How Chinese Game Studios Cut Character Art Cost 80% with AI

The studios you have probably never heard of

When Western coverage of Chinese AI talks about game studios, it usually means miHoYo (Genshin Impact), Tencent, or NetEase. The teams actually driving the per-character cost collapse are two tiers down from that. They are 30 to 120 person studios in Hangzhou, Chengdu, Xiamen, and the Pudong outskirts of Shanghai, mostly shipping mid-budget mobile titles: gacha RPGs targeting Southeast Asia, anime-style card battlers, idle games for the Japanese market, and increasingly mini-games inside WeChat and Douyin.

Their scale matters because it explains why AI hit them harder than it hit the giants. A typical title in this segment ships with 80 to 250 playable characters at launch and adds 6 to 15 per month forever. Each character historically needed concept art, three to five rarity tiers, two or three "skin" variants per year, splash art for banners, chibi sprites, sticker packs for WeChat, vertical-format teaser art for Xiaohongshu, and Douyin-friendly motion clips. A 100-character roster meant something like 4,000 unique art deliverables across a two-year live-ops cycle.

Until 2023 those deliverables came from a mix of in-house artists earning roughly USD 1,500 to 3,500 per month and outsourcing studios in second-tier cities charging USD 200 to 800 per finished character piece depending on rarity. Hero-tier splash art for a flagship character ran USD 2,000 to 5,000 once revisions were counted. By the time a studio hit 200 characters, art was 35 to 50 percent of total production cost.

What changed is not that AI got good. It is that a specific stack of Chinese tools, mostly built between mid-2024 and early 2026, lined up well enough to replace the rough-pass artist while keeping the polish artist. The result, according to operators we spoke with at two Hangzhou studios shipping into the global gacha market, is per-character cost down roughly 75 to 85 percent compared with their 2023 pipeline.

The actual pipeline, step by step

The workflow is consistent enough across studios that it functions almost as a regional standard. The names will be unfamiliar even to Western art directors who follow Stable Diffusion closely.

Step 1: Brief generation

The art lead sits with the planning lead and writes a one-paragraph character brief. Most teams now run that brief through Doubao, ByteDance's Chinese-first chatbot, rather than a Western LLM. Doubao understands the genre conventions of Chinese mobile games (the 二次元 / erciyuan anime aesthetic, the specific archetypes of "cool senior sister," "tragic swordsman," "loli necromancer") with much less prompting than Claude or GPT, and it produces output formatted for Chinese-trained image models. Some studios use DeepSeek-V3 instead, especially after the early 2026 price cuts brought it to roughly USD 0.14 per million input tokens.

The brief gets converted into two artifacts: a Chinese-language reference prompt, and an English-language prompt for fallback to Western models.

Step 2: First-pass image generation

This is where the stack diverges most from Western practice. Three tools handle the bulk of generation:

• Seedream 3.0, ByteDance's image model accessed through Jimeng (即梦), has become the default for studios producing anime-style art. It handles Chinese-language prompts natively, hits character consistency targets that MidJourney still struggles with on Eastern aesthetics, and costs roughly USD 0.04 to 0.08 per generation depending on tier. Studios buy in bulk and average closer to USD 0.03.
• Liblib.art (libliai.com), the dominant Chinese model-sharing platform, hosts tens of thousands of fine-tuned LoRA models contributed by community artists. A studio will train its own character LoRA, usually on a base of Pony Diffusion XL or an Illustrious derivative, then run it through Liblib's hosted ComfyUI for production batches. LoRA training runs roughly USD 8 to 30 depending on dataset size.
• MidJourney via proxy, still used for splash art and "international taste" pieces because its painterly output remains harder to replicate. Studios pay around USD 30 per month for proxy access on top of the base subscription.

A character usually goes through 40 to 200 generations before locking a base image. Total compute cost for that exploration phase: roughly USD 5 to 15.

Step 3: ComfyUI for production iteration

Once a base image is locked, the work moves to a self-hosted ComfyUI install running on a single 4090 or rented compute from autodl.com (the de facto AI GPU rental marketplace, charging about USD 0.40 to 0.70 per hour for a 4090). The studio's character LoRA, plus a face-consistency LoRA trained on the chosen design, lets a single artist crank out the full rarity tier set, three skin variants, and the chibi version in a single afternoon.

This is the step Western studios routinely underestimate. The Chinese teams are not asking AI to nail a one-shot output. They are running custom LoRAs, IP-Adapter, ControlNet pose transfer, and inpainting in a 12 to 20 node ComfyUI graph that took the technical art lead two months to refine. The graph itself is the moat.

Step 4: Human polish

Every studio we spoke with kept the senior artist. The pipeline is two-stage: AI generates, human paints over. A polish pass for a standard character runs 90 minutes to four hours depending on rarity. The artist fixes hands, redraws weapons that AI fumbled, adjusts the face for emotional consistency with the character's personality, and repaints fabric folds where the AI output looks generic.

Studios that tried to skip this step in 2024 got savaged in player community reviews on Bilibili and TapTap. The phrase "AI 味" (AI flavor) became shorthand for lazy work. By mid-2025 the polish step was non-negotiable for any character a player would actually look at for more than three seconds.

Step 5: Animation and motion

Static art ships first, then comes the motion package. Two tools dominate:

• Kling 2.1, from Kuaishou, generates 5 to 10 second character intro animations from a still image plus a motion prompt. Pricing runs roughly USD 0.30 to 0.80 per generation. Studios use it for the "summon animation" players see when pulling the character.
• Vidu, from Shengshu Tech, handles longer scene-style clips and is the preferred tool for character-versus-character battle teaser videos.

For full character animation inside the game, AI is still mostly auxiliary: it produces reference loops that the in-house animator then cleans up in Spine 2D or Live2D Cubism. But the marketing motion content, the stuff that runs on Xiaohongshu and Douyin, is now 70 to 90 percent AI generated according to two studios we spoke with.

Step 6: Marketing asset cascade

This is the step that genuinely has no Western analogue. The same character art gets fed into a marketing pipeline that produces, for one character launch:

• A vertical 9:16 Xiaohongshu post (Xiaohongshu being the Instagram-meets-Pinterest platform that drives most discovery for women-skewed mobile games)
• A Douyin teaser clip cut in Jianying (the domestic Chinese version of CapCut, with deeper AI features than the international build)
• A WeChat sticker pack of 16 to 24 stickers, generated by feeding the locked character art back through Seedream with expression prompts
• A WeChat Channel (视频号) short
• A Bilibili dynamic post with character lore

All five assets get produced from the locked base art in roughly one working day by a single operator using Jianying's AI cut features and a Seedream batch job. The marketing-to-art ratio collapses from a small team of editors to one person.

Cost breakdown in USD

Numbers below are averaged from two studios shipping mid-tier gacha titles, cross-checked against pricing screenshots from Liblib, Jimeng, and Kling as of early 2026. Treat them as roughly accurate to the segment, not precise to any one company.

2023 baseline, traditional pipeline, per standard character (4 rarity tiers, 1 skin):

• Concept art: USD 250 to 400
• Three additional rarity tiers: USD 600 to 1,200
• One skin variant: USD 200 to 400
• Splash art: USD 800 to 2,000
• Chibi sprite: USD 80 to 150
• Marketing asset cut: USD 200 to 500
• Total: roughly USD 2,150 to 4,650 per character

2026 AI-assisted pipeline, same deliverables:

• Compute and tool costs: USD 15 to 35
• LoRA training, one-time per character: USD 10 to 30
• Senior artist polish, four to eight hours total: USD 80 to 200
• Splash art polish, longer pass: USD 150 to 350
• Marketing asset operator, one day: USD 60 to 120
• Total: roughly USD 315 to 735 per character

The midpoint comparison lands at about USD 3,400 traditional versus USD 525 AI-assisted, an 84 percent reduction. Studios reporting "80 percent cost cut" in Chinese press are not exaggerating, though the savings concentrate in the rough-pass labor that used to go to outsourcing studios. Senior artist headcount at the studios we looked at stayed flat or grew slightly.

What Western creators can actually copy

Most of the pipeline is portable. A few pieces are not.

Directly copyable:

• The two-stage AI-then-polish workflow. This is the single highest-leverage idea and it works regardless of base model. Civitai LoRAs, ComfyUI graphs, and SDXL fine-tunes give Western teams the same production substrate.
• Custom character LoRAs trained on locked concept art. The technique is universal, well-documented, and the bottleneck is dataset curation rather than compute.
• Splitting the marketing asset cascade off as a downstream operator role rather than artist work. This is an org chart change as much as a tool change.
• Using a 12 to 20 node ComfyUI graph as the production unit rather than chasing one-shot prompts. The graph carries the studio's style; the prompt is just a parameter.

Hard to replicate without adaptation:

• Seedream and Doubao. Both work best with Chinese-language prompting and Chinese-aesthetic priors. Western teams shipping anime-style art can approximate via Pony, Illustrious, or NoobAI fine-tunes plus a Japanese or English prompt, but the out-of-the-box quality gap on erciyuan styles is real.
• Kling and Vidu. Western alternatives like Runway Gen-3, Pika, and Luma exist, but per-clip pricing is two to four times higher and the camera language differs (Western models trend cinematic-realistic, Chinese models trend anime-stylized).
• Jimeng and Jianying integration. The fact that ByteDance owns CapCut, Jimeng, Doubao, and Douyin lets a studio run an end-to-end pipeline inside one ecosystem. Adobe plus Runway plus ChatGPT is the rough Western equivalent but the seams are visible.
• The marketing asset operator role itself. It depends on Xiaohongshu, Douyin, and WeChat Channels existing as the discovery substrate. Western platforms reward different content shapes, so the cascade has to be redesigned, not just translated.

The most replicable Chinese practice is probably the simplest: pay a senior artist well, give them eight hours of polish budget per character, and let an AI rough pass eat the four days that used to come before that.

Cultural and regulatory caveats

Three things shape the Chinese pipeline that do not apply, or apply differently, in Western jurisdictions.

Algorithm filing. Since August 2023, generative AI services offered to the Chinese public must file algorithm registrations with the Cyberspace Administration of China and pass a content security review. Studios using only internal AI tooling for art production largely sidestep this, but if they expose any generation feature to players (a "generate your own banner" feature, for example) they hit the filing requirement. This is why almost no Chinese game ships a player-facing generative feature even though the internal tooling is mature.

Watermarking and labeling. Rules effective September 2025 require AI-generated content distributed publicly to carry both visible labels and embedded metadata. Marketing material on Xiaohongshu and Douyin that started as AI generation is supposed to be flagged. Enforcement is uneven; the practical effect is that studios run their AI-generated splash art through a polish pass partly to make the "predominantly human-authored" claim defensible.

Banhao scrutiny. The 版号 game license process now asks about AI use in production. Approval still happens routinely for AI-assisted titles, but the studio has to document its human-in-the-loop workflow. This is one structural reason the senior polish artist did not get fired.

Training data. Chinese model providers have trained on data with a lighter copyright footprint than their Western counterparts. This is a benefit for output quality on anime-style work and a risk for any studio planning to ship the same art into US or EU markets where the legal posture around AI training data is shifting. The two studios we spoke with were both quietly building "clean-room" LoRAs trained only on assets they own, specifically to keep options open for Western publishing.

Aesthetic alignment. The dominant style in Chinese mobile games is anime-derived, which is the style AI handles best. Western teams shipping painterly fantasy, gritty realism, or stylized 3D will not see the same cost reduction; the AI rough pass simply is not as strong in those domains yet. The 80 percent number is real for the segment but it is not universal.

The honest summary is that Chinese mid-tier game studios reorganized around AI faster than their Western counterparts because the tools, the art style, the platform incentives, and the labor structure all aligned. The tools are increasingly portable. The other three pieces have to be rebuilt locally.