পাঠ ৩৭ · ৪০-এর মধ্যে · মডিউল ৫
Home / AI Courses / ডিপ লার্নিং / Diffusion model

Diffusion model পরিচিতি

Diffusion models — denoise to generate
৮ মিনিট পড়া মাঝারি · Intermediate PyTorch hands-on

এই পাঠে যা শিখবেন

  • Forward diffusion process — Markov chain noise schedule
  • Reverse process — denoising network ($\epsilon$-prediction)
  • Loss function — simple MSE on noise
  • U-Net architecture — diffusion-এর backbone
  • Conditioning ও classifier-free guidance
  • Stable Diffusion, DALL-E, Sora — landscape

১ · মূল ধারণা — noise → image

একটি real image $x_0$ নিন। ধাপে ধাপে noise add করুন — $T$ step পরে $x_T \approx \mathcal{N}(0, I)$ (pure noise)। এখন প্রশ্ন — উল্টো করা যায় কি? Pure noise থেকে শুরু করে ধাপে ধাপে denoise করে real image-এ পৌঁছানো?

Diffusion-এর insight: প্রতিটি tiny denoising step relatively easy — ছোট noise estimate করো। ১০০০ step করে পুরো image তৈরি কঠিন তবু solvable। Iterative refinement-এর জাদু।

ভাবুন একজন ভাস্কর — মার্বেলের একটি বিশাল blob থেকে ধীরে ধীরে chip চিপ করে ভাস্কর্য বের করেন। এক ছেনি দিয়ে পুরো ভাস্কর্য কাটার চেষ্টা করেন না। Diffusion-ও তাই — প্রতিটি step-এ এক স্তর noise সরিয়ে শেষে clean image।

২ · Forward process — noise schedule

প্রতিটি step:

$$q(x_t | x_{t-1}) = \mathcal{N}(x_t; \sqrt{1 - \beta_t} \, x_{t-1}, \beta_t I)$$

$\beta_t$ — variance schedule (small, e.g., $0.0001 \to 0.02$)। Markov chain — current step শুধু previous-এর উপর depend।

Closed form to step $t$:

$$x_t = \sqrt{\bar{\alpha}_t} \, x_0 + \sqrt{1 - \bar{\alpha}_t} \, \epsilon, \quad \epsilon \sim \mathcal{N}(0, I)$$

যেখানে $\alpha_t = 1 - \beta_t$, $\bar{\alpha}_t = \prod_{s=1}^t \alpha_s$। সরাসরি $x_0$ থেকে $x_t$ — training-এ দ্রুত sample।

৩ · Reverse process — neural denoising

$p_\theta(x_{t-1} | x_t)$ — neural network parameterize। Ho et al. (২০২০) — সরল reformulation: model $\epsilon$-predict করুক:

$$\epsilon_\theta(x_t, t) \approx \epsilon$$

Loss simply MSE:

$$\mathcal{L}_{\text{simple}} = \mathbb{E}_{t, x_0, \epsilon} \left[ \| \epsilon - \epsilon_\theta(x_t, t) \|^2 \right]$$

Training algorithm:

  1. Random sample $t \in [1, T]$।
  2. Random sample $\epsilon \sim \mathcal{N}(0, I)$।
  3. Compute $x_t = \sqrt{\bar{\alpha}_t} x_0 + \sqrt{1-\bar{\alpha}_t} \epsilon$।
  4. Predict $\hat{\epsilon} = \epsilon_\theta(x_t, t)$।
  5. $\nabla \| \epsilon - \hat{\epsilon} \|^2$ — update।

৪ · U-Net — diffusion-এর backbone

Denoising network typically U-NetU-NetEncoder-decoder architecture with skip connections — segmentation-এ designed (Ronneberger ২০১৫), diffusion-এ adopted। Different scale-এ feature combine।:

  • Encoder downsample — multi-scale features।
  • Decoder upsample — image-size output।
  • Skip connections — fine detail preservation।
  • Time embedding $t$ — sinusoidal + MLP, every layer-এ inject।
  • Self-attention — middle layer-এ global pattern।
Diffusion — forward (noise add) ও reverse (denoise) x_0 → x_1 → ... → x_T → ... → x_1 → x_0 Forward (fixed) x_0 clean image x_t partial noise x_T ~ N(0, I) + noise + noise Reverse (learned) x_T noise x_t denoising x_0 generated! U-Net ε_θ predict subtract noise U-Net ε_θ(x_t, t) input: noisy image + step output: predicted noise Loss = ‖ε − ε_θ(x_t, t)‖² — সরল MSE
Forward — fixed Markov chain noise add। Reverse — U-Net learned, প্রতিটি step-এ noise predict করে denoise।

৫ · PyTorch — minimal training step

Python · DDPM training step 
import torch
import torch.nn as nn

T = 1000
betas = torch.linspace(1e-4, 0.02, T)
alphas = 1.0 - betas
alpha_bar = torch.cumprod(alphas, dim=0)

def q_sample(x0, t, noise):
    """Forward — sample x_t from x_0."""
    a = alpha_bar[t].view(-1, 1, 1, 1).sqrt()
    one_minus = (1 - alpha_bar[t]).view(-1, 1, 1, 1).sqrt()
    return a * x0 + one_minus * noise

# Pretend we have a UNet — placeholder
unet = nn.Conv2d(1, 1, 3, padding=1)   # toy stand-in

# One training step
x0 = torch.randn(8, 1, 28, 28)         # MNIST-shaped
t = torch.randint(0, T, (8,))           # random step per sample
noise = torch.randn_like(x0)
xt = q_sample(x0, t, noise)
pred = unet(xt)                         # real UNet would also see t embedding
loss = ((pred - noise) ** 2).mean()
print(f"loss: {loss.item():.4f}")

    


    
৬ · Sampling — pure noise থেকে image

    

      Trained model দিয়ে generation:
    

    
    
      
  1. $x_T \sim \mathcal{N}(0, I)$ (pure noise)।
    2. 
      
  2. $t = T$ থেকে $1$ পর্যন্ত: predict $\hat{\epsilon} = \epsilon_\theta(x_t, t)$, compute $x_{t-1}$ via formula।
    3. 
      
  3. $x_0$ — final image।
    4. 
    

    

      DDPM-এ ১০০০ step lots। DDIM (২০২০) — deterministic skip, ৫০ step-এই comparable quality। Modern samplers — DPM-Solver, Euler, Heun — ১০-২০ step-এই photorealistic।
    


    
৭ · Conditioning — text-to-image

    

      Vanilla DDPM unconditional — random image। Conditioning দিয়ে controllable:
    

    
    
      
  • Class label: "cat" — embedding inject।
    • 
      
  • Text: CLIP encoder → text embedding → cross-attention।
    • 
      
  • Image: ControlNet — pose, sketch, depth।
    • 
    

    

      Classifier-free guidance (Ho-Salimans ২০২২): conditional + unconditional joint train। Sampling-এ:
    

    

      $$\hat{\epsilon} = \epsilon_\theta(x_t, t, c) + s \cdot (\epsilon_\theta(x_t, t, c) - \epsilon_\theta(x_t, t, \emptyset))$$
    

    

      Guidance scale $s$ — ৭-১০ typical। Higher $s$ — prompt-faithful কিন্তু diversity কমে।
    


    
৮ · Stable Diffusion — practical breakthrough

    
    
      
  • Latent diffusion: VAE encode 5১২×5১২ → ৬৪×৬৪ latent — ৪৮x compute reduction।
    • 
      
  • CLIP text encoder: prompt → text embedding।
    • 
      
  • Cross-attention U-Net: text condition inject।
    • 
      
  • Open-source: ২০২২ — community fine-tune flood (LoRA, ControlNet, Dreambooth)।
    • 
      
  • Consumer GPU viable — ৬-৮GB VRAM-এ inference।
    • 
    

    

      DALL-E ২, Midjourney, Imagen — সব diffusion-based। Sora (video) — diffusion + Transformer।
    


    
৯ · GAN বনাম Diffusion

    
    
      
  • Quality: Diffusion সাধারণত winner — sharper, more diverse।
    • 
      
  • Training stability: Diffusion >> GAN। One loss, no adversarial।
    • 
      
  • Mode coverage: Diffusion natural। GAN-এ collapse risk।
    • 
      
  • Inference speed: GAN single-pass, fast। Diffusion iterative, slower।
    • 
      
  • Controllability: Diffusion — guidance, ControlNet — flexible।
    • 
      
  • Likelihood: Diffusion — proper probabilistic model। GAN — implicit।
    • 
    


    

      Diffusion ২০২২+ generative AI-এর dominant paradigm। Image, video, 3D, audio — সর্বত্র। Consistency models, flow matching — even faster variant। L38-এ GPU/CUDA — যা ছাড়া এত big model train অসম্ভব।
    


  


  

    
    
  


  

    
 ভাবনার প্রশ্ন

    
প্রতিটি প্রশ্ন নিজে কিছুক্ষণ ভাবুন — তারপর "→ উত্তর" চাপুন।


    

      
        প্র ০১
        Diffusion-এর iterative process আপাতদৃষ্টিতে inefficient — ১০০০ step কেন? Mathematical ও practical justification কী?
      
      

        

          Iterative refinement — diffusion-এর core insight, weakness নয়।
        

        
Mathematical reason:

        
    
          
  • প্রতিটি step Gaussian transition — analytically tractable।
    • 
          
  • Multiple small steps ≈ approximate complex distribution।
    • 
          
  • Single step approximation poor (impossible practically)।
    • 
          
  • Markov chain-এর gradient strong।
    • 
        

        
Score matching connection:

        
    
          
  • $\epsilon$-prediction = score function approximation।
    • 
          
  • Langevin dynamics — slow convergence।
    • 
          
  • Noise schedule — fast traversal trick।
    • 
          
  • SDE theory underpinning।
    • 
        

        
Why ১০০০ specifically:

        
    
          
  • Original DDPM — empirical choice।
    • 
          
  • Quality-step trade-off optimal।
    • 
          
  • Beta schedule fine-grained।
    • 
          
  • Modern reduced — ১০০-৫০-২০।
    • 
        

        
Acceleration techniques:

        
(১) DDIM (২০২০):

        
    
          
  • Deterministic step skip।
    • 
          
  • ৫০ step competitive।
    • 
          
  • Reformulate as ODE।
    • 
        

        
(২) DPM-Solver:

        
    
          
  • ODE high-order solver।
    • 
          
  • ১০-২০ step photorealistic।
    • 
          
  • SDXL, FLUX-এ default।
    • 
        

        
(৩) Consistency models (২০২৩):

        
    
          
  • Single-step generation।
    • 
          
  • Distillation from pretrained diffusion।
    • 
          
  • GAN-like speed, diffusion-like quality।
    • 
        

        
(৪) Flow matching:

        
    
          
  • SD3, FLUX architecture।
    • 
          
  • Continuous time formulation।
    • 
          
  • Faster inference।
    • 
        

        
Inherent trade-off:

        
    
          
  • Quality $\leftrightarrow$ speed।
    • 
          
  • Iterative refinement gives time to "think"।
    • 
          
  • Single-pass inherently limited।
    • 
          
  • Diffusion's compute = quality।
    • 
        

        
Comparison with GAN:

        
    
          
  • GAN single forward pass — fast inference।
    • 
          
  • Diffusion iterative — slower but better quality।
    • 
          
  • Modern distilled diffusion — close to GAN speed।
    • 
          
  • Quality gap closing fast।
    • 
        

        
Iterative refinement intuition:

        
    
          
  • Hard problem decomposed into easier steps।
    • 
          
  • Each step — local correction।
    • 
          
  • Cumulative — global solution।
    • 
          
  • Like writing — draft, revise, polish।
    • 
        

        

          মূল উপলব্ধি: Iterative — diffusion-এর strength। Mathematical foundation — small step Gaussian tractable। Modern acceleration — distillation, ODE solver। Quality-speed trade-off intrinsic। Iterative refinement universal AI principle (chain-of-thought, beam search analog)।
        

      

    


    

      
        প্র ০২
        Classifier-free guidance scale $s$ — কেন low value-এ image diverse কিন্তু prompt-faithful কম? Mechanism কী?
      
      

        

          CFG — practical diffusion-এর critical hyperparameter। Mechanism subtle।
        

        
Formula recap:

        

          $$\hat{\epsilon} = \epsilon_{uncond} + s \cdot (\epsilon_{cond} - \epsilon_{uncond})$$
        

        
    
          
  • $s = 0$: pure unconditional।
    • 
          
  • $s = 1$: pure conditional।
    • 
          
  • $s > 1$: extrapolation in conditional direction।
    • 
        

        
Why extrapolation works:

        
    
          
  • Conditional model — average match।
    • 
          
  • Extrapolate — push further toward prompt।
    • 
          
  • Gradient direction amplify।
    • 
          
  • Prompt-image distance reduce।
    • 
        

        
Trade-off mechanism:

        
Low $s$ (1-3):

        
    
          
  • Sample broader distribution।
    • 
          
  • Prompt loosely matched।
    • 
          
  • Diverse output।
    • 
          
  • Sometimes off-topic।
    • 
        

        
Medium $s$ (5-8):

        
    
          
  • Sweet spot quality।
    • 
          
  • Prompt-faithful।
    • 
          
  • Reasonable diversity।
    • 
          
  • Production default।
    • 
        

        
High $s$ (10-20):

        
    
          
  • Highly prompt-aligned।
    • 
          
  • Saturated colors।
    • 
          
  • Less diverse।
    • 
          
  • Sometimes artifacts।
    • 
        

        
Visual symptoms by scale:

        
    
          
  • $s = 1$: blurry, generic।
    • 
          
  • $s = 7.5$: clean, prompt-aligned।
    • 
          
  • $s = 15$: oversaturated, "cartoonish"।
    • 
          
  • $s = 30$: artifacts, distorted।
    • 
        

        
Why diversity loss at high $s$:

        
    
          
  • Direction amplified — same target।
    • 
          
  • Variance suppressed।
    • 
          
  • Mode-collapse-like behavior।
    • 
          
  • "Mean prompt image"।
    • 
        

        
Geometric intuition:

        
    
          
  • Probability density shift।
    • 
          
  • Conditional region narrow।
    • 
          
  • High $s$ — peaked near mean।
    • 
          
  • Variance reduction।
    • 
        

        
Application-specific:

        
    
          
  • Photorealistic: $s = 5-7$।
    • 
          
  • Stylized: $s = 8-12$।
    • 
          
  • Specific reference: $s = 12-15$।
    • 
          
  • Ad/marketing: experiment।
    • 
        

        
Recent improvements:

        
    
          
  • Dynamic CFG — schedule across timesteps।
    • 
          
  • Negative prompts — undesirable away।
    • 
          
  • Guidance distillation — single-pass equivalent।
    • 
          
  • SDXL turbo — guidance baked in।
    • 
        

        

          মূল উপলব্ধি: CFG = direction extrapolation, prompt alignment vs diversity trade-off। Sweet spot $s = 7$। High $s$ — quality drop, mode collapse-like। Modern diffusion — guidance fundamental। Hyperparameter tuning art।
        

      

    


    

      
        প্র ০৩
        Bangladesh-এর একটি digital agency Stable Diffusion দিয়ে Bangla wedding card design করতে চাচ্ছে। Custom fine-tune approach কী?
      
      

        

          Practical Bangladesh creative use — Bangla wedding invitation। Cultural specificity critical।
        

        
Use case:

        
    
          
  • Diverse cultural styles — Bengali traditional, modern fusion।
    • 
          
  • Bangla typography integration।
    • 
          
  • Religious motifs (Hindu, Muslim, Christian)।
    • 
          
  • Color palette specific।
    • 
          
  • Repeating client orders — speed essential।
    • 
        

        
Approach 1 — LoRA fine-tune:

        
Pros:

        
    
          
  • Cheap — single GPU, ১-২ ঘণ্টা।
    • 
          
  • Small storage (~১০০MB)।
    • 
          
  • Multiple style — multiple LoRA।
    • 
          
  • Combine via merge।
    • 
        

        
Cons:

        
    
          
  • Limited transformation।
    • 
          
  • Base model bias retain।
    • 
          
  • Bangla typography poor (text rendering hard)।
    • 
        

        
Approach 2 — Dreambooth:

        
    
          
  • Full fine-tune small dataset।
    • 
          
  • Specific style/object embedding।
    • 
          
  • Higher quality LoRA-এর চেয়ে।
    • 
          
  • Compute heavier।
    • 
        

        
Approach 3 — Textual inversion:

        
    
          
  • New token train — concept embed।
    • 
          
  • Lightweight (~কয়েক KB)।
    • 
          
  • Limited stylistic control।
    • 
        

        
Approach 4 — ControlNet:

        
    
          
  • Layout control — wedding card structure।
    • 
          
  • Reference photo guide।
    • 
          
  • Pose-based layout।
    • 
          
  • Production-friendly।
    • 
        

        
Recommended pipeline:

        
Step 1 — Dataset:

        
    
          
  • ৫০-১০০ high-quality wedding card image collect।
    • 
          
  • Manual caption — "Bengali wedding invitation, traditional motifs..."।
    • 
          
  • Style keyword consistent।
    • 
        

        
Step 2 — LoRA train:

        
# Using kohya_ss or diffusers LoRA
from diffusers import StableDiffusionPipeline
from peft import LoraConfig, get_peft_model

lora_config = LoraConfig(
    r=16, lora_alpha=32,
    target_modules=["to_q", "to_k", "to_v"],
    lora_dropout=0.1)

# Train on wedding card dataset
# ~৫০০ steps, learning rate 1e-4

        
Step 3 — Bangla text rendering:

        
    
          
  • Diffusion text quality poor — especially Bangla।
    • 
          
  • Approach: generate background separately, overlay text।
    • 
          
  • InstructPix2Pix for layout edit।
    • 
          
  • Post-process Photoshop/PIL।
    • 
        

        
Step 4 — ControlNet for layout:

        
    
          
  • Sketch input — model fill।
    • 
          
  • Canny edge — structure preserve।
    • 
          
  • Depth — layered design।
    • 
        

        
Workflow integration:

        
    
          
  • Web UI (AUTOMATIC1111 / ComfyUI)।
    • 
          
  • Designer-friendly interface।
    • 
          
  • Batch generation।
    • 
          
  • Manual selection ও refinement।
    • 
        

        
Cost analysis:

        
    
          
  • Training: ~$10 (cloud GPU)।
    • 
          
  • Inference: ~$0.01 per image।
    • 
          
  • Storage: minimal (LoRA <1GB)।
    • 
          
  • Per-customer cost dramatically low।
    • 
        

        
Quality concerns:

        
    
          
  • Hands, faces — diffusion weakness।
    • 
          
  • Wedding card mostly typography + decorative — manageable।
    • 
          
  • Iterative — designer + AI combo।
    • 
        

        
Bangla-specific challenges:

        
    
          
  • Religious motif diversity — multiple LoRA।
    • 
          
  • Cultural subtlety — designer review essential।
    • 
          
  • Color symbolism (red wedding, yellow Mehendi)।
    • 
          
  • Conjunct character text rendering — separate pipeline।
    • 
        

        
Production lessons:

        
    
          
  • AI-generated draft → designer refine।
    • 
          
  • Speed boost ১০x.
    • 
          
  • Cost reduction significant।
    • 
          
  • Cultural accuracy — human in loop।
    • 
        

        
Alternative — Midjourney/DALL-E:

        
    
          
  • No fine-tune needed।
    • 
          
  • Higher cost per image।
    • 
          
  • Less control।
    • 
          
  • Quick start option।
    • 
        

        

          মূল উপলব্ধি: Bangla wedding card — LoRA + ControlNet pipeline practical। Cultural data curation key। Bangla text — separate handle। AI + designer hybrid production winner। Bangladesh creative AI opportunity vast। Cost-effective scale-up।
        

      

    


    

      
        প্র ০৪
        Diffusion bias problems — race, gender stereotype। Sora-এর video generation-এ এই issue magnified। Practical mitigation কী?
      
      

        

          AI ethics — diffusion deployment-এ critical concern। Mitigation evolving।
        

        
Bias sources:

        
    
          
  • Training data — internet-scrape biased।
    • 
          
  • "Doctor" → male, "nurse" → female stereotypes।
    • 
          
  • Race/skin color — Western-centric।
    • 
          
  • Beauty standard narrow।
    • 
        

        
Sora video amplification:

        
    
          
  • Static bias → motion bias।
    • 
          
  • Stereotypical action — "doctor diagnosing", "construction worker"।
    • 
          
  • Cultural setting default Western।
    • 
          
  • Movement pattern stereotype।
    • 
        

        
Documented issues:

        
    
          
  • Stable Diffusion — racial skew documented।
    • 
          
  • Bloomberg study (২০২৩) — occupation bias।
    • 
          
  • "CEO" vs "secretary" gender split extreme।
    • 
        

        
Mitigation strategies:

        
(১) Data curation:

        
    
          
  • Diverse demographic representation।
    • 
          
  • Cultural variety include।
    • 
          
  • Historical bias awareness।
    • 
          
  • Manual auditing।
    • 
        

        
(২) Prompt engineering:

        
    
          
  • Explicit demographic specification।
    • 
          
  • "diverse" keyword often help।
    • 
          
  • Negative prompts — stereotype avoid।
    • 
          
  • Practical but inelegant।
    • 
        

        
(৩) Model fine-tuning:

        
    
          
  • RLHF for bias reduction।
    • 
          
  • Adversarial debiasing।
    • 
          
  • Fairness-aware loss।
    • 
          
  • Active research area।
    • 
        

        
(৪) Post-processing:

        
    
          
  • Generation review pipeline।
    • 
          
  • Demographic balancing।
    • 
          
  • Manual filter।
    • 
          
  • Slow but accurate।
    • 
        

        
(৫) System-level controls:

        
    
          
  • Usage policy enforcement।
    • 
          
  • Sensitive prompt blocking।
    • 
          
  • Watermarking generated content।
    • 
          
  • Provenance tracking।
    • 
        

        
Bangladesh-specific concerns:

        
    
          
  • Bangla face under-represented training।
    • 
          
  • Cultural attire stereotyped।
    • 
          
  • Religious sensitivity — Islamic, Hindu visual norms।
    • 
          
  • Wedding/festival inaccurate often।
    • 
        

        
Practical Bangladesh deployment:

        
    
          
  • Local fine-tuning — Bangla data essential।
    • 
          
  • Cultural review board।
    • 
          
  • Religious advisor consultation।
    • 
          
  • Designer human-in-loop।
    • 
        

        
Industry initiatives:

        
    
          
  • Stability AI — model card disclosure।
    • 
          
  • OpenAI — DALL-E 3 safety system।
    • 
          
  • Adobe Firefly — licensed training data।
    • 
          
  • Watermarking standard (C2PA)।
    • 
        

        
Regulatory landscape:

        
    
          
  • EU AI Act — high-risk AI regulation।
    • 
          
  • US executive order ২০২৩।
    • 
          
  • Bangladesh — emerging framework।
    • 
          
  • Compliance increasing burden।
    • 
        

        
Deepfake concerns:

        
    
          
  • Sora — realistic video misuse risk।
    • 
          
  • Election interference।
    • 
          
  • Non-consensual content।
    • 
          
  • Detection arms race।
    • 
        

        
Detection tools:

        
    
          
  • Forensic AI — diffusion artifact identify।
    • 
          
  • Watermark verification।
    • 
          
  • Metadata authentication।
    • 
          
  • Education public awareness।
    • 
        

        
Ethical use guidelines:

        
    
          
  • Disclose AI-generated content।
    • 
          
  • Avoid deceptive use।
    • 
          
  • Respect consent।
    • 
          
  • Cultural sensitivity priority।
    • 
        

        

          মূল উপলব্ধি: Diffusion bias real, video amplifies। Mitigation multi-layer — data, prompt, fine-tune, policy। Bangladesh — local fine-tune + cultural review। Detection ও watermarking essential। Responsible deployment industry-wide concern। Tech ethics — practical engineering issue।
        

      

    

  


  

    
 অনুশীলন

    
    
      
  1. 
        Schedule compute: $T = 4$, $\beta = (0.1, 0.2, 0.3, 0.4)$ — $\bar{\alpha}_T$ কত?
        
    
          
          
    
            
      
              
    • $\alpha_t = (0.9, 0.8, 0.7, 0.6)$।
      • 
              
    • $\bar{\alpha}_T = 0.9 \times 0.8 \times 0.7 \times 0.6 = 0.3024$।
      • 
              
    • $\sqrt{\bar{\alpha}_T} \approx 0.55$ — original signal weight।
      • 
              
    • $\sqrt{1 - \bar{\alpha}_T} \approx 0.835$ — noise weight।
      • 
              
    • $T = 4$-এ already noise-dominated।
      • 
            
    
          
    
        
    
      
    2. 

      
  2. 
        Code: Pretrained Stable Diffusion দিয়ে ১টি image generate করুন (huggingface diffusers library)।
        
    
          
          
    
            
    from diffusers import StableDiffusionPipeline
import torch

pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float16
).to("cuda")

prompt = "A traditional Bengali village at sunrise, oil painting"
image = pipe(prompt, guidance_scale=7.5,
             num_inference_steps=30).images[0]
image.save("village.png")
    
          
    
        
    
      
    3. 

      
  3. 
        চিন্তা: "Image space" diffusion vs "latent space" diffusion — কেন latent diffusion practical breakthrough?
        
    
          
          
    
            
    ৫১২×৫১২×৩ = ৭৮৬K dim — directly diffuse করা — memory ও compute prohibitive। VAE দিয়ে compress 6৪×৬৪×৪ = ১৬K dim — ~৪৮x reduction। Latent space-এ diffusion fast, decoder একবার চালিয়ে image।
    
            
    এই trick-ই Stable Diffusion-কে consumer GPU-তে চালানোর মূল কারণ। Pretrained VAE freeze করে শুধু diffusion train — modular design।
    
          
    
        
    
      
    4. 
    

  


  

    
 আরও পড়ুন · ABCL TECH-এ আপনার পরবর্তী পদক্ষেপ

    

    

      কোড রানার কাজ না করলে? ব্রাউজারে কাজ না করলে
      Google Colab
      ব্যবহার করুন — Google-এর ফ্রি অনলাইন Python পরিবেশ, শুধু Gmail অ্যাকাউন্ট লাগে।
    

  







  
  
     এই পাঠে ভুল পেয়েছেন? জানান
  





  
    
 পূর্ববর্তী পাঠ

    
পাঠ ৩৬ · GAN