TL;DR: I drew a comic character with one command, asked for the same character in a second scene, and got a stranger back wearing her hoodie. This post is the run that fixed it: four CLI commands, four panels, one lip-synced clip, every step stamped with its price and wall-clock time on 2026-09-17. The panel pipeline cost about eleven US cents.
The pitch you keep hearing is that AI comic drama is the next gold rush. The pitch is not wrong about demand, it is wrong about where the work actually is. Panel one is free-ish and easy; panel two is where your character dies, and almost no beginner tutorial says so out loud.
I'm not here to adjudicate the gold-rush claim. I can't verify anyone's view counts, so I won't cite them. What I can hand you is a measured run on my own machine, with the failure included.
What the gold rush looks like from inside
Before the technique, one piece of context that changes what this topic is. I pulled Zhihu's site-wide hot list on 2026-09-17, thirty items: zero comic-drama entries. Of the seven existing Zhihu questions about AI comic drama, five ask "can I make money from this" and none ask about technique. This is a livelihood vertical, not a mass spectacle, and the people in it ask about income first.
The stratification is blunt. Consumption is still growing: the August top-100 playback increment ran +44% month over month (second-hand, via DataEye). The supply-side dividend is closed: industry tracking puts active production studios at 1,216 falling to 698 within Q1, and revenue per 10k views down roughly tenfold (both second-hand). Tutorials are oversupplied and homogeneous. The empty niche is not "how to draw a panel"; it is what happens after you draw one, which is compliance and cost accounting. Both are below.
The gap nobody's tutorial names
Ask an image model for "a girl with round wire glasses, a bright yellow hoodie, black bob with straight bangs" and you get a perfectly good character sheet. Ask for that same girl on a rooftop at night and you get a different girl: glasses gone, hair a shade off, hoodie redesigned, face subtly someone else's. One broken panel and every panel after it is waste, because a comic is a sequence and a sequence needs one face.
That gap, between one nice illustration and an actual comic, is the whole game.
My setup, pinned so you can reproduce or distrust it: Alibaba Cloud Model Studio CLI bl 1.22.0, image model qwen-image-3.0, TTS cosyvoice-v3-flash, video wan3.0-video. Unit prices read from bl model list --model <name> --output json, not estimated. Runs happened 2026-09-17, 20:41 to 20:48 local time.
Step 1: write the identity down before you draw it
Vague features cannot be locked later. Three identifiable traits go into the prompt, and two flags keep the output usable downstream:
bl image generate --prompt "anime-style character sheet, a girl with round wire glasses, a bright yellow hoodie, black bob hair with straight bangs, half body, front view, calm slightly curious expression, clean lineart, cel shading, pure white background, no text" --model qwen-image-3.0 --size 3:4 --watermark false --prompt-extend false --out-dir . --out-prefix char-sheet
--watermark false, or every downstream panel carries a mark. --prompt-extend false, because prompt expansion invents details and a character sheet needs precision, not surprise.
Cost: ¥0.18, roughly two US cents.
Step 2: the lock is a reference image, not a seed
Here is my failure, in full, because it is the failure most tutorials hand you as advice. My first instinct was --seed. Fix the seed, I reasoned, and the character stays fixed. It does not. A seed reproduces the same image from the same prompt; change the scene description and you get a different person wearing similar clothes. I burned two panels learning that the seed is a reproducibility switch, not an identity switch.
The identity switch is reference-image editing. Feed the sheet from step 1 as --image, and in the prompt change only scene and composition while stating that face, hair, glasses and clothing stay identical:
bl image edit --image ./char-sheet.png --prompt "keep this character's face, hair, round glasses and bright yellow hoodie exactly unchanged; change only scene and composition: she stands at the edge of a rooftop at night, in profile, looking up; background is bokeh city lights and deep blue night sky; comic panel framing, cel shading" --model qwen-image-3.0 --size 3:4 --watermark false --out-dir . --out-prefix panel-1
61 seconds later:
I ran the same sheet through two more scenes to check whether the lock holds or whether I got lucky once:
Glasses, hoodie, bob, face shape: preserved across all three. The outputs even came back with comic panel borders. Per panel: ¥0.20 (¥0.18 edit plus ¥0.02 input), 61 to 70 seconds.
Scope, stated plainly because overclaiming here would be worse than useless: tested is three scenes and four poses (half body, profile, head-down, three-quarter). Untested is cumulative drift past three scenes, full-body action, costume changes, extreme close-ups, multi-character frames. Treat the tested range as tested.
One upgrade path and one trap from the wider community, both consistent with what I measured. The community consensus for consistency is layered stacking, and the turnaround or multi-angle character sheet is its accepted first step; my single sheet held three scenes, and a multi-angle sheet is the recommended upgrade once your storyboard multiplies angles beyond what one view can vouch for. The trap: accessories vanish unless you repeat them in the prompt. A first-hand overseas comparison of single-character references found exactly this, and it matches my own outputs, where the hoodie drawstrings survived only because the prompt kept naming the hoodie.
Step 3: the voice line, and the alignment trap
bl speech synthesize --text "原来城市晚上，这么亮啊。" --voice longcheng_v3 --format mp3 --out voice-1.mp3
Twelve Chinese characters, 3.264 seconds of audio, about ¥0.01. List voices with --list-voices instead of guessing IDs.
The trap: audio length and clip length are independent. My line is 3.26 seconds; in a 5-second clip the character spends the last 1.7 seconds with a closed mouth. Write dialogue that fills the duration, or take the silence as breathing room. Knowing which one you chose is the difference between direction and a bug you cannot find.
Step 4 (optional): make her talk
Movement plus lip sync in one command, with --image and --image-voice paired by position and referenced in the prompt as Image 1 and Audio 1:
bl video ref --prompt "Image 1 stands on the rooftop at night, looking up; night wind moves her hair and hoodie drawstrings; she speaks: Audio 1. Camera slowly pushes in; city lights bokeh and flicker" --image ./panel-1.png --image-voice ./voice-1.mp3 --model wan3.0-video --duration 5 --resolution 720P --ratio 9:16 --watermark false --download clip-1.mp4
Two flags will bite you. --resolution defaults to 1080P, the expensive tier; write 720P on your first clip. And the CLI warns these jobs take 5 to 10 minutes; mine returned in 3 minutes 6 seconds, 720×1280, 150 frames, mouth moving with the audio from about the two-second mark.
The reference budget has hard ceilings, and one number circulating in tutorials is wrong. Officially, wan3.0-video accepts at most 10 reference images; reference video is capped at 5 clips and 15 seconds total; reference audio likewise 5 clips and 15 seconds total; and first-frame or last-frame inputs are mutually exclusive with any reference input in a single request. Images, video and audio are counted independently, which is why the prompt markers read Image 1, Video 1, Audio 1 as separate series. The circulating claim of "up to 20 reference images" is false: 10 is the image cap, and 20 is the total across all reference types.
Treat lip sync as an optional advanced step, not a pipeline requirement, for three reasons. The platform's API layer has dedicated talking-head models (wan2.2-s2v, which handles cartoon subjects, and videoretalk), but the CLI at version 1.22.0 does not wrap them; I checked bl --help, so I am not going to invent commands for models the CLI does not expose. Community measurement puts AI lip sync at draft quality that still needs manual calibration. And overseas sentiment toward AI-dubbed anime has been resistant since the December 2025 discussion cluster on Hacker News. The --image-voice route above is the cheap way in; everything beyond it is deliberate extra work.
This is where the money goes: 5 seconds at ¥0.6 per second is ¥3.00. Which is why my advice runs opposite to most tutorials. Lock the panels cheap first; render video only once the sequence is right. Every screenshot above is a comic panel, and the clip exists only to prove the next step is real.
The budget, and the one cost rule that matters
| Step | Unit price | Usage | Subtotal |
|---|---|---|---|
| Character sheet | ¥0.18/image | 1 | ¥0.18 |
| Panels (reference edit) | ¥0.20/image | 3 | ¥0.60 |
| Voice line | per character | 12 chars | ≈¥0.01 |
| Panel pipeline | 4 images | ≈¥0.8 (~$0.11) | |
| Lip-synced clip (optional) | ¥0.6/second | 5 s | ¥3.00 |
| With one clip | ≈¥3.8 (~$0.53) |
For readers used to token math: there is no prompt-cache lever in image generation. Every panel is a fresh call, so panel count is the budget, full stop. The cache-read economics that dominate LLM bills do not apply here, which makes per-image price transparency the only number worth arguing about. I won't quote your remaining free quota; reading it needs a browser login I can't do for you, so check your own console's quota page.
On the free tier's promise, stated at its boundary: quotas are 90-day, per-model, and independent of each other, and the exact amounts live on each model's console detail page rather than in any single summary I trust. The honest summary is that free quota is enough to run this whole pipeline once as a demo, and not enough to finish one episode for free. At the community-reported 28.7% usable-take rate, thirty free video seconds do not even yield three usable shots.
Three cost numbers circulate in this space and they answer different questions, so here they are side by side without a verdict. Alibaba Cloud's own official comic-creation solution page, which I opened and read directly, quotes a 15-minute deploy and roughly ¥70 per episode with a free trial quota, and never mentions the CLI this post uses; that is the price of a managed, batteries-included route. My bare CLI run above is ¥0.8 for four panels, or ¥3.8 including one five-second lip-synced clip; that is the price of doing each step yourself. And first-hand media reports of beginner earnings put the other side of the ledger at ¥2.28 after thirty days online, ¥16.8 for eight episodes and 40k views, and one creator's summary of per-episode income as "not a few mao, but five-plus yuan." Quote cost and revenue in the same breath and you get nonsense; keep them in separate columns and the picture is at least honest.
A fourth number, labeled as an estimate because it is one: a 90-second episode runs about thirty shots, and at that same 28.7% usable-take rate you render far more than thirty clips to get them, which lands roughly ¥110 to ¥215 per episode at 720P or 480P. Nearly all of it burns on image-to-video retries; images and voice are rounding errors. That estimate, not the ¥0.8 demo figure, is the number to hold against the ¥70 managed-solution quote and the ¥2.28 first-month earnings report.
What you'll actually hit when you publish
Searching beginner questions surfaced two camps: one calling comic drama the last content dividend for ordinary creators, one calling it another course-selling cycle where every tutorial funnels into a paid course. On the second camp there is at least one official datapoint: BOSS Zhipin, a major Chinese job platform, issued a notice on 2026-04-16 naming "recruit-to-train" scams that bait applicants with "high-pay AI comic artist" and "earn 100k on your first drama" promises before collecting training fees. I'm citing the platform's notice, not naming any course or vendor. The counter-move is the same one that applies to the technique: these four commands cost under a dollar and under half an hour to run yourself, so the judgment is worth making firsthand.
Then there is the regulation nobody's tutorial has caught up with. China's NRTA Order No. 16, the Micro-Short Drama Development Management Measures, took effect 2026-09-01, and Article 34 requires a visible AI-disclosure label on every episode of an AI micro-drama. Filing splits by class: Class-1 applicants must hold a Radio & TV Program Production License, while Class-3 goes through platform self-review plus a program number. In practice the label is the part that reaches solo creators first, because it lands on the episode itself. How strictly the entity-qualification side is enforced varies by platform and I found no plain-language official guidance, so confirm against the platform you actually publish on before you act on any of this; treat my summary as a pointer to the order, not as legal advice.
If you publish overseas instead, the disclosure picture is narrower than people fear, and the scope is worth stating exactly. YouTube's disclosure requirement applies to AI that photorealistically alters or generates realistic content; animation and clearly unrealistic content are exempt from proactive disclosure, and non-photoreal AI labels surface in the expanded description rather than on the player. Since May 2026 YouTube also auto-detects videos carrying large amounts of photoreal AI material. For a comic or anime pipeline that materially lowers the compliance burden, and it is the reason this post's screenshots are panels rather than photoreal frames. The exemption is not a free pass elsewhere in the same platform: refusing disclosure on photoreal content can bring labeling, takedown, or removal from the YouTube Partner Program.
On the ownership side, the US position settled negatively and then stopped moving. On 2026-03-02 the Supreme Court declined cert in Thaler v. Perlmutter, leaving purely AI-generated works uncopyrightable for lack of human authorship. What remains protectable runs through three paths: selection and arrangement of the AI output, substantial human modification of it, or expressive human input that is clearly perceptible in the result; mere selection of prompts is not enough. There is no anime-specific rule, and I am not going to claim one. Practically, that pushes value toward the parts of this workflow a human actually authors: the sequence, the edit, the pacing, the panel selection.
The last publishing constraint is the one you choose yourself. Cloning a living artist's signature style for commercial work, or feeding photographs of real people in as reference, carries infringement risk; designing your own character and using your own generated sheet as the sole reference source is the clean path, and it is also the path this post's pipeline assumes.
Your turn
If you have pushed reference-image editing past three scenes, or into multi-character frames, I want the panel number where it broke. Same if you found a cheaper lock that survives a scene change. Comments are the point of posting this.
One expectation to set before you start: everything after the renders is manual. Choosing which takes are usable, shot order, pacing, grading, and the judgment calls in between are yours; FFmpeg will concat, subtitle and mix-export, and that is the extent of its opinion. Budget an evening of editing for every evening of generating.
The CLI lives on the install page; steps 3 and 4 need an API key from the console. And if you'd rather look around the platform first, start here.
Top comments (0)