2nd/10_Wiki/Topics/AI_and_ML/이미지 생성 최적화 (Image Generation Optimization).md

---
id: wiki-2026-0508-이미지-생성-최적화-image-generation-opti
title: 이미지 생성 최적화 (Image Generation Optimization)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Image Gen Optimization, Diffusion Inference Optimization, 이미지 생성 가속]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [ai, image-generation, optimization, inference, diffusion]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: diffusers-tensorrt
---

# 이미지 생성 최적화 (Image Generation Optimization)

## 매 한 줄
> **"매 latency × cost × quality 의 trilemma 를 step reduction, quantization, compilation 으로 동시 해결"**. 2026 의 production image gen 은 distillation (4-step Schnell, Lightning, LCM), quantization (FP8/INT4), graph compilation (TensorRT, torch.compile), batch fusion 을 통해 50-step 30s → 4-step 0.5s 로 압축한다. 매 quality 손실 은 perceptual eval 에서 < 5%.

## 매 핵심

### 매 optimization axes
- **Steps**: 50 → 4 (distillation).
- **Precision**: FP32 → FP16 → FP8 → INT4.
- **Compilation**: eager → torch.compile → TensorRT.
- **Caching**: KV cache, prompt embed cache, latent cache.
- **Resolution**: 1024 → progressive (256→512→1024).
- **Batching**: dynamic batching, continuous batching.

### 매 distillation 기법
- **LCM**: Latent Consistency Model, 4-step.
- **SDXL Lightning**: 1/2/4/8-step variants.
- **Hyper-SD**: 1-step possible.
- **FLUX Schnell**: 4-step out-of-box.
- **DMD2**: distribution matching, single-step quality.

### 매 응용
1. Realtime gen 의 sub-second UX (Krea, Magnific).
2. On-device mobile gen (Core ML, MLC).
3. Mass batch render 의 throughput max.

## 💻 패턴

### Step reduction (LCM-LoRA)
```python
from diffusers import StableDiffusionXLPipeline, LCMScheduler
import torch

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
).to("cuda")
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
pipe.load_lora_weights("latent-consistency/lcm-lora-sdxl")

# 4-step gen
img = pipe(prompt, num_inference_steps=4, guidance_scale=1.0).images[0]
# 50-step (3.5s) → 4-step (0.4s) on A100
```

### torch.compile
```python
pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
pipe.vae.decode = torch.compile(pipe.vae.decode, mode="reduce-overhead")

# warmup
_ = pipe("warmup", num_inference_steps=4)
# 1.4-2x speedup after warmup
```

### TensorRT (production)
```python
# Export → TensorRT engine
from polygraphy.backend.trt import EngineFromNetwork, NetworkFromOnnxPath, TrtRunner

# 1. ONNX export
torch.onnx.export(pipe.unet, dummy_inputs, "unet.onnx",
                  opset_version=17, dynamic_axes={...})

# 2. trtexec build
# trtexec --onnx=unet.onnx --saveEngine=unet.plan --fp16 --memPoolSize=workspace:8192

# 3. Runtime
with TrtRunner(EngineFromNetwork(NetworkFromOnnxPath("unet.onnx"))) as r:
    out = r.infer({"sample": x, "timestep": t, "encoder_hidden_states": h})
# 2-3x faster than torch.compile
```

### FP8 quantization (Hopper / Ada)
```python
from optimum.quanto import quantize, qfloat8, freeze

quantize(pipe.transformer, weights=qfloat8, activations=qfloat8)
freeze(pipe.transformer)
# memory: 24GB → 13GB; latency: 1.3x faster on H100
```

### Prompt embed cache
```python
import hashlib, pickle
from pathlib import Path

class EmbedCache:
    def __init__(self, dir="./.embed_cache"):
        self.dir = Path(dir); self.dir.mkdir(exist_ok=True)

    def get_or_compute(self, prompt, encoder_fn):
        key = hashlib.sha256(prompt.encode()).hexdigest()
        p = self.dir / f"{key}.pt"
        if p.exists(): return torch.load(p)
        emb = encoder_fn(prompt)
        torch.save(emb, p)
        return emb

cache = EmbedCache()
emb = cache.get_or_compute(prompt, pipe.encode_prompt)
# repeat prompt: skip text encoder entirely
```

### Continuous batching (server)
```python
# vLLM-style continuous batching for diffusion (sdxl-batched-server)
from collections import deque
import asyncio

class BatchedServer:
    def __init__(self, max_batch=8, wait_ms=20):
        self.q = deque(); self.max_batch = max_batch; self.wait_ms = wait_ms

    async def submit(self, prompt):
        fut = asyncio.Future(); self.q.append((prompt, fut))
        return await fut

    async def loop(self):
        while True:
            await asyncio.sleep(self.wait_ms/1000)
            if not self.q: continue
            batch = [self.q.popleft() for _ in range(min(len(self.q), self.max_batch))]
            prompts = [p for p,_ in batch]
            imgs = pipe(prompts).images
            for (_, fut), img in zip(batch, imgs): fut.set_result(img)
```

### Progressive resolution
```python
# Cascade: 256 → 512 → 1024
img_lo = pipe(prompt, height=256, width=256, num_inference_steps=8).images[0]
img_md = img2img_pipe(prompt, image=img_lo, strength=0.5,
                       height=512, width=512, num_inference_steps=8).images[0]
img_hi = img2img_pipe(prompt, image=img_md, strength=0.3,
                       height=1024, width=1024, num_inference_steps=8).images[0]
# Total cost < single-pass 1024
```

### MLX (Apple Silicon)
```python
import mlx.core as mx
from mlx_diffusion import StableDiffusion

sd = StableDiffusion("stabilityai/sdxl-turbo", float16=True)
img = sd.generate("a cat", n_steps=4, n_images=4)
# M3 Max: 4-step 1024px in ~1.2s
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| latency critical | distill (4-step) + TensorRT |
| memory tight | FP8/INT4 quantize |
| Apple device | MLX |
| repeat prompts | embed cache |
| many concurrent | continuous batch |
| highest quality | full 50-step + xformers |

**기본값**: 4-step LCM/Lightning + torch.compile + FP16, escalate to TRT for >10 RPS.

## 🔗 Graph
- 부모: [[AI 이미지 생성 (AI Image Generation)]]
- Adjacent: [[TensorRT]] · [[torch.compile]] · [[오픈소스 이미지 모델 미세 조정 및 배포]]

## 🤖 LLM 활용
**언제**: bottleneck profiling interpretation, kernel fusion plan, deploy config.
**언제 X**: low-level CUDA kernel writing — Triton/cutlass docs 직접 참조.

## ❌ 안티패턴
- **Optimize before profile**: nvtx/torch profiler 없이 추측.
- **Over-distillation**: 1-step 이라 quality cliff — perceptual eval 누락.
- **Quantize without calib**: dynamic quant 만으로 quality 폭락.
- **Single-process bottleneck**: GIL 무시한 sync server.

## 🧪 검증 / 중복
- Verified (LCM paper Luo 2023, SDXL Lightning ByteDance 2024, NVIDIA TRT-LLM docs).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — distillation + quantize + compile stack. |