Split View: 비전 모델 개발·파인튜닝 완전 가이드 2026 — CNN, ViT, DETR, SAM 2, VLM 까지 실전 의사결정 트리

The 2026 Vision Model Development & Fine-Tuning Guide — CNN, ViT, DETR, SAM 2, VLMs and a Real Decision Tree

Prologue — The Real Question a 2026 Vision Engineer Faces

A one-line ticket on a Monday morning in May 2026:

"Mark defective parts in our factory line camera feed. 12 classes, 300k images/day, accuracy 95% or better, runs on a Jetson Orin Nano at 30 fps."

In 2018 the answer would have been obvious — ResNet50 backbone, RetinaNet head, COCO pretraining, fine-tune on your data. Done.

In 2026 there are roughly eight answers.

1. Just fine-tune YOLOv11/12.
2. Use RT-DETR for transformer-based detection.
3. Run SAM 2 for masks and stack a classifier on top.
4. Prompt a vision foundation model like Florence-2.
5. Send the photo to Gemini 2.5 Vision or Claude Vision and parse the natural-language result.
6. Extract CLIP embeddings and do kNN classification.
7. Use OWLv2 / Grounding DINO for text-prompted zero-shot detection.
8. Pipeline two or three of the above.

This guide is about when, why, and how to choose among those eight. Not "the latest SOTA wins" — a decision tree across data size, accuracy target, latency budget, and operating cost.

The core skill of a 2026 vision engineer is no longer "train a model." It is "decide which model to train, and whether to train at all."

1. Architecture Families — From CNN to VLM at a Glance

Every vision model is "image to something." The "something" decides the family.

Memorize this table. The next eight chapters expand each row.

Core principle: every vision model ultimately "sees the image as a sequence of tokens" — a CNN as a spatial grid, a ViT as a patch sequence, SAM together with prompts, a VLM as input tokens to an LLM. Representation defines the model.

2. CNNs Are Not Dead — Where They Win in 2026

Despite marketing claims that ViT ate everything, CNNs are very much alive in 2026. Especially in these situations.

Pick a CNN when

1. Data is small — under 10k labeled images. ViT without pretraining barely learns.
2. You deploy at the edge — Jetson, Coral, mobile. A ConvNeXt-Tiny beats a ViT-Tiny at the same FLOPs.
3. Latency is brutal — sub-millisecond. A small CNN can hit 0.5 ms on a GPU.
4. Resolution is huge — 4K medical imaging. ViT patch counts explode.

One-line model loading with timm

PyTorch Image Models (timm) remains the de-facto standard for vision backbones in 2026. Over 1,000 pretrained backbones, one line.

import timm
import torch

# ConvNeXt v2 large, ImageNet-22k pretrain, 22k-1k fine-tune
model = timm.create_model(
    'convnextv2_large.fcmae_ft_in22k_in1k_384',
    pretrained=True,
    num_classes=12,  # our task's class count
)

# The model tells you the input transform it expects
data_config = timm.data.resolve_model_data_config(model)
transform = timm.data.create_transform(**data_config, is_training=True)

# Tensor shape including the batch dim is `B x C x H x W`
x = torch.randn(2, 3, 384, 384)
logits = model(x)  # shape: 2 x 12

Use timm.list_models('convnext*', pretrained=True) to see candidates. The pretrained_cfg contains mean, std, input_size, and crop_pct, so transforms stay consistent.

CNN training recipe — making small data work

• Pretrain to head-only training to full fine-tuning — three stages.
• Mixup, CutMix, RandAugment — almost mandatory below 10k labels.
• EMA (Exponential Moving Average) — 1–2pp of validation accuracy for free.
• Cosine schedule with a short warmup — OneCycleLR works too.
• AdamW with weight decay 0.05 — forget the old SGD.

3. ViT — The Champion When Data Is Plentiful

A ViT slices an image into patches like 16x16, treats them as a sequence, and stacks a Transformer on top. The key insight was "a model with less inductive bias beats a CNN if you give it enough data."

ViT variants worth knowing in 2026

• Swin Transformer v2 — window attention, efficient at high resolution.
• DeiT III — data-efficient training recipe.
• EVA-02 — masked image modeling, scaled to 22B parameters.
• DINOv2 — self-supervised, a powerful backbone trained without labels.
• SigLIP / SigLIP 2 — contrastive learning, strong image-text embeddings.

When to pick a ViT

One-line ViT classifier with Hugging Face transformers

from transformers import AutoImageProcessor, AutoModelForImageClassification
import torch
from PIL import Image

processor = AutoImageProcessor.from_pretrained('facebook/dinov2-large')
model = AutoModelForImageClassification.from_pretrained(
    'facebook/dinov2-large',
    num_labels=12,
    ignore_mismatched_sizes=True,  # reinitialize the head
)

img = Image.open('part.jpg').convert('RGB')
inputs = processor(images=img, return_tensors='pt')
with torch.no_grad():
    outputs = model(**inputs)
pred = outputs.logits.argmax(-1).item()

Swap facebook/dinov2-large for microsoft/swinv2-large-patch4-window12-192-22k and the API is identical.

4. Object Detection — YOLO vs DETR vs RT-DETR

Detection solves "what is where" simultaneously. As of 2026 the practical space splits in two.

The YOLO family — overwhelming production share

From Ultralytics YOLOv8 through v12, plus YOLO-NAS, YOLOv9, YOLOv10. For speed and deployment ergonomics, YOLO still wins.

from ultralytics import YOLO

# Load a pretrained model
model = YOLO('yolo11x.pt')

# Fine-tune on your dataset
results = model.train(
    data='factory_parts.yaml',  # train/val paths plus class names
    epochs=100,
    imgsz=640,
    batch=32,
    device=0,
    optimizer='AdamW',
    lr0=0.001,
    cos_lr=True,
    patience=20,  # early stopping
    project='runs/factory',
)

# Export to ONNX (edge deployment)
model.export(format='onnx', dynamic=True, simplify=True)
# Export to TensorRT (Jetson)
model.export(format='engine', half=True)

factory_parts.yaml:

path: /data/factory
train: images/train
val: images/val
names:
  0: scratch
  1: dent
  2: discoloration
  3: missing_screw

YOLO's weaknesses: NMS-based detection struggles on dense or tiny objects, and global context is weaker than DETR-class models.

The DETR family — NMS-free transformer detection

DETR removed NMS by using object queries and Hungarian matching. In 2026 the practical choice is RT-DETR.

from transformers import RTDetrV2ForObjectDetection, RTDetrImageProcessor

processor = RTDetrImageProcessor.from_pretrained('PekingU/rtdetr_v2_r50vd')
model = RTDetrV2ForObjectDetection.from_pretrained(
    'PekingU/rtdetr_v2_r50vd',
    num_labels=12,
    ignore_mismatched_sizes=True,
)

YOLO vs DETR — the decision

When OpenMMLab fits

The mmdetection, mmsegmentation, and broader OpenMMLab ecosystem shines in research and experimentation. You can compare 50-plus detection models in one codebase and swap backbones with a config line. But the learning curve is steep, deployment is a separate toolchain, and for a production-first model Ultralytics ships much faster.

5. Segmentation and SAM 2

Segmentation is "where, per pixel." In 2026 SAM 2 is the starting point for almost every case.

SAM 2 — promptable segmentation for both images and video

Released by Meta in July 2024, SAM 2 takes a click, box, or mask prompt and "segments any object, then auto-tracks it across video frames." The core ideas:

• Unified image + video — one model for both.
• Memory attention — past-frame representations are stored to track objects over time.
• Promptable — segmentation is driven by user input (clicks, boxes), not predefined masks.

from sam2.build_sam import build_sam2
from sam2.sam2_image_predictor import SAM2ImagePredictor
import numpy as np
from PIL import Image

sam2 = build_sam2('configs/sam2.1/sam2.1_hiera_l.yaml', 'sam2.1_hiera_large.pt')
predictor = SAM2ImagePredictor(sam2)

img = np.array(Image.open('part.jpg'))
predictor.set_image(img)

# One click to a mask
point_coords = np.array([[450, 320]])
point_labels = np.array([1])  # 1 = foreground
masks, scores, _ = predictor.predict(
    point_coords=point_coords,
    point_labels=point_labels,
    multimask_output=True,
)

How to actually use SAM 2 — "SAM segments, something else labels"

SAM 2 does not tell you what something is. It says "this mask is object 1, that mask is object 2." For semantic labels you need a two-stage pipeline.

1. SAM 2 produces masks.
2. Crop each mask and classify with CLIP or SigLIP.

This is the 2026 standard recipe for "zero-shot segmentation." Works without any labeled data.

Where HQ-SAM, MobileSAM, and EfficientSAM fit

• HQ-SAM — sharper boundaries for medical and satellite imagery.
• MobileSAM / EfficientSAM — lightweight for edge devices.
• SAM 2.1 — improved video tracking precision.

6. VLMs — The Era of Talking to Images in Natural Language

A Vision-Language Model "turns images into input tokens for an LLM." Users ask in natural language; the model answers in natural language.

The major VLMs in 2026

Using a VLM "just by prompting"

When you barely have labels, when classes change weekly, or when "explain why" is required — sending a photo and a system prompt to a VLM gives the highest ROI.

import anthropic
import base64

client = anthropic.Anthropic()

with open('part.jpg', 'rb') as f:
    img_b64 = base64.standard_b64encode(f.read()).decode()

resp = client.messages.create(
    model='claude-opus-4-7',
    max_tokens=512,
    system=(
        'You are an automotive parts inspector. Look at the photo and answer ONLY '
        'with this JSON. schema: {"defect": one of [scratch, dent, discoloration, '
        'missing_screw, none], "severity": one of [low, medium, high], '
        '"reasoning": short string}'
    ),
    messages=[{
        'role': 'user',
        'content': [
            {'type': 'image', 'source': {'type': 'base64', 'media_type': 'image/jpeg', 'data': img_b64}},
            {'type': 'text', 'text': 'Inspect this part.'},
        ],
    }],
)
print(resp.content[0].text)

VLM limits — they are not always the right answer

• Cost — 1 to 5 cents per image. 300k images/day means 90k–450k USD/month. Compare that to training a custom model.
• Latency — 200 ms to 2 s. Useless at 30 fps on edge.
• Non-determinism — slightly different answers on the same photo. Threshold-based decisions need calibration.
• Ignoring JSON — must be enforced via response_format or tool use.
• Data governance — can the photo legally go to an external API? Loop in legal before you ship.

7. Which Task to Which Architecture — The Matrix

Eight chapters compressed into one table. When picking the first model, this table alone gets you 80% of the way there.

One-line rule: Small data, large pretrained backbone plus a small head. Large data, custom training. No labels, VLM or embeddings.

8. Data — How Much, How, From Where

The cliche "data matters more than the model" is still true in 2026.

How much data you actually need (rough numbers)

Public datasets still alive in 2026

• ImageNet-22k / 1k — the unchanging benchmark for classification pretraining.
• COCO 2017 — detection, keypoints, captions; still the standard.
• Open Images V7 — 9M+ images with weak labels.
• LAION-5B / DataComp — large image-text pairs for CLIP-style training (check copyright).
• LVIS — 1,200+ classes, long-tail detection.
• ADE20K, Cityscapes, Mapillary — segmentation.
• DocVQA, ChartQA, InfographicVQA — document and chart VQA.
• OpenX-Embodiment, Ego4D — robotics and first-person video.
• SA-1B — SAM training data, 1.1B masks.

Labeling tools — a practical 2026 comparison

The 2026 labeling secret: auto-label with SAM 2 plus Grounding DINO first, then have humans correct in Roboflow or CVAT. Labeling time drops 5–10x.

9. Train vs Fine-Tune vs Prompt — The Cost Model

Rule of thumb: as data grows, training becomes economical; as inference traffic grows, owning the model becomes economical. You have to balance both axes.

LoRA / QLoRA — the practical standard for VLM fine-tuning

Full fine-tuning a VLM is painful even with 4x 80 GB VRAM. LoRA is the answer.

from transformers import AutoModelForVision2Seq, AutoProcessor
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer, SFTConfig
import torch

model_id = 'Qwen/Qwen2.5-VL-7B-Instruct'
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForVision2Seq.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map='auto',
)

lora_cfg = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=['q_proj', 'k_proj', 'v_proj', 'o_proj'],
    lora_dropout=0.05,
    bias='none',
    task_type='CAUSAL_LM',
)
model = get_peft_model(model, lora_cfg)
model.print_trainable_parameters()  # usually 0.1–1% trainable

cfg = SFTConfig(
    output_dir='runs/qwen-vl-defect',
    num_train_epochs=3,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=1e-4,
    lr_scheduler_type='cosine',
    warmup_ratio=0.03,
    bf16=True,
    gradient_checkpointing=True,
    logging_steps=20,
    save_strategy='epoch',
)

# train_dataset is a sequence of {"image": PIL.Image, "messages": [...]}
trainer = SFTTrainer(model=model, args=cfg, train_dataset=train_ds, processing_class=processor)
trainer.train()

QLoRA (4-bit quantization plus LoRA) lets you fine-tune a 7B VLM on a single 24 GB GPU. A 70B-class model fits on a single 80 GB card.

10. Deployment — ONNX, TensorRT, Core ML, TFLite

Training is the end of the beginning. A bad deployment choice multiplies your inference cost tenfold.

Recommended formats by target

PyTorch to ONNX to TensorRT flow

import torch
import torch.onnx
from ultralytics import YOLO

model = YOLO('runs/factory/best.pt')

# 1) ONNX
model.export(format='onnx', dynamic=True, simplify=True, opset=17)

# 2) TensorRT (NVIDIA GPU) — directly supported by Ultralytics
model.export(format='engine', half=True, workspace=4)

For manual conversion, trtexec --onnx=model.onnx --saveEngine=model.engine --fp16 is the simplest path. INT8 quantization needs a calibration dataset.

Core ML and TFLite — mobile

# Core ML
import coremltools as ct
mlmodel = ct.convert(traced_model, inputs=[ct.ImageType(shape=(1, 3, 224, 224))])
mlmodel.save('Model.mlpackage')

# TFLite (PyTorch to TF, ai-edge-torch recommended)
import ai_edge_torch
edge_model = ai_edge_torch.convert(model.eval(), sample_inputs)
edge_model.export('model.tflite')

11. Failure Modes — The Real Problems You Hit in Production

The field aphorism: "If accuracy isn't moving, don't swap the model — go look at the data. Four times out of five, it's the data."

12. The Decision Tree — Decide in 30 Seconds

When a new vision problem lands on your desk, ask these in order.

1. "Can a VLM solve this in one prompt?" — Hand-feed Claude/Gemini Vision ten photos. If you get 90% accuracy, that's your baseline.
2. "Do you have labels?" — No, then VLM or zero-shot (CLIP, Grounding DINO, SAM 2).
3. "Do you have 10k+ labels?" — Yes, train your own. No, pretrain plus fine-tune.
4. "Does it run at the edge?" — Yes, CNN or a small YOLO. No, ViT is fair game.
5. "30 fps or higher?" — Yes, YOLO plus TensorRT/Core ML. No, DETR-class is fine.
6. "Do you need to explain the decision?" — Yes, VLM or attention visualization. No, a normal model.
7. "Are errors expensive?" (medical, autonomous driving) — Yes, ensembles, calibration, human-in-the-loop.

Seven questions handle 80% of the first-model decision.

Epilogue — Checklist, Anti-Patterns, Coming Next

Pre-launch checklist for your first vision model

• Have you looked at label distribution? (Class imbalance must be known before training.)
• Are train and val splits free of time- or source-overlap?
• Do you have a baseline? (Simplest model, or human accuracy, or majority class.)
• Have you measured one-prompt VLM accuracy and cost?
• Have you compared training cost against inference cost?
• Do you know the memory and latency limits of your deployment target?
• Have you measured inter-annotator agreement (IAA)?
• Is there a hand-curated set of 100 corner cases as a separate eval?
• Monitoring — how do you detect production drift?
• Rollback — can you go back to the previous model when this one breaks?

Common anti-patterns

• "What's the latest SOTA?" as the first question — go look at your data first.
• Starting from a ViT with no pretraining — almost always fails below 10k labels.
• Tuning while peeking at the val set repeatedly — that's training data. Hold out a real test set.
• Shipping on mAP alone — also check per-class PR curves, small-object metrics, and inference latency.
• Trusting VLM output without post-processing — JSON-schema validation, fallbacks, caching.
• Forcing one model to handle every class — split frequent classes from the long tail.
• Treating augmentation as an afterthought — often augmentation matters more than model choice.
• Train-eval transform mismatch — the number-one debug time sink.

Coming next

• "Operating vision models — drift detection, A/B, canary, and active learning loops."
• "Auto-labeling with VLMs — a Claude/Gemini/Qwen-VL pipeline that automates 90% of annotation."
• "Edge vision inference — running the same model on Jetson, Coral, iPhone, and Android."

References

Architecture papers

• ViT — "An Image is Worth 16x16 Words" — https://arxiv.org/abs/2010.11929
• Swin Transformer v2 — https://arxiv.org/abs/2111.09883
• ConvNeXt v2 — https://arxiv.org/abs/2301.00808
• DINOv2 — https://arxiv.org/abs/2304.07193
• DETR — "End-to-End Object Detection with Transformers" — https://arxiv.org/abs/2005.12872
• RT-DETR — https://arxiv.org/abs/2304.08069
• SAM — https://arxiv.org/abs/2304.02643
• SAM 2 — https://arxiv.org/abs/2408.00714
• LLaVA — https://arxiv.org/abs/2304.08485
• Qwen2.5-VL — https://arxiv.org/abs/2502.13923
• InternVL — https://arxiv.org/abs/2312.14238
• Grounding DINO — https://arxiv.org/abs/2303.05499
• Florence-2 — https://arxiv.org/abs/2311.06242
• LoRA — https://arxiv.org/abs/2106.09685
• QLoRA — https://arxiv.org/abs/2305.14314

Tools and libraries

• timm (PyTorch Image Models) — https://github.com/huggingface/pytorch-image-models
• Hugging Face transformers — https://huggingface.co/docs/transformers
• Ultralytics YOLO — https://docs.ultralytics.com/
• OpenMMLab MMDetection — https://github.com/open-mmlab/mmdetection
• Segment Anything 2 — https://github.com/facebookresearch/sam2
• PEFT (LoRA/QLoRA) — https://github.com/huggingface/peft
• TRL (SFT) — https://github.com/huggingface/trl

Labeling tools

• Label Studio — https://labelstud.io/
• CVAT — https://www.cvat.ai/
• Roboflow — https://roboflow.com/
• V7 Darwin — https://www.v7labs.com/
• Encord — https://encord.com/

Datasets

• ImageNet — https://www.image-net.org/
• COCO — https://cocodataset.org/
• Open Images V7 — https://storage.googleapis.com/openimages/web/index.html
• LAION — https://laion.ai/
• LVIS — https://www.lvisdataset.org/
• ADE20K — https://groups.csail.mit.edu/vision/datasets/ADE20K/
• SA-1B — https://ai.meta.com/datasets/segment-anything/

Deployment / inference

• ONNX Runtime — https://onnxruntime.ai/
• NVIDIA TensorRT — https://developer.nvidia.com/tensorrt
• Apple Core ML Tools — https://github.com/apple/coremltools
• ai-edge-torch (PyTorch to TFLite) — https://github.com/google-ai-edge/ai-edge-torch
• OpenVINO — https://docs.openvino.ai/

VLM APIs

• Anthropic Claude Vision — https://docs.anthropic.com/en/docs/build-with-claude/vision
• Google Gemini Vision — https://ai.google.dev/gemini-api/docs/vision
• OpenAI Vision — https://platform.openai.com/docs/guides/vision