AIGC Landscape:
Image · Video · Audio · 3D · Digital Humans

1. What is AIGC? Five sub-domains at a glance

AIGC is the application layer of generative AI — not "chatting" but "producing". By modality, mainstream scenarios fall into 5 sub-domains:

• Image generation: text / image → image (Stable Diffusion, Midjourney, FLUX)
• Video generation: text / image → video (Sora 2, Veo 3, Kling)
• Audio & music: text → voice / music (Suno, ElevenLabs)
• 3D generation: text / image → 3D models (Tripo3D, Hunyuan3D 2.0)
• Digital humans: text / audio → virtual-human video / live (HeyGen, MuseTalk)

In 2025-2026, all five sub-domains have simultaneously entered a "product-explosion phase": each has 1-2 flagship products, a handful of open-source models, and an entire workflow ecosystem. The key to learning AIGC is not memorizing the product list, but understanding the three shared layers: prompt engineering + control tools + workflow orchestration.

2. Image generation: from Stable Diffusion to FLUX and Midjourney V7

Image generation was the first AIGC sub-domain to break out; 2024-2026 saw three rounds of iteration — "closed-source catches up to open-source → open-source overtakes → closed-source takes the lead again".

2.1 Closed-source SaaS: the commercial default

2.2 Open-source models: commercial-friendly, locally deployable

2.3 GUI tools: who actually runs these models?

Open-source models need GUI tools to be used daily. The 2025 landscape:

• ComfyUI (58k+ stars, node-based workflow): maximum controllability + reproducibility, steep learning curve, ideal for complex production pipelines. Runs SD 1.5 on a 4 GB GPU.
• Automatic1111 / Forge / SD.Next (154k+ stars): classic WebUI, easy to start, the richest plugin ecosystem (ControlNet, LoRA, one-click install).
• Fooocus (38k+ stars): minimalist SDXL-based UI — "as simple as Midjourney, but free and open-source".
• InvokeAI: professional-grade, supports canvas inpainting and node workflows.
• Draw Things: native macOS / iOS, optimized for Apple Silicon.

2.4 Control and editing: SD's real moat

The biggest limit of closed-source models is the lack of "fine-grained control" — this is the moat the open-source ecosystem has built over 5 years:

• ControlNet (lllyasviel, GitHub 30k+ stars): use Canny edges, depth maps, pose skeletons, etc. to control composition precisely. The de-facto standard for "character consistency + multi-view generation".
• IP-Adapter (Tencent ARC Lab): "a LoRA of one image", used for style transfer and face consistency. The Plus Face version in 2024 dramatically improved facial fidelity.
• Segment Anything (Meta SAM / SAM 2): zero-shot image segmentation — auto-generate inpainting masks.
• InstructPix2Pix: edit images with natural-language instructions, natively supported in ComfyUI.
• LoRA / LyCORIS: lightweight fine-tuning — train a custom style or character from dozens of images, file size typically 10-200 MB.

These control tools are the reason "the open-source ecosystem is still competitive" — even the strongest closed-source model can rarely support arbitrary pose / depth / edge maps.

3. Video generation: from Sora 2 to the Kling + Veo + Seedance three-way fight

OpenAI Sora kicked off the video-generation era in February 2024. By 2025-2026, the field has split into "overseas flagships + Chinese contenders + open-source chasers".

3.1 Overseas flagships: physical-consistency arms race

3.2 Chinese contenders: full reversal in 2025-2026

2024-2025 was the "comeback year" for Chinese video models. On physical consistency, character dynamics, and pricing, multiple products have decisively overtaken their overseas counterparts:

3.3 Selection guide: which video model should you use?

4. Audio and music: TTS, music generation, voice cloning

Audio AIGC is the most "productized" sub-domain — Suno and ElevenLabs are "out-of-the-box" with no GUI required.

4.1 TTS (text-to-speech): from tool to Voice Agent

4.2 Music generation: from "AI-sounding" to "studio-grade"

4.3 Voice cloning & voice conversion

• Open-source RVC / So-VITS-SVC: clone with just 5-10 minutes of audio, GitHub 30k+ stars, the de-facto standard for live-stream voice changing.
• OpenVoice (MyShell): cross-language cloning, 2-second sample, supports tone / emotion / rhythm control.
• XTTS-v2 (Coqui): multilingual TTS cloning, Apache 2.0 commercial license.
• GPT-Sovits: Chinese-English voice cloning, great for live streams and podcasts.
• CosyVoice / F5-TTS (Alibaba): open-source mainstays for Chinese TTS.

5. 3D generation: from text-to-model to NeRF / 3DGS asset pipelines

3D generation is the youngest but fastest-growing AIGC sub-domain. From 2024-2026 it has moved from "experimental toy" to "production tool", with gaming, e-commerce, and VR/AR adopting it at scale.

5.1 Text / image → 3D model: from single image to PBR assets

5.2 NeRF / Gaussian Splatting (3DGS)

3DGS won the SIGGRAPH 2023 Best Paper; in 2024-2026 it became the de-facto successor to NeRF — real-time rendering + fast training + low VRAM.

• 3D Gaussian Splatting reference implementation (graphdeco-inria, 18k+ stars): the official CUDA implementation.
• awesome-3D-GS (mrnerdube, 3k+ stars): papers + open-source project list — must-read.
• gsplat (nerfstudio-project): PyTorch-optimized implementation, 2-3× faster than the original.
• Polycam / Luma AI / KIRI Engine: mobile 3DGS scanning apps, 4.7+ App Store ratings.

The killer feature of 3DGS is "photo-realistic real-time rendering" — it is becoming the foundation of the next generation of gaming, AR/VR, and digital-twin tech.

5.3 Game / VR asset generation

Gaming is the biggest landing site for 3D AIGC. The mainstream workflow:

Concept art (SD/FLUX) → Tripo3D / Meshy model → Blender retopology → Substance textures → Unity / Unreal integration
                ↓
        Combine with 3DGS for scene reconstruction — drastically cuts environment-art cost

Empirically, 3D AIGC compresses single-character modeling from 3 weeks to 2 hours.

6. Digital humans and virtual avatars: from recorded video to real-time interaction

Digital humans are the most commercially mature AIGC track — Synthesia was valued at $2.1B in 2024, HeyGen hit $35M ARR, and Lemon Slice just raised a $10.5M seed. Capital and demand are both there.

6.1 Video digital humans: pre-recorded scenarios

6.2 Real-time interactive digital humans: virtual customer service / live stream

• MiniMax Chat (Xingye / Talkie): role-play + long-context real-time conversation — China's consumer-facing digital-human standard-bearer.
• ByteDance Doubao Virtual Human: built into the Doubao app, customizable enterprise digital employees.
• Baidu Digital Human: Baidu Intelligent Cloud Xiling — focuses on live commerce and enterprise services.
• Lemon Slice-2 (YC + Matrix invested): 20B-parameter diffusion model, single-image virtual humans, real-time 20 FPS streaming on a single GPU.

6.3 Open-source real-time digital-human stacks

In real-time interaction, open source has overtaken closed source — real-time scenarios have extreme demands for latency, self-hosting, and customization:

Real-time digital-human core stack (open-source) is essentially a "TTS + LLM + digital-human rendering + lip-sync" 4-piece suite:

# The minimal real-time digital-human Agent
from RealtimeTTS import TextToAudioStream
from openai import OpenAI
from musetalk_runner import MuseTalkRealTime
import sounddevice as sd

# 1. Load the MuseTalk real-time model (30+ FPS on a V100)
musetalk = MuseTalkRealTime(avatar_image="avatar.png")

# 2. TTS streaming (Edge-TTS / CosyVoice / ElevenLabs)
tts = TextToAudioStream(engine="edge", voice="zh-CN-Xiaoxiao")

# 3. LLM inference + tool calls
llm = OpenAI()

def chat(user_text):
    stream = llm.chat.completions.create(
        model="gpt-4o", stream=True,
        messages=[{"role": "system", "content": "You are a digital-human assistant"},
                  {"role": "user", "content": user_text}]
    )
    for chunk in stream:
        text = chunk.choices[0].delta.content or ""
        # 4. TTS → audio → MuseTalk lip-sync → stream
        for audio_chunk in tts.feed(text):
            musetalk.feed_audio(audio_chunk)
            frame = musetalk.get_frame()  # 30 FPS video frames
            yield frame, audio_chunk

7. Workflows and ecosystem: from point tools to production pipelines

Wiring the 5 sub-domains together is what makes you an "AIGC engineer" rather than an "AIGC player". Here are the two highest-frequency production pipelines of 2025-2026.

7.1 ComfyUI node-based workflow: stitch every tool together

ComfyUI is the "Photoshop" of AIGC engineers — a canvas where every model, LoRA, ControlNet, IP-Adapter, AnimateDiff, and Video Helper can be visually orchestrated.

Core node ecosystem (2025):

• CheckpointLoader / DualCLIPLoader / LoraLoader: model and LoRA loading.
• ControlNetLoader + ApplyControlNet: pose / edge / depth control.
• IPAdapter Unified Loader + IPAdapter Advanced: character consistency.
• AnimateDiff + SVD_img2vid_Conditioning: still image to video.
• IPNMS / SAM nodes: automatic mask generation.
• Video Helper Suite: video frame batch processing.

7.2 Pipeline #1: from script to a 60-second short

# A full production pipeline (pseudocode)
1. Script: Claude / GPT-5 writes the script, split into 5-10 shots
2. Character sheet (prompt sheet): for each character, write a "consistency prompt + IP-Adapter reference image"
3. Text-to-image: ComfyUI + SDXL / FLUX, generate 1080p per shot
4. Inpainting: Inpaint node to fix hands, text, brand logos
5. Video: AnimateDiff 5s short shots or SVD 4s transitions
6. Voice: ElevenLabs / CosyVoice clones per character
7. Lip-sync: MuseTalk aligns the voiceover to the mouth
8. Edit: CapCut / Premiere adds subtitles, BGM, transitions
9. Deliver: export a 1080p 60s short

The same workflow took 5-10 person-weeks in 2024; in 2025 a single person can do it in 1-2 days with ComfyUI + closed-source APIs.

7.3 Pipeline #2: multi-shot consistency (comics / short drama)

The core challenge of multi-shot / comic / short drama is "the same character cannot look different in different scenes". The 2025 best practice is the "IP-Adapter FaceID + ControlNet OpenPose + LoRA combo":

1. Prepare 5-10 character reference images (different angles, different expressions)
2. Use IPAdapter Plus Face to lock the facial features
3. Use ControlNet OpenPose to control the pose of each frame
4. Use LoRA to lock the art style / outfit
5. Generate shot by shot — character consistency > 95%

Advanced: InstantID / IP-Adapter FaceID v2 / PuLID for even finer ID consistency.

7.4 Copyright & compliance: AIGC's unavoidable boundaries

From 2024-2026, AIGC copyright has moved from "theoretical debate" to "real lawsuits":

• Training data: Suno / Udio sued for "YouTube stream-ripping"; Anthropic settled for $1.5B over pirated books (2025).
• Generated content: the US Copyright Office has clarified that purely AI-generated content is not copyright-protected, but human-AI collaboration is partially protected.
• Commercial licensing: Adobe Firefly (licensed data), FLUX schnell (Apache-2.0), Stable Audio commercial license — prefer these three for commercial use.
• Watermarking: Google SynthID, Adobe Content Credentials (CAI C2PA standard) — the industry is adopting them.
• Portrait rights: HeyGen / Sora 2 Cameos both require "live verification" to prevent misuse. China's Interim Measures for the Management of Generative AI Services requires training data and generated content to be filed and compliant.

Risk is manageable for individual use; for enterprise commercial use, choose only models with clearly licensed training data (Firefly / FLUX schnell / Adobe Stock dataset) + human review on critical assets + C2PA watermarking.

8. Data dashboard: AIGC industry landscape

As of October 2025 (sources: Artificial Analysis, company websites, CB Insights, The Information):

The AIGC industry in 2025 totals roughly $15-20B, projected to break $40B in 2026 (enterprise ToB deployment + short-video content production).

9. Beginner path: from zero to shipping

Compressed into a "4-week learning plan", one sub-domain per week:

Week 1: image generation (most mature, richest ecosystem)

1. Sign up for Midjourney, run 50 prompts, understand "aesthetic styles" and "negative prompts".
2. Locally install ComfyUI + FLUX schnell (Apache-2.0, low VRAM), get a basic text-to-image working.
3. Add ControlNet + IP-Adapter and make a "character consistency" test image.
4. Deliverable: 10 high-quality images + 1 character-consistency workflow.

Week 2: video generation (hottest of 2025)

1. Sign up for Kling 2.6, run 20 shots, understand the difference between "image-to-video" and "text-to-video".
2. Compare with Sora 2 / Veo 3 to feel the generation gap in physical consistency.
3. In ComfyUI, install Video Helper Suite, get AnimateDiff + SVD pipeline working.
4. Deliverable: 3 short 15-30s videos (character-consistent + voiced).

Week 3: audio + digital humans

1. In ElevenLabs, record 30 seconds of your own voice, clone it, and generate multilingual TTS.
2. Use HeyGen to make a 60-second digital-human intro video, translate it into 5 languages.
3. Locally install MuseTalk + LiveTalking and ship a real-time digital-human conversation demo.
4. Deliverable: 1 multilingual digital-human video + 1 real-time interaction demo.

Week 4: 3D + workflow integration

1. Use Tripo3D / Hunyuan3D 2.0 to output a few character 3D models, import them into Blender.
2. In ComfyUI, string together the full workflow: character reference image → IP-Adapter → text-to-image → Inpaint fix → AnimateDiff to video → ElevenLabs voice → MuseTalk lip-sync.
3. Deliverable: 1 60s complete AIGC short + the full ComfyUI workflow file.

10. Decision guide: which tool to pick?

Five sub-domains, each with a "3-step selector" flow.

10.1 Image generation

• Just need one image, no fiddling → Midjourney V7 or Recraft V3
• Need fine-grained control (pose / character consistency) → ComfyUI + FLUX + IP-Adapter + ControlNet
• Commercial-compliance priority (brand assets) → Adobe Firefly Image 4
• Free + commercial + local → FLUX.1 schnell + ComfyUI

10.2 Video generation

• English ad / cinematic feel → Sora 2 / Veo 3 / Runway Gen-4.5
• Chinese scenarios (talking head / live stream) → Seedance 2.0 / Kling 2.6
• Character dynamics (SOTA) → Kling 2.0/3.0
• Image-to-video (given first frame) → Kling 2.0 or Vidu Q2
• Budget < ¥1/sec → Wan 2.6 / Seedance 2.0 (720P) / Kling 2.5 Turbo

10.3 Audio / music

• Commercial TTS / digital-human voiceover → ElevenLabs / MiniMax Voice Design
• Real-time streaming (< 300ms) → Cartesia Sonic / Inworld TTS-1.5
• Music generation (studio-grade) → Suno v5.5 / Udio
• Music with no copyright risk for commercial use → Adobe Firefly Music / Stable Audio 2
• Open-source voice cloning → RVC / GPT-Sovits / CosyVoice

10.4 3D generation

• Game asset generation → Tripo3D v3.0 + Blender
• Open-source + PBR → Hunyuan3D 2.0/2.5 (Tencent)
• Ultra-fast single-image reconstruction → Stable Fast 3D / TripoSR (open-source 0.5s)
• NeRF / 3DGS scene reconstruction → 3DGS original / Luma AI / Polycam

10.5 Digital humans

• Pre-recorded digital-human video → HeyGen Avatar 3.0 (best for Chinese / English / Japanese / Korean)
• Enterprise + 230+ preset avatars → Synthesia 2.0
• Real-time interaction / virtual customer service → LiveTalking (open-source) + MiniMax Chat
• Live streaming / anime VTuber → Open-LLM-VTuber / ChatVRM + MuseTalk

11. Chapter summary

• ✅ AIGC's five sub-domains — image, video, audio/music, 3D, digital humans — all entered a "product-explosion phase" in 2025-2026, with clear open-source / closed-source tiers in every track.
• ✅ Image: closed-source Midjourney V7 + open-source FLUX 1.1 Pro/Ultra Raw compete on two tracks. ComfyUI + ControlNet + IP-Adapter is the must-have stack for control-class applications.
• ✅ Video: Sora 2 / Veo 3 overseas flagships + Kling / Seedance / Vidu / Hailuo Chinese contenders + Wan / CogVideoX open-source chasers. In 2025 Chinese models overtake on physical dynamics and price.
• ✅ Audio: Suno v5.5 (studio-grade + voice cloning), ElevenLabs (Voice Design + multilingual), MiniMax Speech form the TTS/music big three. Open-source voice cloning RVC / GPT-Sovits is the standard.
• ✅ 3D: Tripo3D v3.0 + Hunyuan3D 2.5 (Tencent open-source) + TRELLIS (Microsoft) are the main text/image-to-3D options. 3DGS is becoming the foundation of the next generation of scene reconstruction.
• ✅ Digital humans: HeyGen / Synthesia / D-ID are commercially mature. LiveTalking / MuseTalk / Open-LLM-VTuber open-source real-time stacks overtake closed source.
• ✅ The real moat is "workflow orchestration" — ComfyUI visualizes the wiring of all models and tools, AnimateDiff + SVD + IP-Adapter enables character-consistent multi-shot narratives.
• ✅ Copyright and compliance: Suno / Udio are already being sued by major labels. For commercial use, prefer models with clearly licensed data (Firefly / FLUX schnell / Stable Audio).