ComfyUI + AnimateDiff Open-Source Animation Workflow Research

1. Introduction

ComfyUI is a free, open-source, node-based generative-AI workflow GUI and backend developed by comfyanonymous starting in January 2023. Among the 8 subjects in this group, ComfyUI is the only one where all six layers are built entirely by the user: it has no default context policy, no default toolset, no default orchestration, no default asset store, no default evaluation, and no default governance.

This is both its greatest weakness (an extremely high barrier to entry) and its greatest strength (full controllability). For this group's core thesis, ComfyUI is an extreme contrast sample: when L4 is built entirely by yourself, what does continuous production of AI comic dramas actually require? The practice of the ComfyUI community provides a concrete answer — a character image bank, character LoRAs, style anchor images, and versioned workflow JSON.

1.1. Basic Project Information

Version note: ComfyUI uses rolling releases, with no traditional semantic version numbers. This document does not state a "version number" and instead uses GitHub star count and node count as indicators of development status.

1.2. Positioning & Design Philosophy

ComfyUI's positioning is a "free, open-source, node-based generative-AI workflow GUI + backend".

The original author's design philosophy (verbatim):

"Everyone is trying to make easy to use interfaces. Let me try to make a powerful interface that's not easy to use."

This sentence precisely captures the fundamental difference between ComfyUI and commercial platforms: commercial platforms optimize "ease of use", while ComfyUI optimizes "capability ceiling". For a scenario like AI comic drama that requires fine-grained control over every layer, this trade-off is reasonable — but it also means users must bear the full cost of building Harness themselves.

1.3. Pricing

Hidden cost checklist (any part for which no official pricing was found is marked [To be filled]):

• GPU hardware: recommended configuration is RTX 4090 + 16–24GB VRAM
• Model storage: cumulative footprint of base models, LoRAs, ControlNet, Motion Module, VAE, etc.
• Electricity and operation/maintenance
• Labor: workflow construction and debugging (this is the biggest hidden cost)

1.4. Open Formats

ComfyUI's open format is the most thorough in this group: the code is open source, workflows can be exported as JSON, it can run fully offline, and any node can be added by yourself.

2. Glossary of Terms

3. Feature Description

3.1. Node-Based Workflow

All of ComfyUI's capabilities are exposed as nodes. As of December 2024 it had 1,674 built-in node types, and in 2026 the official claim is 60,000+ nodes.

The node chain of a typical animation workflow:

Load Checkpoint
  → AnimateDiff Loader（mm_sd_v15_v2.ckpt、beta_schedule linear、motion_lora_v1S.safetensors）
  → CLIP Text Encode（正/负提示词）
  → AnimateDiff Sampling
  → VAE Decode
  → Frame Interpolation（FILM，film_net_fp32.pt，multiplier 2）
  → Video Concat
  → Video Encode（mp4/h264，fps 24，crf 18）

Common negative prompt terms: low quality, blur, painting, sketch, cartoon, anime, text, watermark, signature, jpeg artifacts, bad anatomy, extra limbs, extra fingers

3.2. AnimateDiff Animation Generation

Signal flow of the AnimateDiff + IP-Adapter scheme:

参考图像 → IP-Adapter（风格） → AnimateDiff + SD 模型 → VAE Decode → Video Save
驱动视频/动作 ────────────────────────────────┘

Key node packages:

Advanced techniques:

• VACE all-in-one video editing: motion transfer, video extension, first–last frames
• Depth ControlNet: keeps spatial structure consistent
• Same Motion Module: keeps inter-frame coherence (different Motion Modules cause motion-style jumps)
• SeedVR2: upscaling

"Infinite animation length": it is underpinned by sliding context window technology; as long as VRAM permits it can keep generating by scrolling. This is the only scheme in the group not bound by a fixed duration cap (versus the 15–16 seconds of commercial platforms).

Prompt Travel: keyframe descriptions are set on the timeline and the image evolves smoothly with the camera, used for storyboard transitions. This is the group's most direct "timeline context" mechanism.

3.3. Character & Style Anchoring

IPAdapter's weight 0.6–0.8 is the group's most practical consistency technique: use one style anchor image to unify the whole series' style without training a LoRA. This greatly lowers the barrier to character/style consistency — no need to train a model for each character.

Common failures and countermeasures:

3.4. Full Novel-to-Comic-Drama Pipeline

Core logic: novel text → AI storyboard parsing → batch generation of characters/scenes → unified style output

Five modules:

1. Text processing module

• Load Text (loads .txt/.md) → LLM Text Parser (local LLM or API breaks it down into storyboards) → Text to JSON (structured storyboard array)
• Storyboard prompt template output format: [scene number] - scene description / characters appearing / character actions / shot type (wide/medium/close-up/extreme close-up) / emotional atmosphere

1. Character generation module

• Character Definition (character-definition-table JSON, including basic description / clothing / hairstyle / expression library)
• → Batch Text Prompts → Load Checkpoint (recommended: anything-v5, meinamix, ghostmix, ponyDiffusionV6XL)
• → Load LoRA (strength_model 0.8~1.0) → CLIP Text Encode
• → KSampler (steps 20–30, cfg 7.0–8.0, DPM++ 2M Karras) → VAE Decode → Save Image

1. Scene generation module

• Scene Prompt Builder
• → IPAdapter Apply (weight 0.6–0.8, unify the whole series' style with one style anchor image)
• → ControlNet Apply (canny/depth/openpose/scribble, weight 0.8–1.2)
• → Batch KSampler → Batch Save Image

1. Character consistency maintenance module

• Character Image Bank (character image bank dictionary) → Character Detector → Image Paste → Inpaint Node (denoise 0.5–0.7)

1. Batch processing and automation module

• Batch Manager → JSON Iterator → Condition Switch (conditional branching: indoor/outdoor/fight take different ControlNets and Motion LoRAs) → Progress Logger

4. Platform Architecture

图 4-1｜ComfyUI 六层架构：从产品形态层到已上线模型

数据来源：基于本文分析绘制的示意图。

4.1. Overall Architecture

┌──────────────────────────────────────────────────────────────────┐
│  产品形态层  ComfyDesktop（本地）· ComfyCloud · ComfyAPI（端点化）   │
│              · ComfyEnterprise                                     │
├──────────────────────────────────────────────────────────────────┤
│  编排层  节点 DAG · JSON Iterator · Batch Manager                  │
│          Condition Switch（条件分支）· Progress Logger              │
├──────────────────────────────────────────────────────────────────┤
│  控制信号层  ControlNet（canny/depth/openpose/scribble）             │
│              IPAdapter · SparseCtrl · Inpaint · VACE                │
├──────────────────────────────────────────────────────────────────┤
│  节点与模型层  Load Checkpoint · Load LoRA · AnimateDiff Loader      │
│                KSampler · VAE Decode · FILM 插帧 · SeedVR2 超分      │
│                VideoHelperSuite · WD14 Tagger                       │
├──────────────────────────────────────────────────────────────────┤
│  资产层  角色图像库（Character Image Bank）· 角色 LoRA · 风格锚点图   │
│          工作流 JSON（可版本化）                                     │
├──────────────────────────────────────────────────────────────────┤
│  已上线模型  Seedance 2.5 · LTX 2.5 · Wan Animate 2 · MiniMax H3    │
└──────────────────────────────────────────────────────────────────┘

4.2. Node & Model Layer

ComfyUI's model layer is replaceable: base models (Checkpoint), LoRAs, Motion Modules, VAEs, ControlNet models and upscaling models are all chosen by the user. Models such as Seedance 2.5, LTX 2.5, Wan Animate 2 and MiniMax H3 are already online.

This "pluggable models" property lets ComfyUI accommodate any new model — including commercial-platform models (such as MiniMax H3's open-source weights). This is a flexibility commercial platforms cannot provide.

4.3. Control Signal Layer

The control signal layer is ComfyUI's core advantage over commercial platforms: ControlNet (canny/depth/openpose/scribble), IPAdapter, SparseCtrl, Inpaint and VACE are all available simultaneously, and they are combinable and weightable.

Commercial platforms usually wrap these control signals into fixed features (such as "motion control" or "first–last frames"), while ComfyUI treats them as freely wireable primitives.

4.4. Product Form Layer

ComfyAPI deserves special attention: it lets self-built complex workflows be directly exposed as HTTP endpoints, callable by external systems (including Agents). In function this is equivalent to PixVerse's CLI/API path, but the workflow is your own.

5. Harness Design

5.1. L1 Context Engineering Layer

ComfyUI's L1 is fully programmable, which is unique in this group:

Assessment: ComfyUI's L1 is not "strong" but "programmable" — it presets no policy, yet gives you the primitives to implement any policy. You can implement something like Hailuo's context compression (self-built compression nodes) or Seedance-like multimodal reference (multiple IPAdapters injecting in parallel).

Cost: every policy must be implemented by yourself. Having no defaults means there is no out-of-the-box experience.

5.2. L2 Tools & Execution Layer

ComfyUI's L2 is the fullest in this group: 60,000+ nodes (official claim).

Core tool list:

Assessment: no commercial platform's tool breadth can match it. But having all tools also means high selection cost — beginners faced with 60,000+ nodes will be at a loss.

Programmability: ComfyAPI can turn workflows into production endpoints, equivalent to API-ification. This is the group's engineering-integration path second only to PixVerse CLI/Skills, and the workflow is entirely your own.

5.3. L3 Orchestration & Control Layer

ComfyUI's L3 is the most flexible in this group: the node graph is DAG orchestration.

Assessment: ComfyUI is the only scheme that simultaneously has conditional branching, loops, batch processing, visual editing and endpoint-ification. PixVerse Canvas has visualization but no customization; Vidu has frame-level control but no node graph; Kling has intelligent storyboards but is not visualizable.

The dual identity of workflow JSON: it is both an orchestration definition and a versionable state carrier (see Section 5.4). This is ComfyUI's most underestimated feature — it turns orchestration itself into a git-manageable text file.

5.4. L4 Memory & State Layer

ComfyUI's L4 is the most controllable in this group: character image bank + character LoRA + style anchor image + workflow JSON.

Workflow JSON is ComfyUI's most distinctive L4 contribution: it turns "how this entire production pipeline is configured" into a text file that can be committed to git, code-reviewed, rolled back and reused. This is the only scheme in this group that achieves "orchestration is state, state is code".

For AI comic dramas, this solves a problem common to commercial platforms: when you need to redo Episode 12 three months later, can you still precisely reproduce the generation configuration of that time? With ComfyUI, the answer is yes (workflow JSON + random seed + model version); with most commercial platforms, the answer is not necessarily.

Gaps:

1. No plot state machine: character appearance can be anchored, but cross-episode plot state (character relationships, injuries, timeline) still has to be maintained outside ComfyUI.
2. No standard for the asset store: the Character Image Bank is a community practice, not a platform feature, and each team implements it differently.
3. No cross-session automatic retrieval: calling up character cards requires manual wiring or self-built retrieval logic.

5.5. L5 Evaluation & Observability Layer

ComfyUI's L5 is the weakest in this group: it depends on manual review and community benchmarks.

Available evaluation methods:

Gap list:

1. No automated quality-evaluation metrics (no FVD, no CLIP similarity, no automatic computation of character-consistency rate).
2. No regression-set mechanism — after model or parameter updates, there is no automated way to verify whether "a configuration that used to produce good images still does".
3. No trajectory tracking — although Progress Logger can record progress, it does not include structured logs of generation quality, time cost or resource consumption.
4. No A/B capability.

This is the layer ComfyUI most needs to be built in-house. Recommended approach: freeze 20–50 representative shots (including close-ups, action, night scenes, and scenes with multiple characters) into a regression set, and re-run and manually compare them after any model or parameter change; teams with engineering capability can introduce automated metrics (such as the CLIP-embedding similarity between character reference images and generated frames) as a quantitative proxy for consistency rate.

5.6. L6 Governance & Security Layer

ComfyUI's L6 has three distinct mechanisms:

Assessment: ComfyUI's L6 is strong in data sovereignty and auditability. All assets, models and intermediate products stay local and never pass through a third party; the code is open source and auditable; there is no risk of a platform shutting down.

Gaps:

1. No built-in compliance-labeling capability — ComfyUI does not automatically add an AI-generated synthetic-content label to output videos. Per the Measures for Labeling AI-Generated Synthesized Content and GB 45438—2025, users must build labeling nodes themselves.
2. No permission or audit mechanism — the community edition has no user system and no operation audit.
3. No cost guardrails — local deployment looks "free", but GPU time and labor are real costs, and there is no platform-level quota control.
4. GPL-3.0 copyleft — if you do secondary development based on ComfyUI and distribute it, mind license compliance.

Building compliance labels in-house is a hard requirement; see Section 6.3 for specifics.

5.7. Six-Layer Capability Matrix

6. Case Studies

6.1. Case 1: Five-Stage 3D Character Animation Pipeline

• Background: a static character reference image needs to be turned into coherent 3D-style animation while keeping the character's identity stable in every frame.
• Approach (GitHub project sanchitwadehra/Effecient-3D-Character-GenAI-Animation-Pipeline):

1. Character input: IPAdapter loads the reference image (Prep for ClipVision) and injects the character's identity into every frame
2. Driving video: the driving video is sampled into frames
3. Pose estimation: OpenPose (via ControlNet) extracts the pose skeleton frame by frame
4. Generation: AnimateDiff generates animation frames conditioned on the pose skeleton + IPAdapter character embedding
5. Render & combine: KSampler denoising → VAE decoding → Video Combine assembly

• Results: an 8-configuration benchmark (4 motion modules × LCM toggle), with v3sd15mm + LCM (8 steps) as the best. The tech stack includes ComfyUI, AnimateDiff-Evolved (v3_sd15_mm / mm_sd_v15_v2 / AnimateDiff-Lightning), ControlNet + OpenPose, ComfyUI_IPAdapter_plus, WD14 Tagger, EpicRealism (SD1.5), VideoHelperSuite and LCM.
• Harness reading: the essence of this pipeline is three-layer conditional stacking — character identity (IPAdapter) + motion skeleton (OpenPose/ControlNet) + motion style (Motion Module). Commercial platforms wrap these three layers into a single button; ComfyUI lets you adjust the weight of each layer separately. That is precisely what "controllable" concretely means.

6.2. Case 2: Automated Full Novel-to-Comic-Drama Pipeline

• Background: turning web-novel IP into comic dramas requires handling long texts, many characters and many scenes, while keeping a unified art style across the whole series.
• Approach: a five-module full pipeline (see Section 3.4) — text processing (LLM breaks it down into structured storyboard JSON) → character generation (LoRA strength 0.8–1.0) → scene generation (IPAdapter style anchor weight 0.6–0.8 + ControlNet weight 0.8–1.2) → character consistency maintenance (character image bank + Inpaint denoise 0.5–0.7) → batch processing and automation (Batch Manager + JSON Iterator + Condition Switch).
• Results: no specific throughput or consistency figures have been publicly disclosed; marked [To be filled]. The common failure countermeasures given by the community are: inconsistent character → raise LoRA strength to 1.0; inconsistent style → IPAdapter weight to 0.8; inaccurate motion → add ControlNet-openpose; slow → lower steps to 20.
• Harness reading: Condition Switch is the key to this pipeline — indoor/outdoor/fight take different ControlNets and Motion LoRAs, showing that ComfyUI's orchestration can do "scene-aware dynamic branching". This is the only scheme in this group that supports such conditional branching.

6.3. Case 3: Building a Compliance Watermark Node

• Background: the Measures for Labeling AI-Generated Synthesized Content took effect on September 1, 2025, and GB 45438—2025 imposes quantitative requirements on video explicit labels (start frame, corner position, font height no less than 5% of the shortest side, lasting no less than 2 seconds) and metadata implicit labels (AIGC field). ComfyUI has no built-in labeling capability, and among the 8 subjects in this group no commercial platform publicly declares a built-in label compliant with this standard.
• Approach: insert a self-built labeling node before ComfyUI's Video Encode node:

1. Explicit-label node: overlay text or graphics containing the elements "artificial intelligence (or AI)" + "generated/synthesized" on the edge or corner of the start frame (optionally including the end and intermediate frames); the font height must be no less than 5% of the shortest side length of the frame; at normal playback speed it must last no less than 2 seconds.
2. Implicit-label node: write the AIGC metadata field when wrapping into mp4, in the format {"AIGC": {"Label": "...", "ContentProducer": "...", "ProduceID": "...", "ReservedCode1": "...", "ContentPropagator": "...", "PropagateID": "..."}}; Label values: belongs to = 1, possible = 2, suspected = 3.

• Results: this group's search found no publicly available ready-made ComfyUI compliance-labeling node implementation; marked [To be filled].
• Harness reading: this case illustrates the two-sided nature of open-source solutions at L6 — you must build it yourself, but you actually can. Commercial platforms may already have built it (or may not, and you cannot verify), but you cannot confirm or modify it. For strictly-compliant scenarios, "you can build it and prove it" is often more reliable than "the platform says it is built in".

7. Summary

7.1. Strengths

1. All six layers fully programmable and self-built: no platform-preset constraints; any policy can be implemented.
2. L2 has the fullest toolset: 60,000+ nodes covering the full chain from loading, control and editing to encoding, upscaling and wrapping.
3. L3 is the most flexible: node DAG + conditional branching + loops + batch processing + ComfyAPI endpoint-ification.
4. L4 is the most controllable and versionable: character image bank + character LoRA + style anchor image + workflow JSON (git-manageable, rollback-able, reproducible).
5. The sliding context window enables "infinite animation length": the only scheme in this group not bound by a fixed duration cap.
6. IPAdapter style anchors need no training: unify the whole series' art style with one image (weight 0.6–0.8), without training a LoRA for each character.
7. Data sovereignty and auditability: fully local offline operation, GPL-3.0 code is auditable, no risk of platform shutdown.
8. Pluggable models: can integrate any available model such as Seedance 2.5, LTX 2.5, Wan Animate 2 and MiniMax H3.
9. Zero subscription cost: the software is free.

7.2. Limitations

1. Extremely high barrier to entry: the original author himself admits it is "a powerful interface that's not easy to use". You must understand many concepts such as base models, LoRA, ControlNet, Motion Module and sampler parameters.
2. L5 is the weakest: no automated quality evaluation, no regression set, no trajectory tracking; quality judgment relies entirely on the human eye.
3. L6 has nothing built in: it does not provide AI-generated synthetic-content labeling (must be built in-house); no permissions or audit; no cost guardrails.
4. High hidden costs: GPU (recommended RTX 4090 + 16–24GB VRAM), storage, electricity and labor (workflow construction and debugging).
5. No official SLA: the community edition makes no support commitments.
6. Fragmented asset ecosystem: the Character Image Bank is a community practice rather than a platform standard, making reuse across teams difficult.
7. No plot state machine: as with all platforms, cross-episode plot state must be maintained externally.
8. Rolling releases: no semantic version numbers; environment reproduction depends on manual record-keeping.
9. GPL-3.0 copyleft: distributing secondary development requires attention to license compliance.

7.3. Applicability Boundaries

7.4. Selection Recommendations

• The core reason to choose ComfyUI is controllability and data sovereignty, not cost. If your core need is "every layer must be modifiable, provable and reproducible", ComfyUI is the only choice.
• Put workflow JSON under version control: this is ComfyUI's greatest hidden value. Create a dedicated repository per series, managing workflow JSON, character-definition-table JSON, storyboard JSON, random seeds and model versions. This constitutes your real L4 assets.
• Prefer IPAdapter style anchors and use LoRA cautiously: style anchors (weight 0.6–0.8) unify the whole series' art style without training; LoRA (strength 0.8–1.0) is stronger but costs training. Get things working with anchors first, then add LoRA for major characters.
• You must build L5 yourself: establish a regression set of 20–50 representative shots (including close-ups, action, night scenes and multi-character shots), and re-run and compare after any model or parameter change. Teams with the means can introduce automated metrics (such as the CLIP-embedding similarity between reference images and generated frames) as proxies for consistency rate.
• You must build the L6 compliance-labeling node yourself: insert two nodes before Video Encode — an explicit label (start-frame corner, font height no less than 5% of the shortest side, lasting no less than 2 seconds) and an implicit label (AIGC metadata field) — and automatically validate every output.
• Use ComfyAPI as your engineering outlet: expose mature workflows via ComfyAPI as production endpoints, wired into your own orchestration system (including storyboard Agent, asset Agent and compliance Agent).
• Prepare hardware per the recommended configuration: RTX 4090 + 16–24GB VRAM, generation resolution kept at 768×768; do not underestimate storage needs (base models + LoRA + ControlNet + Motion Module + VAE accumulate significant footprint).
• Consider a hybrid approach: ComfyUI does not have to be the only solution. Use ComfyUI for character/style anchoring and batch generation (controllable, zero marginal cost), use commercial platforms (such as Vidu or Hailuo) to supplement difficult shots, and use ComfyUI to uniformly do labeling and wrapping.

Information Gap Statement

1. ComfyUI official version number: it uses rolling releases with no traditional semantic version number. This document substitutes the GitHub star count and node count and does not state a version number.
2. Pricing of ComfyCloud / ComfyAPI / ComfyEnterprise: no public price lists were found; marked [To be filled].
3. Ready-made compliance-labeling nodes: this group's search found no public ComfyUI compliance-labeling node (compliant with GB 45438—2025); marked [To be filled].
4. The scope difference between 60,000+ nodes and 1,674 node types: the former is the official 2026 claim (presumably including community custom nodes), the latter is the built-in node count as of December 2024; the two use different statistical scopes and have not been officially clarified.
5. "Fully local offline with no data upload": a community claim not confirmed in official documentation; marked [To be verified].
6. Full data of the 8-configuration benchmark: only the best is known to be v3sd15mm + LCM (8 steps); the specific metrics of the other 7 configurations have not been disclosed.
7. Actual throughput and consistency data of the full novel-to-comic-drama pipeline: community practice-sharing provided no quantitative results; marked [To be filled].
8. ComfyUI's adaptation to domestic chips / Xinchuang environments: no relevant information was found.
9. How AnimateDiff cooperates with the latest video models (Seedance 2.5, Wan Animate 2, MiniMax H3): Comfy's official site lists these models as live, but there is no description of how they cooperate with AnimateDiff workflows.
10. Comfy Org's governance structure and commercial sustainability: founded in June 2024; its business model and long-term support commitments have not been publicly disclosed.

8. References

1. Comfy official site — https://comfy.org/
2. AI Wiki, "ComfyUI" — https://aiwiki.ai/wiki/comfyui
3. AIBars, "ComfyUI" open-source library entry (2026-09-10 data) — http://www.aibars.net/en/library/open-source-ai/details/725005680951562240
4. GitHub — sanchitwadehra/Effecient-3D-Character-GenAI-Animation-Pipeline — https://github.com/sanchitwadehra/Effecient-3D-Character-Animation
5. GitHub — comfyanonymous/ComfyUI — https://github.com/comfyanonymous/ComfyUI
6. Open-Source Model Encyclopedia, "Workflow Usage Guide" (AnimateDiff + IP-Adapter scheme) — cited from Section 9.2 of Search Report R11
7. Practice Sharing, "ComfyUI + AnimateDiff Local Animation Workflow" — cited from Section 9.2 of Search Report R11
8. Practice Sharing, "Novel-to-Comic-Drama ComfyUI Workflow" — cited from Section 9.3 of Search Report R11
9. Baidu Baike, "AI Comic Drama" — https://baike.baidu.com/item/AI%E6%BC%AB%E5%89%A7/68788906
10. Baidu Baike, "Keyframe Animation" — https://baike.baidu.com/item/%E5%85%B3%E9%94%AE%E5%B8%A7%E5%8A%A8%E7%94%BB/10223838
11. 360 Baike, "Keyframe" — https://baike.so.com/doc/6737995-32354145.html
12. Measures for Labeling AI-Generated Synthesized Content by the CAC and three other departments (CAC Circular [2025] No. 2) — https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
13. Mandatory national standard GB 45438—2025, "Cybersecurity Technology — Methods for Labeling AI-Generated Synthesized Content" — https://www.tc260.org.cn/upload/2025-03-15/1742009439794081593.pdf
14. CNTV News, "How Will the Labeling, Dissemination and Review of AI-Generated Synthesized Content Develop?" — https://news.cctv.cn/2025/03/15/ARTI3wMX1ohsE7LsLT4qBVrv250315.shtml
15. Wancai Animation Master, "9 Beginner Knowledge Points for Quickly Understanding 2D Animation Video Production" — https://www.animiz.com.cn/blog-13952.html
16. Southern Metropolis Daily, "MiniMax's Third-Generation Video Model Is Open-Sourced, with the World's No. 1 Video-Editing Capability" (including ComfyUI day-one adaptation info) — https://www.toutiao.com/article/7669739176429912595/
17. The Paper, "1914 Yuan to Produce One Episode? Comic Dramas Are Still Trapped in Hidden Costs" — https://www.thepaper.cn/newsDetail_forward_33993493