Making AI short drama solo: three routes, real costs
I have one desktop with a 4070 and want to produce 3-minute AI short drama episodes alone. Is that feasible, what does it cost, and how long?
1 · What the situation actually calls for
Feasible — but the bottleneck is not image generation. It is character consistency and shot continuity. Those decide whether the cut is watchable, and whether you need to own the pipeline.
Three minutes per episode, one person, a local 4070: that means you don't need to rent cloud GPUs, but you also can't expect one-click output. The real cost is time, not money.
An overlooked precondition: short drama is narrative, not clip assembly. A watchable storyboard and dialogue come first; tools only amplify what's there. The reverse produces handsome footage nobody finishes.
2 · Three routes, compared
Route A: all-in-one platform (Toonflow / huobao-drama)
- Best for
- First-timers who want one finished episode within a week
- Cost
- Software $0 (open source); API usage roughly $2–6 per episode
- Time
- 3–5 days for the first episode
- Difficulty
- Low to moderate — no model knowledge needed
Start here. You give up control over style, but you learn within two weeks whether this is worth continuing — a more valuable answer than any technical choice.
Route B: build the pipeline in ComfyUI
- Best for
- Anyone producing a series needing a locked style and cast
- Cost
- Software $0; electricity is the main cost; 12 GB VRAM is enough
- Time
- 3–7 days to build, then 1–2 days per episode
- Difficulty
- High — nodes, samplers and LoRA basics required
Only pays off after episode three. Starting here, most people abandon during setup without producing a single episode.
Route C: outsource visuals, own the narrative
- Best for
- Writers who don't want to learn the models
- Cost
- External quotes typically $40–280 per episode depending on quality
- Time
- Depends on queue, typically 3–7 days per episode
- Difficulty
- Low, but you must be able to judge output
Only makes sense once you've proven the content gets traction and time costs more than money. Before you have an audience, it's the most expensive route.
3 · Recommendation
A then B: finish one full episode on an all-in-one platform to validate story and pacing, then move to ComfyUI to lock the style for a series. Skipping straight to self-hosting is the most common way these projects die.
4 · Day-by-day plan
- Day 1: write a 3-minute scene breakdown, 8–12 shots. Don't touch any tool yet.
- Day 2: produce one 15-second clip in Toonflow or huobao-drama. Confirm picture and audio both work.
- Days 3–4: complete all shots and lock one reference set per character, reused across the episode.
- Day 5: voiceover via GPT-SoVITS, aligned line by line to the storyboard.
- Day 6: edit, caption and score. This carries about a third of perceived quality — don't compress it.
- Day 7: watch it three times end to end and redo whatever makes you want to stop watching.
5 · What actually goes wrong
- Character drift: the same face changing between shots is the first flaw an audience notices. Carry the same reference set into every generation; a prompt alone won't hold a face.
- Shots that don't cut: each looks fine alone, jarring in sequence, because generation had no knowledge of the previous shot's end. Fix it by producing two adjacent shots at a time and checking the cut.
- Compliance: short-video platforms require AI-content labelling, and commercial use depends on each model's licence. Open weights do not mean unrestricted commercial use — check each one before distributing.
- Don't chase tools. This category turns over every few months. Your assets are the script, the character bible and audience feedback — not someone's node graph.
6 · Verdict
One person can do this at roughly a week per episode, with cash cost of $2–6 in API calls or zero if fully local. The real barrier is writing three minutes someone finishes; tools only set the visual floor. Ship one episode on an all-in-one platform before investing in a self-hosted pipeline.
Want this for your own situation?
Pro members describe what they're trying to build and get the same format back — route comparison, real costs, a day-by-day plan and the pitfalls specific to their constraints.