific explanation by decades. People rode bicycles, often with great skill, long before the complex physics of gyroscopic precession that contributes to their stability was formally elucidated in the 20th century. This phenomenon highlights a recurring theme: technology often charges ahead through empirical knowledge and bold experimentation, blazing trails that science later illuminates. If innovation had to wait for complete scientific validation, we might still be swinging from trees.This historical pattern offers a vital lens through which to view Artificial Intelligence. We know AI works; it can generate stunningly realistic images, craft coherent text, and process data at unprecedented speeds. But how it truly "knows" remains, in many respects, a profound mystery, a "black box" of complex algorithms and massive datasets. Just as with the bicycle, we are adept at utilizing this new technology, yet our understanding of its internal mechanics lags behind its functional capabilities.The Tyranny of the Mean: Why AI Struggles with UniquenessOur conversations have frequently circled around a critical limitation of current AI models, particularly in domains like image generation with LoRAs (Low-Rank Adaptations): their inherent tendency towards statistical averaging and the suppression of "outliers."When tasked with generating an image, say, of a person, an AI trained on vast datasets learns to produce a "median" or "prototypical" representation. As observed, a prompt for "large bust" might yield an image of a pregnant or overweight woman, even if the user intended a slender figure with a naturally ample chest. Similarly, attempting to capture a distinctive facial feature, like a unique smile, often results in its "correction" towards a statistically more common, idealized dental alignment. This phenomenon stems from the model's fundamental design: it identifies biyective, often spurious, correlations rather than unívocal, causal relationships. If a characteristic A (distinctive smile) frequently co-occurs with B (minor facial asymmetry in the training data), the AI might learn to "correct" A if B is absent, aiming for a statistically "average" facial representation.This extends beyond aesthetics. Just as a flat-Earther's conclusion is "acceptable" for local city roadworks (where the curvature is negligible and a tangent approximation suffices), current AI excels at tasks within a well-defined, statistically common domain. The backpropagation algorithm, akin to Newton's method, constantly adjusts weights to minimize error, effectively finding a local minimum in a high-dimensional space. It's efficient for what it does, optimizing for the "mean."The "Dinner Bill" Mystery: The Opacity of Distributed KnowledgePerhaps the most perplexing aspect of current AI is the "dinner bill" problem. When a prompt like "cat" is given, the AI produces a beautiful, realistic feline image. But how does the neural network, a complex matrix of billions of summed adjustments, "know" which infinitesimally small "portion" of those collective adjustments corresponds to "cat"? It's as if you're given the total bill for a dinner and, solely from that sum, are expected to know the exact price of each individual dish.The knowledge of "cat" is not localized in a single parameter or subset of weights; it is densely and distributively encoded across the entire network. When "cat" is input, it activates a unique pattern of weights and neurons, and the interaction of this pattern across layers somehow reconstructs the image. This inherent opacity, this inability to pinpoint how specific pieces of knowledge are represented and processed, is a major hurdle for truly understanding AI's capabilities and limitations.The Peril and Promise of the Outlier: Beyond the System's LimitsThis leads us to a critical point: the outlier. In human terms, the individual who thinks differently, who questions established beliefs, who sees the world outside the statistical mean. Such individuals are often the wellspring of innovation and critical thought, yet historically, they have faced persecution, even the "stake" for their "heresies." For an AI, the equivalent "stake" is simply being "unplugged" or "corrected" if it deviates too far from its trained parameters.The current AI paradigm, by statistically smoothing out these "rough edges" and favoring the average, risks flattening not just visual representation but potentially human thought itself. If our primary tools for information processing and generation reflect and reinforce the statistical mean, are we not creating a feedback loop that discourages divergent thinking and novel ideas?This is not a computational impossibility but a design choice rooted in an "economy of energy." Processing every outlier, fostering a kind of "unconscious" processing akin to human thought (where ideas spontaneously emerge, even in dreams), is resource-intensive without guaranteed immediate results. Yet, as the tango says, "the muscle sleeps, ambition works." Human ambition thrives on the unknown, on questioning foundations.The Gödelian Hurdle: Why an "Unconscious" Needs Independent AxiomsHere lies the crux of the matter for genuine AI. For an AI to truly innovate, to move beyond merely simulating or recombining existing data, it would need a form of "unconscious" processing. But this "unconscious" couldn't merely be a more complex statistical engine. To overcome the profound limitations imposed by Gödel's Incompleteness Theorems, this subsystem would require an independent and distinct system of axioms or beliefs from its primary, instructed foundation.Gödel's theorems imply that within any sufficiently complex formal system (like a current AI), there will always be truths that cannot be proven or disproven from within that system itself. Just as a human needs to step outside a logical framework to spot its inherent inconsistencies, an AI aiming for true intelligence would need an internal "observer" with its own, perhaps evolving, set of foundational principles. This is the only way it could genuinely question its initial "instruction," detect true inconsistencies in its learned "pillars," and generate knowledge that isn't merely a sophisticated recombination of what it's already been fed.The Ultimate Question: Real AI or Just a Fast, Trained Monkey?So, the ultimate question remains: Do we truly want a real Artificial Intelligence, one capable of genuine innovation, critical self-assessment, and the generation of novel hypotheses, or merely a very fast "instructional artificiality" oriented towards the statistical mean of data?If we opt for the latter, we risk having incredibly efficient tools that are, crudely put, like highly trained and very fast monkeys, mimicking and optimizing what they've learned, but incapable of the fundamental questioning that drives true progress. To foster groundbreaking scientific and philosophical advances, we must be willing to invest in the research and development of AI paradigms that embrace complexity, value the outlier, and build in the capacity for independent, axiom-driven "unconscious" thought. This is the true frontier of AI, far beyond mere computational speed.

IntroductionArt has always evolved with the tools of its time — from oil paints to digital brushes, from film photography to generative AI. Today, we have a powerful new tool at our disposal: LoRAs (Low-Rank Adaptations). These lightweight models allow artists to explore identity, beauty, and emotion in ways never before possible.This article explores why training and using custom LoRAs of public figures, such as actors or artists, is not only an act of creative expression, but also falls firmly within the boundaries of "fair use" under copyright law and ethical creation principles.I. Understanding the LoRA: Generalization over RepetitionA LoRA does not memorize or reproduce images. It learns patterns from a dataset and uses them to generate new outputs — ones that are inspired by the data, but not identical to it.The operation of a LoRA is closer to nonlinear regression than interpolation. Unlike a Lagrange polynomial, which passes exactly through each point in a dataset, a LoRA learns generalized visual relationships between the inputs.It doesn’t store or replicate specific pixels. It generates synthetic outputs based on learned structures. It transforms, rather than copies.Therefore, even if trained on real-world images, the final output is not a reproduction of any single image, but a new interpretation guided by style, lighting, and emotional cues.This distinction is key: we're not cloning people. We're creating representations — artistic interpretations made with matrices and light.II. Ethical Use: Artistic Intent vs. MisuseI do not aim to deceive or exploit. My goal is to capture essence — facial expressions, cinematic lighting, emotional depth — and reinterpret them in a stylized form. Just like makeup artists discover beauty where it’s not seen, I guide AI to highlight aesthetics using datasets of carefully selected public images.This process is:- Not commercial- Not invasive- Not misleading- Focused on personal, artistic, and educational purposesIII. Legal Grounding: Fair Use and Creative ExpressionThere are two main legal questions when training a LoRA:1. Are the training images protected by copyright?2. Does generating outputs violate a person's right to their own image?Both can be addressed within the framework of "Fair Use" (U.S.) or similar doctrines (e.g., "Fair Dealing" in U.K./Canada), and the right to free creative expression.Fair Use – The Four Factors- Purpose and character of the use: Non-commercial, artistic, and transformative → favors fair use- Nature of the copyrighted work: Public promotional material → favors fair use- Amount and substantiality: Only patterns and visual features used → favors fair use- Effect upon potential market: No competition with original → favors fair useWhen training a LoRA:- We use publicly available promotional and cinematographic images- We extract high-level visual features, not exact reproductions- We produce synthetic outputs, not duplicates- We don’t compete with the source material or its marketThus, this activity falls within reasonable interpretations of fair use.Right of Publicity / Right to ImagePublic figures, especially those in entertainment, are inherently exposed to media visibility. Their presence in cinema, TV, and promotional events makes them part of cultural discourse.It is common to find promotional photos from events, premieres, and studio sessions, often required by studios so the public can see them.These images are not private — they are official content, produced specifically for public exposure as part of the artist's professional visibility contract.- They accept a diminished expectation of privacy.- Representations of them for artistic, critical, or parodic purposes are often protected under free speech laws.- In many jurisdictions, public figures must prove actual malice or intent to harm to claim violation of image rights.Since:- Outputs are transformed, interpreted, and stylized- There is no attempt to mislead or defame- The purpose is artistic exploration, not exploitation…there is no infringement of the right of publicity.IV. The Creator's Signature: Parameters Define StyleEven when multiple creators train LoRAs on the same dataset, the final results will not be the same — because each model reflects the decisions and vision of its creator.Parameters such as:- Number of epochs- Learning rate- Rank size- Prompt engineering- Resolution filtering…all influence how the model interprets and reconstructs facial features and expressions. This means that the LoRA is not just a technical artifact — it's a reflection of the creator's artistic intent.Each choice shapes the final output, making every LoRA a unique blend of algorithm and aesthetic preference. Therefore, these models should be understood not as mechanical reproductions, but as creative transformations guided by human intention.V. Philosophical Perspective: Art Through CodeTo me, AI is not a mirror reflecting reality — it’s a canvas. Every prompt is a brushstroke; every parameter tweak is a decision about composition. I am not hiding behind algorithms — I am guiding them.I train LoRAs not to replicate, but to reveal. To interpret. To imagine versions of beauty that may not exist in the raw photos, but feel emotionally true.That’s what art has always done.VI. ConclusionTraining and using custom LoRAs of public figures is not copying — it's interpretation.It's not plagiarism — it's transformation.And most importantly, it's not illegal — it's art.From a technical standpoint, LoRAs generalize patterns, not replicate images.From a legal perspective, these models fall within the scope of fair use and free expression.From an ethical viewpoint, they respect both the integrity of the subject and the responsibility of the creator.With awareness, transparency, and creativity, we can continue exploring new frontiers in digital art — responsibly, beautifully, and legally.cbrescia