Print-friendly | Download PDF | Save for Exam Day
This comprehensive cheat sheet covers all 7 NCA-GENM exam domains with formulas, architecture details, comparison tables, and key facts. Based on the official NVIDIA exam blueprint (25/20/15/15/10/10/5 weighting).
Domain 1: Experimentation (25%)
Evaluation Metrics Quick Reference
| Metric | Measures | Scale | Formula Intuition | When to Use |
|---|---|---|---|---|
| FID | Quality + diversity | Lower = better | Distance between real and generated image distributions | Comparing two generators overall |
| IS | Quality + diversity | Higher = better | Classifier confidence on generated images | Quick quality check |
| CLIP Score | Text-image alignment | Higher = better | Cosine similarity of CLIP embeddings | Does image match prompt? |
| BLEU | N-gram overlap | 0-1, higher = better | Precision of n-gram matches with reference | Caption evaluation |
| CIDEr | Caption consensus | Higher = better | TF-IDF weighted n-gram similarity | Best for captioning |
| METEOR | Word alignment | 0-1, higher = better | Precision + recall with synonyms | Captioning with paraphrases |
FID (Frechet Inception Distance)
Diffusion Model Hyperparameters
| Parameter | What It Controls | Default/Typical | Low Value Effect | High Value Effect |
|---|---|---|---|---|
| Guidance Scale | Prompt adherence vs diversity | 7.5 | Creative, diverse, less accurate | Strong adherence, may over-saturate |
| Inference Steps | Quality vs speed | 30-50 | Faster but lower quality | Better quality, diminishing returns |
| Scheduler | Denoising trajectory | DDPM | Baseline, slow | DDIM/DPM-Solver: fewer steps needed |
| Seed | Reproducibility | Random | N/A | Same seed = same output |
| Negative Prompt | What NOT to generate | Empty | No constraints | Removes unwanted features |
Guidance Scale Cheat Sheet
Scale 1.0 → No guidance (unconditional generation)
Scale 3-5 → Creative, diverse, less prompt-accurate
Scale 7-8 → Balanced (default for most use cases)
Scale 10-12 → Strong prompt adherence
Scale 15+ → Over-saturated, artifacts likely — avoid
Scheduler Comparison
| Scheduler | Steps Needed | Speed | Quality | Notes |
|---|---|---|---|---|
| DDPM | 1000 | Slowest | High | Original, rarely used for inference |
| DDIM | 20-50 | Fast | Good | Deterministic sampling option |
| Euler | 20-30 | Fast | Good | Simple, effective |
| DPM-Solver++ | 15-25 | Fastest | Good | State-of-the-art efficiency |
| UniPC | 10-20 | Very fast | Good | Unified predictor-corrector |
Fine-Tuning Methods
| Method | Trainable Params | Data Needed | Best For | Compute |
|---|---|---|---|---|
| LoRA | <1% | 50-500 images | Style transfer | Low |
| DreamBooth | ~All (+ prior) | 3-10 images | Subject personalization | High |
| Textual Inversion | 1 embedding | 5-15 images | Lightweight concept learning | Very low |
| Full Fine-Tuning | 100% | 1000+ images | Domain adaptation | Very high |
Prompt Engineering for Image Generation
Effective prompt structure:
[Subject] [Details] [Style] [Quality modifiers]
Example: "A golden retriever puppy, sitting in a field of wildflowers,
oil painting style, soft lighting, highly detailed, 4k"
Negative prompt examples:
"blurry, low quality, distorted, deformed, watermark, text,
oversaturated, bad anatomy, extra limbs"
Prompt weighting (Stable Diffusion syntax):
(important concept:1.3) → 30% more emphasis
(less important:0.7) → 30% less emphasis
Preparing for NCA-GENM? Practice with 455+ exam questions
Domain 2: Core ML and AI Knowledge (20%)
Vision Transformer (ViT) Architecture
Processing Pipeline:
Image (224×224) → Split into patches (14×14 grid of 16×16 patches)
→ Linear projection (each patch → embedding vector)
→ Add position embeddings (learnable, not sinusoidal)
→ Prepend CLS token
→ Transformer encoder (self-attention across all patches)
→ CLS token output → Classification head
Key Numbers:
- ViT-Base: 86M params, 12 layers, 768 hidden dim, 12 heads
- ViT-Large: 307M params, 24 layers, 1024 hidden dim, 16 heads
- Standard patch size: 16×16 pixels
- Standard input: 224×224 → 196 patches + 1 CLS token = 197 tokens
CLIP Architecture
Dual Encoder Structure:
Text → Text Encoder (Transformer) → Text Embedding (512-d)
→ Cosine Similarity
Image → Image Encoder (ViT/ResNet) → Image Embedding (512-d)
Contrastive Training:
- Batch of N text-image pairs
- N correct matches (diagonal) should have HIGH similarity
- N² - N incorrect matches (off-diagonal) should have LOW similarity
- Loss: symmetric cross-entropy on the similarity matrix
Zero-Shot Classification:
1. Create text prompts: "a photo of a [class]" for each class
2. Encode all text prompts → text embeddings
3. Encode the image → image embedding
4. Compute cosine similarity between image and each text embedding
5. Predicted class = highest similarity
CLIP Key Facts:
- Does NOT generate images — only aligns text and image representations
- Trained on 400M text-image pairs from the internet
- Enables zero-shot transfer without any fine-tuning
- CLIP Score uses this model to evaluate text-image alignment
Diffusion Models
Forward Process (Adding Noise):
x_0 (clean image) → x_1 → x_2 → ... → x_T (pure Gaussian noise)
Each step: x_t = √(α_t) × x_{t-1} + √(1 - α_t) × ε
where ε ~ N(0, I)
Reverse Process (Learned Denoising):
x_T (noise) → x_{T-1} → ... → x_1 → x_0 (generated image)
Neural network learns: ε_θ(x_t, t) ≈ ε (predict the noise)
Latent Diffusion Model (Stable Diffusion) Pipeline
Text Prompt → CLIP Text Encoder → Text Embeddings
↓ (cross-attention)
Random Noise → [U-Net Denoiser × N steps] → Denoised Latent
↓
VAE Decoder → Generated Image
Key Components:
- VAE Encoder: Compresses 512×512 image → 64×64 latent (8× spatial reduction)
- U-Net: Predicts noise in latent space, conditioned on text via cross-attention
- CLIP Text Encoder: Converts text prompt to embedding vectors
- VAE Decoder: Converts denoised latent back to pixel space
- Scheduler: Controls the denoising trajectory (step size, noise removal rate)
Cross-Attention vs Self-Attention
| Feature | Self-Attention | Cross-Attention |
|---|---|---|
| Q, K, V source | All from same input | Q from one modality, K/V from another |
| Purpose | Model relationships within a sequence | Fuse information across modalities |
| In Stable Diffusion | Image latent attends to itself | Image latent (Q) attends to text (K, V) |
| Result | Spatial coherence in image | Text conditioning of image generation |
Variational Autoencoder (VAE)
Architecture:
Input Image → Encoder → μ, σ (mean, std of latent distribution)
→ Reparameterization: z = μ + σ × ε (ε ~ N(0,1))
→ Decoder → Reconstructed Image
Model Architecture Comparison
| Architecture | Type | Input Modalities | Key Use Case | Examples |
|---|---|---|---|---|
| ViT | Encoder | Image | Classification, features | ViT-B/16, DINOv2 |
| CLIP | Dual Encoder | Text + Image | Alignment, retrieval | OpenCLIP, SigLIP |
| Stable Diffusion | Diffusion + VAE | Text → Image | Image generation | SD 2.1, SDXL |
| DALL-E | Diffusion/Autoregressive | Text → Image | Image generation | DALL-E 3 |
| LLaVA | Vision-Language | Image + Text → Text | Visual QA, captioning | LLaVA-1.5, LLaVA-NeXT |
| Whisper | Encoder-Decoder | Audio → Text | Speech recognition | Whisper Large-v3 |
Domain 3: Multimodal Data (15%)
Image Preprocessing Pipeline
| Step | Operation | Typical Values | Purpose |
|---|---|---|---|
| 1. Resize | Resize to model size | 224×224 (ViT), 512×512 (SD) | Match expected input |
| 2. Center Crop | Crop center region | Square crop | Remove borders |
| 3. Normalize | Scale pixel values | ImageNet: mean=[0.485, 0.456, 0.406] | Standardize input |
| 4. To Tensor | Convert to tensor | Channel-first (C, H, W) | Framework compatibility |
Data Augmentation Safety Guide
| Augmentation | Safe for Multimodal? | Risk |
|---|---|---|
| Horizontal Flip | Usually safe | Breaks "left/right" descriptions |
| Random Crop | Safe if subject visible | May crop out described objects |
| Color Jitter (mild) | Usually safe | Breaks color-specific descriptions |
| Color Jitter (strong) | Unsafe | "Red car" becomes blue |
| Rotation (small) | Safe | Breaks orientation descriptions |
| Vertical Flip | Usually unsafe | Breaks gravity, spatial relations |
| Cutout/Erasing | Risky | May remove described objects |
Audio Data Representations
| Format | Description | Dimensions | Used By |
|---|---|---|---|
| Waveform | Raw amplitude over time | 1D signal | WaveNet, SoundStream |
| Spectrogram | Time × Frequency | 2D (like image) | General audio models |
| Mel Spectrogram | Perceptual frequency scale | 2D (like image) | Whisper, audio transformers |
| MFCC | Compact spectral features | Feature vectors | Traditional speech |
Dataset Quality Checklist
- Text-image alignment: Captions accurately describe image content
- Resolution: Images meet minimum model requirements
- Diversity: Balanced representation across categories
- Deduplication: No exact or near-duplicate pairs
- Filtering: NSFW, low-quality, and watermarked images removed
- Language quality: Captions are grammatically correct, descriptive
- Scale: Sufficient examples for the task (thousands for fine-tuning, millions for pre-training)
Domain 4: Software Development (15%)
NVIDIA Tools Reference
| Tool | What It Does | When to Use It |
|---|---|---|
| NeMo | Full framework for building and training AI models | Custom multimodal model development |
| Picasso | Cloud service for visual content generation | Enterprise image/video generation |
| NIM | Pre-optimized model deployment microservices | Quick production deployment |
| Triton | High-performance inference server | Multi-model serving at scale |
| TensorRT | Inference optimization engine | Maximum GPU inference speed |
Hugging Face Diffusers Key API
# Load pipeline
from diffusers import StableDiffusionPipeline, DDIMScheduler
pipe = StableDiffusionPipeline.from_pretrained(
"stabilityai/stable-diffusion-2-1",
torch_dtype=torch.float16
).to("cuda")
# Change scheduler
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
# Generate image
image = pipe(
prompt="a castle on a hill, fantasy art",
negative_prompt="blurry, low quality",
num_inference_steps=30,
guidance_scale=7.5,
generator=torch.Generator("cuda").manual_seed(42)
).images[0]
# Save
image.save("output.png")
CLIP Usage Pattern
from transformers import CLIPModel, CLIPProcessor
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
# Zero-shot classification
inputs = processor(
text=["a photo of a cat", "a photo of a dog"],
images=image,
return_tensors="pt"
)
outputs = model(**inputs)
similarity = outputs.logits_per_image # shape: [1, 2]
predicted_class = similarity.argmax() # 0 = cat, 1 = dog
Tool Selection Decision Tree
Need to TRAIN a custom model? → NVIDIA NeMo
Need enterprise image GENERATION? → NVIDIA Picasso
Need to DEPLOY a model quickly? → NVIDIA NIM
Need high-throughput MODEL SERVING? → Triton Inference Server
Need to OPTIMIZE inference speed? → TensorRT
Need to PROTOTYPE with diffusion? → Hugging Face Diffusers
Need to EVALUATE text-image match? → CLIP (via Transformers)
Master These Concepts with Practice
Our NCA-GENM practice bundle includes:
- 7 full practice exams (455+ questions)
- Detailed explanations for every answer
- Domain-by-domain performance tracking
30-day money-back guarantee
Domain 5: Data Analysis and Visualization (10%)
Visualization Techniques
| Technique | What It Shows | How to Create | Interpretation |
|---|---|---|---|
| Attention Maps | Where model looks in image | Extract attention weights, overlay on image | Bright regions = high attention |
| Grad-CAM | Class-relevant image regions | Gradient-weighted activation maps | Highlights decision-relevant areas |
| t-SNE | Embedding clusters in 2D | Apply t-SNE to CLIP embeddings | Similar items cluster together |
| UMAP | Embedding structure in 2D | Faster alternative to t-SNE | Preserves global structure better |
| Training Curves | Loss over time | Plot loss vs steps/epochs | Should decrease smoothly |
| FID over Training | Generation quality over time | Compute FID periodically | Should decrease, then plateau |
Training Curve Diagnosis
| Symptom | Diagnosis | Fix |
|---|---|---|
| Training loss drops, val loss rises | Overfitting | More data, regularization, early stopping |
| Both losses remain high | Underfitting | More capacity, more training, lower LR |
| Loss spikes or NaN | Divergence | Lower learning rate, gradient clipping |
| Loss plateaus very early | Learning rate too low | Increase LR, use warmup schedule |
| FID stops improving | Training saturated | Stop training, adjust architecture |
Embedding Space Quality Indicators
Good CLIP embedding space (t-SNE visualization):
- Clear clusters by semantic category
- Similar concepts nearby (e.g., "dog" and "puppy" close)
- Text and image embeddings for matching pairs overlap
- Smooth transitions between related concepts
Poor embedding space:
- Random scatter with no clear structure
- Unrelated concepts mixed together
- Large gaps between text and image embeddings of same content
Domain 6: Performance Optimization (10%)
Quantization Reference
| Precision | Bytes/Param | Memory (1B) | Speed vs FP32 | Quality Loss |
|---|---|---|---|---|
| FP32 | 4 | 4 GB | 1x (baseline) | None |
| FP16 | 2 | 2 GB | ~2x | Negligible |
| BF16 | 2 | 2 GB | ~2x | Negligible (better for training) |
| INT8 | 1 | 1 GB | ~3-4x | Minimal with calibration |
| INT4 | 0.5 | 0.5 GB | ~4-6x | Noticeable, needs calibration |
Diffusion Speed Optimization
| Technique | Speedup | Quality Impact | Difficulty |
|---|---|---|---|
| FP16 inference | ~2x | Negligible | Easy |
| Fewer steps (50→25) | ~2x | Mild | Easy |
| Fast scheduler (DPM++) | ~2-3x fewer steps | Minimal | Easy |
| TensorRT compilation | ~2-3x | None | Medium |
| Step distillation | ~4-8x | Requires distilled model | Hard |
| Consistency models | ~1-4 steps | Slight quality trade-off | Hard |
Serving Optimization
| Strategy | What It Does | When to Use |
|---|---|---|
| Dynamic Batching | Groups incoming requests | Multiple concurrent users |
| Model Caching | Keeps model in GPU memory | Repeated model use |
| Request Queuing | Manages request overflow | High traffic |
| Horizontal Scaling | Multiple GPU instances | Exceeding single-GPU capacity |
| Async Processing | Non-blocking generation | Long-running tasks (image gen) |
Domain 7: Trustworthy AI (5%)
Key Concepts
| Concept | Definition | Example |
|---|---|---|
| Visual Bias | Model generates stereotypical images | "CEO" generates only white male images |
| Content Safety | Filtering harmful generated content | NSFW classifier before output delivery |
| Watermarking | Invisible markers in generated images | Proves AI origin for provenance |
| Deepfakes | Realistic fake face/video generation | Used for misinformation, fraud |
| Data Provenance | Tracking training data sources | Important for IP compliance |
| Consent | Permission for data usage | Especially for face images |
Bias Detection Methods
- Prompt diversity testing: Generate images for the same role/concept across demographics
- Demographic analysis: Measure representation in generated outputs
- Comparative evaluation: Compare outputs for "doctor" vs "nurse" for gender distribution
- Red teaming: Deliberately test edge cases and sensitive prompts
- Automated classifiers: Use attribute classifiers to measure output distributions
Content Safety Pipeline
User Prompt → Prompt Safety Check → [If safe] → Generate Image
→ Output Safety Check → [If safe] → Deliver
→ [If unsafe] → Block + Log
→ [If unsafe] → Block + Notify User
Watermarking Methods
| Method | Visibility | Robustness | Use Case |
|---|---|---|---|
| Visible watermark | User sees it | Easy to remove | Previews, demos |
| Invisible watermark | Imperceptible | Survives crops, compression | Production, provenance |
| Metadata embedding | In file metadata | Easy to strip | Basic tracking |
| Spectral watermark | In frequency domain | High robustness | Research, verification |
Quick Reference: Numbers to Remember
| Fact | Value |
|---|---|
| ViT-Base patch size | 16×16 pixels |
| ViT-Base input size | 224×224 pixels |
| ViT-Base patches | 196 (14×14 grid) |
| CLIP training data | 400M text-image pairs |
| CLIP embedding dim | 512 |
| Stable Diffusion latent size | 64×64 (from 512×512 input) |
| Typical guidance scale | 7.5 |
| Typical inference steps | 20-50 |
| FP16 memory savings | 2× over FP32 |
| INT8 memory savings | 4× over FP32 |
| NCA-GENM exam time | 60 minutes |
| NCA-GENM questions | 50-60 |
| NCA-GENM cost | $125 |
| NCA-GENM validity | 2 years |
Exam Day Quick Tips
- Largest domain is Experimentation (25%) — know metrics and hyperparameters cold
- CLIP appears across multiple domains — architecture, evaluation, visualization
- Diffusion models are the core generative technology — understand forward/reverse process
- ViT treats images as sequences of patches — not pixels, not arbitrary regions
- Cross-attention fuses text and image — text provides K/V, image provides Q
- FID = quality + diversity (lower better) vs CLIP Score = alignment (higher better)
- Guidance scale 7-8 is the default — 1.0 means no guidance, 15+ causes artifacts
- LoRA for style, DreamBooth for subjects, Textual Inversion for lightweight concepts
- FP16 is almost always a free speedup — negligible quality loss
- Never leave a question blank — no penalty for guessing
For more detailed coverage:
- NCA-GENM Complete Guide
- NCA-GENM Exam Domains
- How to Pass NCA-GENM First Attempt
- 4-Week Study Plan
- Practice Tests
Ready to Pass the NCA-GENM Exam?
Join thousands who passed with Preporato practice tests
Instant access30-day guaranteeUpdated monthly