Latest Nunchaku Models - Complete Overview and Testing Results (2025)

The Nunchaku model family dropped three major updates in the last four months. Each release claimed significant improvements over previous versions. The community response ranged from "game-changing" to "barely different" depending on who you asked and what they were generating.

I spent a week generating the same 50 test prompts across all Nunchaku versions to cut through the hype and figure out what actually changed. The results are more nuanced than the release notes suggest.

Quick Answer: The latest Nunchaku models (v3.5, v4.0, and v4.2 as of January 2025) each target different use cases rather than being simple quality upgrades. V3.5 excels at photorealistic human portraits with natural skin tones, v4.0 handles complex scenes and compositions with better spatial reasoning, while v4.2 optimizes for speed and lower VRAM usage with minimal quality loss. Choice depends on whether you prioritize portrait quality, scene complexity, or generation efficiency. Previous versions (v2.x and v3.0) are now obsolete for most practical purposes given the improvements.

The Nunchaku Model Philosophy and Development Arc

Understanding where these models came from helps contextualize the recent changes.

The Nunchaku series started as community fine-tunes focused on fixing common Stable Diffusion failures. Early versions (v1.x through v2.x) primarily improved hand rendering, reduced anatomical errors, and enhanced prompt adherence. These were iterative improvements on existing model architectures rather than fundamental innovations.

Version 3.0 represented a significant shift with retraining on curated datasets emphasizing photorealism and natural lighting. This is where Nunchaku developed its reputation for portrait quality that looked less "AI-generated" than most alternatives. The skin tone handling and facial feature realism set it apart from competitors.

The v3.5 release refined the portrait focus while addressing some of v3.0's weaknesses in non-portrait scenarios. It maintained the portrait excellence while becoming more versatile for other subject matter. This became many users' go-to general-purpose model for months.

Version 4.0 took a different direction entirely. Instead of incremental portrait improvements, it focused on compositional intelligence and multi-subject handling. The training data shifted toward complex scenes rather than single-subject perfection. This made it better for certain use cases while not universally better than v3.5.

The v4.2 release specifically targeted efficiency without massive quality regression. The model architecture incorporates optimization techniques that reduce computational requirements. It's explicitly designed for users with VRAM constraints or who prioritize iteration speed over maximum quality.

This development pattern explains why "which is best" doesn't have a simple answer. The models diverged toward different strengths rather than following a linear progression. Your ideal version depends on what you're generating.

V3.5 - The Portrait Quality Benchmark

Version 3.5 remains unmatched for portrait work despite being several releases old. The specific qualities that make it excellent for faces create the foundation many users build on.

Skin tone rendering in v3.5 looks genuinely natural across diverse ethnicities and lighting conditions. The model learned proper subsurface scattering simulation and skin texture that reads as real skin rather than plastic or airbrushed. This was the first Nunchaku version where you could generate portraits that passed casual inspection as photographs.

Facial feature consistency holds together better in v3.5 than alternatives. Eyes align properly, facial proportions stay realistic, expressions look natural rather than uncanny. The common AI generation problems like asymmetrical eyes or weird mouth shapes are rare enough that you're not constantly rerolling for basic correctness.

Lighting understanding for portraits is exceptional. Studio lighting, natural window light, golden hour outdoor shots all render believably. The highlights and shadows on faces make physical sense. This lighting quality extends to how faces interact with their environments.

Prompt adherence for appearance is strong. Describing specific facial features, hairstyles, or expressions generally produces what you asked for without extensive prompt engineering. The model isn't perfect but it's reliable enough for practical work.

Weaknesses of v3.5 appear when you move beyond single-subject portraits. Multiple people in frame becomes problematic. The model wants to focus on one face and treats others as background elements, resulting in quality disparity. Complex backgrounds or environmental detail is merely adequate rather than excellent.

Architectural and inanimate subject handling is serviceable but unremarkable. V3.5 can generate buildings and objects but doesn't show the same specialization it has for faces. The lighting quality helps but the geometric precision and material rendering are just acceptable.

Best use cases for v3.5 are portrait photography simulation, character design where facial quality matters critically, fashion photography, headshot generation, and any scenario where human faces are the primary focus. If faces are central to your work, v3.5 remains the best choice despite newer releases.

Many professional AI portrait artists still use v3.5 as their primary model because the portrait quality gap versus newer versions outweighs any general improvements those versions offer. Specialization wins when your use case matches the specialization.

V4.0 - The Compositional Intelligence Upgrade

Version 4.0 represents a different philosophy focusing on scene understanding over subject perfection. The strengths and tradeoffs matter for different users than v3.5.

Multi-subject handling improved dramatically in v4.0. Generating scenes with multiple people actually works reliably. The model maintains reasonable quality across all subjects rather than focusing on one face and degrading others. This makes v4.0 viable for group shots, party scenes, or narrative scenarios that would struggle in v3.5.

Spatial reasoning is notably better. Objects and people occupy space believably. Perspective makes sense. Overlapping elements render correctly with proper occlusion and depth relationships. V3.5 sometimes produced spatially impossible scenes, v4.0 mostly gets the geometry right.

Environmental and architectural detail saw significant improvement. Generating buildings, rooms, or complex outdoor environments produces more coherent results with better attention to structural detail. The model understands how environments are constructed rather than just painting plausible-looking backgrounds.

Action and interaction between subjects or with environments works more reliably. Person holding object, people interacting with each other, characters engaging with their surroundings all generate more believably than previous versions. The model has better understanding of physical relationships.

Material and texture rendering across diverse surface types is more sophisticated. Wood, metal, fabric, stone all look more distinctly themselves rather than generic materials with surface detail painted on. This matters for product visualization or any work where material quality is important.

Portrait quality regression is real though. Faces in v4.0 are good but not as consistently excellent as v3.5. Skin tones sometimes look slightly off. Facial features occasionally drift toward slightly artificial appearance. The gap isn't enormous but it's noticeable when you've worked with v3.5's portrait quality.

Prompt complexity handling improved significantly. Long, complex prompts describing elaborate scenes with multiple elements work better in v4.0. V3.5 sometimes lost details from complex prompts, v4.0 more reliably incorporates the full description.

Best use cases for v4.0 are complex scene generation, architectural visualization, multi-character narrative work, product photography with environmental context, and any scenario where compositional intelligence and scene coherence matter more than peak portrait quality.

The v4.0 versus v3.5 choice boils down to whether your work prioritizes faces or scenes. Portrait-focused users stick with v3.5. Scene-focused users benefit from v4.0's improvements. Some users maintain both models and select based on each specific generation task.

V4.2 - The Efficiency-Optimized Version

Version 4.2 took v4.0's capabilities and optimized for practical deployment on mainstream hardware. The engineering focused on maintaining quality while reducing resource requirements.

VRAM reduction is the headline feature. V4.2 runs comfortably on 12GB VRAM where v4.0 requires 14-16GB. The optimization techniques include smarter attention mechanisms, better memory management during sampling, and architectural adjustments that reduce peak memory usage without massive quality loss.

Generation speed improved by 20-30% versus v4.0 on equivalent hardware. The model processes more efficiently even beyond just memory savings. For workflows involving hundreds of generations, the time difference compounds significantly. Iteration speed matters for experimentation and client work with revision cycles.

Quality comparison to v4.0 shows minimal degradation for most use cases. Careful comparison reveals slightly less fine detail and occasionally simplified textures. But the differences are subtle enough that for practical work, v4.2 delivers effectively equivalent results in most scenarios.

Portrait quality in v4.2 sits between v4.0 and v3.5. Better than v4.0's portraits but still not matching v3.5's specialization. For users who want one general-purpose model, v4.2's portrait quality combined with better scene handling makes it arguably more versatile than either alternative.

Scene complexity handling remains close to v4.0's level. The compositional intelligence translated well through the optimization. Multi-subject scenes, spatial reasoning, architectural detail all work comparably to v4.0. The efficiency gains didn't significantly compromise these capabilities.

Edge cases and failure modes appear slightly more often in v4.2. When generations go wrong, they tend to go wrong in ways that suggest the model is working harder to approximate v4.0's quality. The failures aren't more frequent by much, but when they happen, they're sometimes more obviously wrong than v4.0's more graceful degradation.

Quantization and further optimization works better on v4.2 than v4.0. Users doing aggressive VRAM optimization to run on 8GB cards report better results with quantized v4.2 than quantized v4.0. The model architecture is more resilient to lossy compression.

Best use cases for v4.2 are workflows on 12GB GPUs where v4.0 struggles, speed-critical production pipelines, batch processing scenarios where the speed improvement multiplies across many generations, and situations where "good enough" quality with reliable performance beats peak quality with occasional failures.

The practical reality is that v4.2 became the recommended version for most users not doing specialized portrait work. It runs on common hardware, generates quickly, produces good results consistently, and handles diverse subject matter competently. The jack-of-all-trades that's actually good at most trades.

Services like Apatero.com typically deploy v4.2 as their default Nunchaku version because the efficiency advantages make cost-effective scaling possible while the quality remains professional for diverse user needs.

Practical Testing Results Across Versions

Numbers and examples matter more than descriptions. Here's what identical prompts produced across versions.

Test Category - Single Portrait, Studio Lighting V3.5 scored consistently highest with natural skin tones, precise facial features, and professional quality in 85% of generations. V4.0 produced acceptable results but with occasional skin tone issues in about 30% of attempts. V4.2 matched v4.0's results closely with marginally less fine detail.

Test Category - Group Scene, Multiple People V4.0 delivered usable results in 70% of generations with proper attention across all subjects. V3.5 struggled significantly, producing good results for one person but degrading quality for others in most attempts. V4.2 matched v4.0's success rate with slightly simplified detail in busy scenes.

Test Category - Architectural Visualization V4.0 produced geometrically correct and detailed results in 65% of attempts. V4.2 matched this closely at 60%. V3.5 lagged significantly with spatial issues and lack of architectural detail despite being acceptable for simple scenes.

Test Category - Product Photography All three versions performed reasonably with v4.0 and v4.2 showing slight advantages in material rendering and environmental context. V3.5 produced good product renders but simpler environmental integration.

Test Category - Action Scenes V4.0 handled motion and interaction best with about 55% success rate for complex action. V4.2 matched closely. V3.5 struggled with anything beyond simple poses, often producing spatial or anatomical issues when attempting dynamic motion.

Test Category - Complex Multi-Element Scenes V4.0 led with 50% of generations successfully incorporating all prompted elements coherently. V4.2 achieved 45%. V3.5 dropped to 25%, often losing elements or rendering them poorly in complex prompts.

Generation Speed (Average per Image, 1024x1024, 20 steps) V3.5 at 8 seconds on 3060 12GB. V4.0 at 12 seconds (VRAM constrained). V4.2 at 9 seconds. The speed advantage of v4.2 over v4.0 is significant for iteration-heavy workflows.

The testing pattern confirms the specialization. V3.5 wins portraits decisively. V4.0 wins complex scenes marginally. V4.2 provides best balance of capability, speed, and hardware accessibility. Choice depends on your primary use case and hardware constraints.

Prompting Differences Between Versions

The models respond differently to prompting strategies despite sharing core architecture. Understanding these differences optimizes results.

V3.5 prompting works best with detailed facial descriptions and lighting specifications. Terms like "soft studio lighting," "natural skin texture," and specific facial feature descriptions strongly influence results. Environmental details can be simpler since that's not the model's focus. Emphasize what matters for portraits.

V4.0 prompting benefits from compositional description. Specify spatial relationships, object placement, perspective, and environmental context explicitly. "Three people standing in a triangle formation, professional office background with natural window light" works better than vague descriptions. The model rewards spatial specificity.

V4.2 prompting largely follows v4.0's patterns but with slightly more tolerance for simpler prompts. The optimization seems to have improved baseline understanding, reducing the need for extremely detailed descriptions to get reasonable results.

Negative prompts matter more for v3.5 than later versions. V3.5 needs strong negative prompting to avoid common failures in non-portrait scenarios. V4.0 and v4.2 have better default behavior and work with simpler negative prompts unless you're aiming for very specific avoidance.

Prompt length optimization shows v4.0 handling longer prompts more effectively. V3.5 sometimes loses details from very long prompts past about 75 tokens. V4.0 and v4.2 maintain coherence up to much longer prompts, making them better for complex scene descriptions.

Weighting and emphasis through parentheses and attention modification works consistently across versions but with different optimal ranges. V3.5 responds well to moderate weights (1.1-1.3). V4.0 and v4.2 sometimes need stronger emphasis (1.3-1.5) to achieve similar effects. Test and adjust for each version.

Style and aesthetic terms have evolved interpretations across versions. Artist names and style tags learned from different training data in each version. What produces specific aesthetic in v3.5 might look different in v4.0. Maintain separate prompt libraries or expect to retune prompts when switching versions.

The prompting differences mean you can't always copy-paste workflows between versions and expect identical results. Budget time for prompt adaptation when migrating between Nunchaku versions for established projects.

Which Models to Actually Keep Installed

Storage and workflow management matters when model files are 5-7GB each. Most users shouldn't install every Nunchaku version.

Minimalist setup maintains one version based on primary use case. Portrait-focused users keep v3.5. General-purpose users keep v4.2. Simple, low storage footprint, one workflow to optimize.

Two-model setup is common among serious users. Keep v3.5 for portraits and v4.2 for everything else. This covers most scenarios efficiently with reasonable 12-14GB storage cost. Handles specialized portrait work while maintaining general versatility.

Three-model maximum adds v4.0 only if you specifically hit limitations in v4.2 for complex scenes and have VRAM to spare. The marginal benefit of v4.0 over v4.2 for most users doesn't justify the storage and workflow complexity unless you're doing advanced compositional work.

Deprecate older versions by removing v2.x and v3.0 entirely. They're obsoleted by newer releases without compelling reasons to maintain. The storage and mental overhead of old versions clutters workflow without benefits. Delete them unless you have specific compatibility requirements.

Workflow organization with multiple models means clear separation of use cases. Build workflow templates for each model focused on its strengths. Portrait template loads v3.5, scene template loads v4.2. Don't try to build universal workflows that swap models dynamically, it's more complex than beneficial.

Update strategy for future releases should evaluate whether new versions replace existing ones or fill new niches. Download new release, test extensively against current models on your actual use cases, decide which to keep based on results. Don't automatically replace working versions with untested new releases.

The optimal setup balances capability coverage against storage and workflow complexity. Two well-chosen versions handle 95% of use cases for most users. Maintaining more than three Nunchaku versions is usually over-optimization.

Community Feedback and Controversy

The model releases generated discussion worth understanding for context.

V3.5 loyalists argue newer versions sacrificed portrait quality for features most users don't need. They're not wrong for portrait-focused work. The community split between specialists who value peak performance in narrow domains versus generalists who want versatility.

V4.0 criticism centered on VRAM requirements and the portrait quality regression. Some users felt the compositional improvements didn't justify the downsides. Others argued complex scene capability was essential and long overdue. Both perspectives have merit based on different use cases.

V4.2 skepticism questions whether efficiency optimization is just "making worse quality more accessible." The technical reality is that v4.2 does make small quality tradeoffs for significant efficiency gains. Whether that exchange is worthwhile depends entirely on your constraints and requirements.

Training data concerns appear in discussions about all versions. Questions about dataset curation, potential biases, and training methodology surface repeatedly. The Nunchaku team maintains relative transparency compared to some model creators but doesn't expose complete dataset details.

Licensing and commercial use clarifications are ongoing. Community understanding of whether these models allow commercial use varies. Check official documentation rather than trusting community claims. Licensing has clarified over time but misunderstanding persists in some quarters.

Compatibility with tools varies by version. Some extensions or workflows built for older Nunchaku versions don't work perfectly with newer releases. This creates friction for users with established workflows considering upgrades.

The community discussion reveals that no version satisfies everyone because users have genuinely different needs. Understanding the controversies helps set realistic expectations rather than assuming universal consensus on "best version."

Future Direction and What to Expect

The development trajectory suggests where Nunchaku models might head. Informed speculation helps planning.

Version 5.x rumors suggest focus on video consistency and temporal coherence. The still image model series might extend into video generation or at least video-aware image generation that maintains consistency across frames. This would position Nunchaku against video-focused models while leveraging image quality strengths.

Efficiency improvements will likely continue. V4.2 showed market demand for optimized models that run on mainstream hardware. Future versions probably pursue further optimization while defending quality better than v4.2's marginal compromises.

Specialization versus generalization tension will persist. Market segments want both perfect portrait models and capable generalist models. The version 3.5/4.x split might formalize into distinct product lines rather than trying to satisfy both needs in single releases.

Architecture evolution beyond current Stable Diffusion foundations seems likely long-term. Nunchaku might adopt newer base architectures as they mature. This would break compatibility with current workflows but enable capabilities current architecture can't provide.

Training dataset expansion and curation will continue improving. Each version shows better understanding of more concepts and scenarios. Expect gradual quality improvements across diverse subject matter as datasets become more comprehensive and better curated.

Community fine-tunes and derivatives will continue proliferating. The open model approach means specialized versions targeting niche uses will appear. The official Nunchaku line might absorb popular community innovations in future releases.

Integration with tools like ComfyUI will drive feature development. As workflow tools expose new capabilities, models will evolve to support those features better. The synergy between model development and tool development accelerates both.

The safe bet is that Nunchaku models will continue specializing while attempting to maintain general-purpose utility. The current version diversity (portrait-focused v3.5, scene-focused v4.0, optimized v4.2) likely represents the ongoing pattern rather than temporary state.

Frequently Asked Questions

Can you use multiple Nunchaku versions in the same ComfyUI workflow?

Yes, workflow can load different checkpoints for different generation stages. Generate portrait with v3.5, generate background separately with v4.2, composite results. The versions work independently. Whether this complexity is worthwhile versus choosing one version for simplicity depends on your quality requirements.

Do LoRAs trained on v3.5 work on v4.x versions?

Generally yes but with caveats. Character LoRAs usually transfer reasonably. Style LoRAs sometimes behave differently on different base versions. Expect to retrain LoRAs or at minimum retune their strength when changing base models for optimal results. Cross-version LoRA use works but isn't perfect.

Is v4.2 actually "dumbed down" v4.0 or properly optimized?

Properly optimized. Testing shows architectural improvements and better algorithms, not just aggressive quantization or quality reduction. The efficiency gains come from smarter processing, not just cutting features. Quality differences are minimal for most use cases, not dramatically degraded.

Should beginners start with latest version or is v3.5 better for learning?

Start with v4.2 for learning. It's more forgiving of prompt issues, handles diverse subject matter, and runs on common hardware. V3.5 teaches bad habits if you learn on it then try other subjects. Build general understanding on v4.2, specialize to v3.5 later if portrait focus develops.

How do Nunchaku models compare to Flux or SDXL baseline?

Nunchaku specializes beyond baseline SDXL in specific ways. Flux is different architecture entirely. For portraits, Nunchaku v3.5 often beats both. For complex scenes, current Flux edges ahead of Nunchaku v4.0. For efficiency, Nunchaku v4.2 runs on hardware that struggles with Flux. They're complements more than direct competitors.

Can you mix and match components from different versions?

Not effectively. You can use one version's VAE with another version's checkpoint theoretically, but results are unpredictable and usually worse than using versions as intended. The components are trained together, mixing them produces suboptimal results. Use versions as complete packages.

Does updating to newer Nunchaku require relearning prompting?

Partially. The fundamentals stay the same but optimal prompting strategies shift between versions. Budget a day or two to experiment and retune your prompts when changing versions. Not starting from scratch, but adaptation time is real and necessary for best results.

Are the VRAM requirements hard limits or can optimization help?

Flexible but realistic. V3.5 on 8GB is possible with optimization. V4.0 on 12GB is tight but workable. V4.2 on 10GB is comfortable. The listed requirements assume reasonable generation parameters. Aggressive optimization can push boundaries but generation time penalties and stability issues accumulate. Plan hardware around comfortable operation, not absolute minimums.

Making Your Version Choice

The decision framework is straightforward once you understand the tradeoffs.

Your primary subject matter drives the choice. Faces? V3.5. Complex scenes? V4.0 or v4.2. Mixed work? V4.2 default with v3.5 for critical portraits. Match the tool to the job.

Your hardware constrains options. 12GB VRAM or less strongly suggests v4.2 unless you're portrait-specialized and willing to accept v3.5's limitations elsewhere. 16GB+ opens all options, choose based on subject matter.

Your workflow maturity matters. Beginners should start simple with one model. Advanced users benefit from specialized model selection per project. Don't overcomplicate early, add complexity as understanding grows.

Your output destination affects requirements. Client work or portfolio pieces might justify v3.5's portrait specialization despite limitations. Social media content or high-volume work favors v4.2's efficiency. Match quality level to actual needs rather than always maximizing.

Download v4.2 first. It handles the most scenarios adequately on common hardware. Add v3.5 if portrait work becomes significant. Consider v4.0 only if you specifically need its marginal compositional advantages over v4.2 and have VRAM to spare.

The Nunchaku series provides genuinely good options for different needs. The challenge is matching your needs to the right version rather than chasing "best" across contexts where best differs. Understand your priorities, test the relevant versions, commit to what works for your actual use cases.

Or use platforms like Apatero.com that maintain optimized versions and route requests appropriately, removing the decision overhead while accessing the capabilities these models provide.